DE2149520A1 - Device for generating a linking scheme for character recognition - Google Patents

Description

2H9520

Boeblingen, October 1, 1971 jo-bue

Applicant: International Business Machines

Corporation, Armonk, N.Y. 10504

Official file number: New registration File number of the applicant: Docket RO 970 013

Device for generating a linking scheme for character recognition .

The invention relates to a device for generating a linkage scheme for the recognition of a number of input pattern classes. It also relates to the classification of Input patterns, such as alphanumeric characters, into a number of categories or classes, with a retroactive effect the establishment of the establishment of the links is provided. The facility itself is programmed in a memory General purpose computer implemented with a graphic display device.

Memory-programmed data computers essentially follow a relatively small repertory when performing their tasks of basic instructions that are executed at extremely high speed. Any information processing task that the data computer is to perform must, however, be expressed as a program composed of these basic instructions is. The instructions can be divided into two main types:

209816/1373

- Operation instructions which inform the data computer that it should request a given operation and

- Branch or logical instructions that the data processor specify where to find the next instruction to be executed.

With the help of these two types of instructions, a data processor can Execute programs that are both operationally composed (by having many operation instructions) and also have a logically complex structure (in that they contain many branch instructions).

However, humans have shared abilities in these areas. He can operationally bind program tasks in such a way that he can use them cannot actually perform such as path calculations; and he can carry out tasks so logically complex that he she cannot program, such as playing chess.

The objects of the invention include the limits of applicability of a data computer into the range of tasks that both humans and the data computer perform can, but which humans cannot practically carry out on their own. For this it is necessary that the data computer for writes his own program for a specific task. For example, a compiler allows a desired one to be specified Task in a highly structured language that is more or less similar to natural languages. The compiler translates then sometimes in a very complicated way the description of the input work in a sequence of directly executable steps in machine code. Since the internal structure of the human brain is very different from today's data computers, the (last) machine code is usually so different from the input representation that it is from a People cannot be understood who does the job first

209816/1373

Docket RO 970 013

2H9520.

described. Since people build data computers and not the other way around, It is assumed that machine decisions are made in a simple manner, while human decisions are made in obscure, mysterious "intelligent" ways. Thus, although it is relatively easy for humans to learn to make machine-like decisions, it has been found that it is very difficult to design machines that can make human-like decisions. In some cases, however, it is possible for the machine to learn to perform tasks that were believed to be only human.

The object of the invention is now also to provide a system for the design of a machine that imitates a person who imitates a machine himself, although he is only one has vague knowledge of the internal structure of the machine.

The machine recognition of patterns is of great practical importance. Most of the knowledge to date is recorded in a non-machine readable form, and much of the information is still in this form today laid down. However, the growing demands in the fields of commerce and research require that a considerable one Part of this information is made available for automatic data processing. Pattern and character recognition machines There are currently, but it takes thousands of man-hours to make machines that few Recognize classes of characters (or other patterns) in a small number of types of types. Since also the development of the recognition logic circuits is performed in a language that corresponds to the object language of the machine to be designed is very similar, this requires the use of carefully trained individuals who put their human skills into one in a very ineffective way.

It was also previously known to use data computers in the design of logic circuits for pattern and character recognition machines

Docket HO 970 013 209816/1373

2U9520

use. The simulation of full or partial Drafting, for example, is used to a very large extent for debugging and analyzing performance these recognition engines are used. There have also been retroactive graphic displays for real-time analysis of patterns and even used to develop the detection logic itself. The graphical representations obtained from these systems, however, touch on abstract data that form the input patterns relate to, for example, two-dimensional representations of space vectors of certain features, histograms and the like. However, all of these known methods require carefully trained operators and a great deal of time. None this method is ultimately suitable for mass processing necessary for commercial applications.

It is therefore a further object of the invention to avoid these disadvantages of the known character recognition. This includes in particular primarily the task of specifying a facility that allows the establishment of the link and the continuation of the states which performs pattern recognition.

This facility should be easy to understand and used by those with little or no experience have in the construction of links and have no other knowledge in the field of data computers. To this The task is also to provide a device that takes advantage of human intuition for the design of internal machine structures exploit.

For a device for generating a linking scheme for the recognition of a number of input pattern classes / s see the Invention in that there are provided:

an input device for the input pattern scans, a display device with a matrix of variables, Docket RO 970 013 209816/1373

2U9520

that of a number of predetermined parts of the input noise samples correspond, wherein the display device selectively a visible image of the states, which a number of variables assigned for a current scan are generated,

an actuator for the selection of a number of variables whose states form the picture and for which Specification of a name from a number of names of the input pattern classes and

a computing device connected to the input device, the display device, and the operating device and which is a program for accepting the present pattern scan from the input device for the transmission of the states of the selected variables to the display device, for the storage of the data, which represent the named variables and their states, in a recognition logic circuit with a given Format and for storing the specified name in the recognition association circuit.

Further features, advantageous configurations and developments the subject matter of the invention can be found in the subclaims.

The advantage that can be achieved by the present invention is that by means of the proposed economic device, logical identification links built up by inexperienced personnel can be.

Fig. 1 shows a schematic representation of the system according to the invention with various types optically represented images;

Fig. 2 shows an illustration of a logical decision

209816/1373

2U9520

dungsbaumes and a data structure for this tree;

Fig. 3 shows a sequence table of an inventive Computer programs and

FIGS. 4 to 11 show individual flow tables in FIG. 3

shown blocks.

System description

When an operator uses a design technique, draws a processing unit or a computer records its movements and uses them to construct a recognition algorithm, the is of a format acceptable for rapid execution by a character recognition engine. This machine can be hardwired Be a logic unit, a special processing unit or a general purpose computer. The operator's ability to build Good detection logic depends on its skill in recognizing the pattern and not in his skill in building logic circuits or logic machines. He does not need any knowledge of the internal structure of the machine he is constructing or the machine on which he is working Draft prepared.

The system shown in Figure 1 includes a general purpose central processing unit (CPU) 101, a source 102 for input patterns, an output device such as e.g. B. the printer 103 for generation a record 106 of the completed links, a graphical one Display device 104 and light pen actuator 105. Pattern source 102 is referred to as a tape file; it but can also be another source such as B. a scanner can be used. In the same way, the output device 103 the form of a disk file, tape device, or other

Docket ro 970 013 209816/1373

2U9520.

assume similar unity. At least some of the functions of the light pen 105 can be performed by instruments such as a keyboard, a Roikugelgeber etc. are taken over. A stored program 300 coordinates the mutual influencing of the units 101 to 105, generates data for the visible image on the optical display device 104 and constructs the recognition links according to the invention.

If the program has been entered correctly, an image appears on the screen of the display device 104. Various household operations but are necessary beforehand. First, the light pen 105 is placed on one of the points 110 to create a link memory to hold the current set of links. A dark mark appears at the point 111 corresponds to this buffer and the name of the buffer appears in position 112. Next, the pen 105 is on one of the Points 120 placed to select a video buffer for holding the order of the input patterns to be processed. A Another character appears at point 121, which corresponds to the switched on Video buffer. For later use in the program, the buffers can be named under any name on list 130 with a pattern can be loaded which corresponds to the name directly above. After selecting the activated buffer, the Set the light pen 105 to the word "READ" in the operation list 140. Then the pen will point to the letter 131 in the list 130 led. Thereby, a pattern representing the letter "A" is loaded into the buffer 121. A picture of this pattern can be examined on the matrix 150 by placing the highlight on the word "VIDEO" on the operation list 140. The image appears as a series of vertical stripes 152 overlying dots 151. The picture can be erased through Touch the word "DELETE" with the light pen. Next, the Α pattern is transferred to the buffer at A 131 by the word "TRANSFER" and touch points 121 and 122 with the light pen. The buffer 121 remains the active video buffer. For the remaining characters in list 130, this becomes

Docket RO 970 013 209816/1373

2H9520

Process repeated. The system is now ready to begin building a link to recognize patterns that are to be belong to the classes specified by name in list 130.

The following is an overview of the basic tasks to be carried out Procedural steps. The CPU 101 fetches an input pattern from of tape file 102 and set it on an internal board. The pattern is in the form of 1024 variable or binary number parts in an arrangement 32 χ 32. Each number represents a location with X and Y coordinates in the pattern. The value of a each number (0 or 1) represents the intensity (white or black) of the pattern at the point with the relevant coordinates. At the beginning the visible picture on the board is empty and has only the 151 points to indicate the positions of the variable sample parts. The calculator simulates previous experiments and conclusions on the pattern and puts the results in the matrix 150. The CPU 101 can be made to depict the pattern intensity at the selected points 153 and record certain data in a link tree. On the other hand, an attempt can be made to identify the pattern. The CPU 101 then compares the proposed or provisional one Name with a current, d. H. actual name that the specimen and records bear in the tree. Another Patterns can be retrieved from tape file 102 with or without performing the above operations. Certain others Operations to aid the discovery process are also available. When the discovery link is complete a record 106 of that link is read out.

A pattern is selected for recognition by one of several methods. Touching "NEXT" in the operation list 140 selects the next pattern from tape file 102 and places it in the active video buffer. If any parameter "BELIEB" is on, any number of file patterns will

Docket RO 970 0X3 2.09816 / 1373

2H9520

skipped before a pattern is selected. The status of this parameter fluctuates between on and off as soon as the command is given "PLEASE." activated by the light pen 1O5. Activation the command "AUTO" in the operation list 140, briefly too Called list 140, the CPU 101 causes several coughs to be read in sequence until it finds one which is linked by the link is incorrectly referred to as already established. This command becomes more useful later in the procedure. the Activation of "READ" followed by a name from list 130 causes CPU 101 to search the sample file until it finds a pattern with the specified name. During the actual setup process, this command can be used for voting and debugging logic for a particular, hard-to-identify character or to differentiate between two similar characters contribute. The operator can reset or rewind the tape file by touching "BACK. Or" BACK ".

If the matrix 150 is empty at the beginning of the build cycle, at least some locations of the input pattern must be checked before identification can be attempted. When the light pen 105 touches one of the points 151, the state or intensity of the pattern at that point is recorded: "0 ^N stands for white while" X "stands for black. The intensity symbols 153 remain after the light pen is lifted. Select For example, if you place a point approximately in the middle of the matrix 150, you can almost instantly distinguish between characters such as 0, U, C etc. and A, H, X etc. If the point is white ("0") and you then select a point approximately in the middle at the top of the matrix, you can distinguish between 0 and H, etc. Any number of points 151 can be selected in any order. To undo the last choice you made, activate " NODE "in the list 140, the corresponding node is removed from the tree and the character is deleted from the matrix 150. In order to delete the whole series of tests carried out on the current sample,

Docket ro 970 013 209816/1373

2U9520

which takes all nodes out of the tree by touching "AST", and so the corresponding characters in the matrix are deleted again. Pressing the "TREE" command triggers an acoustic signal Alarm signal off and a second actuation of "TREE" deletes the whole tree and its graphically represented Sign 153.

When enough points have been checked to identify the current pattern, this pattern can be given a tentative name by adding the light pen 105 is placed on the corresponding letter in the list 130. The computer checks this preliminary Names with the actual name that accompanies the pattern. If the name is correct, it will be attached to the recognition link appended and a different pattern can be chosen for checking and / or identification. If the classification is incorrect, an acoustic signal (not shown) is emitted and the CPU 101 awaits the next command. If the Input patterns in the tape file 102 are not named or if a classification is to be overridden then enforced touching "ID" followed by a letter from list 130 an identification.

If a number of tests (or even no tests) have been performed on the current pattern or if a correct identification has been achieved, a different input pattern can be selected by one of the methods described above. The CPU 101 then reads another pattern from the file 102 into the active video buffer 121 and the cycle begins again. The image indications by the points 153 of the locations checked for all of the preceding patterns are sequentially transferred to subsequent patterns. Whether a checked point contains a 0 or an X, however, always depends on the status (white or black) of the current pattern that is available at the respective point. The graphic representation of previous test results can be deleted by selecting the "BLNK" command. This deletion avoids confusion during rapid processing of a

209818/1373

Docket RO 97O 013

2U9520.

large number of patterns by the "AUTO" function when searching of an error pattern. As the process continues, there will be an increasing number of samples as the number of exams decreases of each individual pattern identifiable. When an acceptable recognition rate has been reached, the program can exit are activated by touching "END" in the list 140 with the light pen 105. This command receives the complete recognition link. It can be output from the CPU 101 directly in the form of a printed record 106 or on another Medium such as B. a plate stack (not shown) can be obtained.

The first stages of the procedure can be particularly unusual for Character sets or inexperienced operators can reduce the number of exams by trying different ones Types of auxiliary images used for displaying images on the Matrix 150 are available. The image "VIDEO" has already been described. Although the CPU 101 normally only uses the previously checked Entering the current pattern into the active video buffer, the whole pattern can be presented for review at any time will. It sometimes happens that the pattern e.g. B. very washed out, the real name wrong or the pattern not part of a is class shown in list 130, so recognition should not continue for this pattern. Other important ones Reasons for the pictorial representation of the entire running pattern can also arise. Choosing the command "SHOW" and the subsequent designation of one of the video buffers 120 results in a vertical stripe image similar to image 152 in FIG Matrix 150. Such an image is used to select the points 151, which support the decision as to whether the current pattern is a belongs to a certain class or not. If checkpoints 151 for the Differentiation of two classes from each other should be chosen, a second pattern 154 can be superimposed on the "SHOW" or "VIDEO" image by selecting "OLAY" from the list 140, followed by touching a video buffer point 120 with the light pen 105. The image 154 is made up of horizontal stripes

'> 09816/1 3 73

Docket RO 970 013 "

2H9520

together. Changed the existence of "OLAY", "VIDEO" and / or "SHOW" does not include the configuration of symbols 153 in matrix 150 that represent the previous checkpoints. These characters are are not deleted and represent intensity states of the current pattern and not those of the patterns shown. The "VIDEO", "SHOW" - and "OLAY" images are deleted by pressing the "DELETE" command. If there is more than one of these images all in matrix 150 are deleted.

Various additional operations are available for the detection of distorted ones or incorrectly registered samples are available. Handwritten characters include B. very often. The operations "+ X", "-x", "+ Y" and "-Y" shift the current pattern accordingly up, down, right and left within the active video buffer. The matrix 150 represents the combination results for the characters moved, but the link itself does not change under the conditions of the move. the operations selectable by the instructions "CHAIN" and "CYCLE" on the other hand, transform the link while leaving the current pattern unchanged. Transformations are foreseen for moving, skewing and rotating the link relative to the pattern. The activation of the "CHAIN" command by the Pen 105 initiates a series of ordered transformations. The repeated actuation of "CYCLE" causes the application of a predetermined sequence of transformations at a specific point in time. The configuration of the symbols 153 in the matrix 150 will of course change as each transformation is applied. One pressing the chain a second time releases all transformations from one another.

Structure of the links ( Fig. 2)

According to the above description, an operator with little knowledge of computers or character recognition machines can

Docket RO 970 013 709816/1373

2H9520

Shortcuts for such machines using only common ones of well-known concepts. The operator does not need to know the actual form of the link different types of linkage structure can be used for this purpose. Another area of application of the present invention is seen in the recognition of fonts by general purpose computers. Digital decision trees with multiple outputs can be effectively used and searched in such computers, and especially this type of linkage is thus chosen for a more detailed examination.

2 shows a schematic representation of a decision tree 200 with multiple outputs. The tree 200 is addressed through the input 201, in which a first node 211 of a set of internal Resume or Check Node 210 is addressed. The knot 211 tests a particular pattern variable for its state or value. Though the changeable take different forms they are described as sample parts or locations represented by X and Y coordinates. Every point in the 32x32 matrix 150 of FIG. 1 represents such a location. Node 211 retrieves the value or parity of the pattern variable at the point Xl, Yl. If the parity is 1 (black), branch 212 to node 213 is taken and the cycle repeats. If parity is 0 (white), branch 214 to node 215 is taken. The tree 200 is exceeded by the internal nodes 210 until a branch, such as e.g. B. branches 216, 217 or 218, completed is. These and other branches lead to a series 220 of exit nodes. The finished tree contains at least an exit node for each class in the set of the class to be recognized. Two or more outputs such as B. the nodes 221 and can denote the same character or the same class. Other outputs such as nodes 223 and 224 can be reject nodes indicating that the pattern from the respective set of test results obtained on reaching these nodes cannot be classified. If an exit node has the lukewarm

970 ο »209816/1373

The tree 200 is always built up by replacing the reject nodes with check nodes. the Selecting a new pattern "as the current pattern causes the tree 200 to be traversed through a set of nodes and branches, which correspond to the intensity values or parities of the current pattern until one reaches a connection or output node 220 hits.

The data structure 230 for a set of trees 200 includes one . Number of pages 231, and each page a number of double words 232. Depending on the storage medium used, each link tree 200 can be set on one or more pages 231. Page and double word structure can be made transparent to the programmer. Each double word 232 contains two Individual words or node words 233. For an internal or test node 210, each individual word 233 has a structure 240. The fields

241 and 242 contain 6-bit representations of the X and Y coordinates the part of the current sample to be tested. The remaining fields contain a pointer to a subsequent pair of nodes, d. H. to another double word 232. The pointer has a page address 243 and a double word address 244. Since the parity of each pattern position is either 0 or 1., this number can be

) for addressing one of the double words 233 in each word pair 232 be used. If z. For example, it is assumed that the single word 240 represents the node 211, then the fields contain 241 and

242 the coordinate values Xl, Yl, the fields 243 and 244 the address a double word 232 containing a pair of single words 233 which designate nodes 213 and 215 respectively. Which of these two single words actually chosen is dictated by the parity (0 or 1) of the pattern at the point Xl, Yl. Output nodes 220 are identified as such by the existence of zeros in field 251 of single word 250. A die Code representing sample class is located in field 252. The exact form of this code is arbitrary. For the sake of simplicity however, you select the value "000000" in field 252, indicating that it is a rejection node.

Docket RO 970 013 209816/1373

2U9520

Structure of the program

The program 300 shown in FIG. 3 is started with the start block 301. Block 302 initializes a number of parameters and subroutines used in the remainder of the program also set and generate video and link buffers Data for the image representations on the display device 104. Its exact functions and its structure depend on the particular Language and the forms in which the program is held. Block 303 then senses the coordinates of the inputting light pen (LPI) 105 on the display device 104 and decodes the operation to be performed. A decision with multiple branches directs control to the various blocks to carry out the selected operation.

For the sake of simplicity, the operations have been divided into several general categories. The basic operation blocks include Blocks for connecting a test node (500), testing a pattern name (600), retrieving the next pattern (700) and pictorial representation of the link (703, FIGS. 5 to 7). Blocks 500, 600 and 700 are described in more detail below. The block 305 actuated by the light pen ("ID" on the list 140) packs a code similar to that selected from the list 130 Name corresponds to, in the field 252 of a current node word 250. The "END" actuated block 306 draws the active one Link tree 400 on output record 106. Line 308 from this block then terminates the program at Block 309.

Routines 900 for handling the file 102 of input patterns is described in connection with FIG. 9. Transformation operations 1000, video display routines 1100 and management functions 400 are described in connection with FIGS. 10, 11 and 4. Line 310 from blocks 500, 600, 305, 306, 307, 900,

Docket ro 97o 013 209816/1373

2H9520

1000, 1100 and 400 return control to block 303 where subsequent operations are decoded for further execution will. I. E. the program 300 decodes a command for the corresponding operation and expects in the manner of an interpretive one Compiler the next command.

4 shows various management functions 400 with more general Use in program 300. The selection (410) of an active video buffer point 121 is made by sensing (411) the coordinates of the light pen 105, moving (412) a character to these coordinates, setting (413) the active buffer and Return to line 310. A link buffer is selected (420) by sensing (421) the coordinates of the light pen 105 and moving (422) a character to point 111. If the link in the selected buffer is already in main or core memory is (output 423 of decision block 424), the link for the current pattern is skipped (800) and the Control returned on line 310. If the selected link is not in the core memory (output 425), the Old link currently in the core memory on a disk storage device (not shown) or another large storage area retained (426) and read the selected link into core memory. Control then returns over the line 428 back to block 800.

The selected link can, according to the description given in connection with FIG. 1, by the commands "NODE", "AST" and "TREE" can be modified. When TREE is selected by light pen 105, routine 430 triggers an audible signal (431). Since this is a somewhat dangerous operation, routine 430 (432) waits for a second choice (433) of the "TREE" command. When the second choice is made (434) will be all node words 240 and 250 of tree 200 are reset to their initial value (435). If the tree command is not re-elected negative output 436 returns control directly on line 310. For removing a number of knots

Docket ro 970 013 -. '098 1 (5/137 3

the "AST" routine 440 calls (441) the pointer 244 back to the last previous node word 250. The pointer 244 was previously stored in a special place by the routine 800. The word, which had represented the last connection node 220, but which has now become an internal Node 210 is fetched (442) and its pointer is deleted (443). This effectively removes all of the nodes in the The row is broken off by removing the entrance to the first node in the row. The "NODE" routine 450 initiated by a clear signal from light pen 105, clears only the immediately preceding node in a row. The pointer 244 on the The last selected node retained by routine 500 is retrieved (451) and thereby the retrieval (452) of the last one Node. Entry to this node is then interrupted by clearing (453) its pointer. All routines 410 through 450 return control to block 303 via line 310.

5 shows the routine 500 for connecting a test or an internal node to the tree 200. The block 501 senses the Χ, Υ coordinates of the light pen 105 on the matrix 150. Block 502 inverts the active transformation into the form of a new coordinate pair X ¹ and Y *. Transformations and their inversions are described in connection with FIG. Block 503 generates a pointer value corresponding to the address of the next free double word 232 available in the memory area previously set aside for tree 200. There are numerous conventional memory allocation routines used to accomplish this task. Block 505 assembles the pointer and transformed coordinates into a single word format. Block 505 retrieves the parity bit (intensity value) of the input pattern at the transformed location X ¹ , Y ¹ . Parity bit and pointer together define the address of a single word 233 of the next available double word. The block 506 sets the pointer assembled and the values for X ¹ and Y ¹ in the fields 241 to 244 of the lau f node definition 240. The block 507 loading

Docket ro 970 o »209816/1373

2U9520

preserves the pointer and parity values for the current node word as well as its address (i.e. the pointer value of the previous node word). These preserved values are used in the "NODE" operation 450 in FIG. 4 is used. Control then passes back on line 310.

6 shows details of the pattern name checking routine 600. The coordinates of the light pen 105 on the list 130 define a provisional name for the sample of the current pattern (601). Block 602 retrieves the actual name ψ of this pattern from file 102. If the provisional name matches the actual name (output 603, block 604), the name is inserted into field 252 of the current node word (605). When the values of all node words are initialized to 0 by block 302, the remainder 251 of the word designates this as a connection node word 250. If provisional and incorrect name are not the same (output 601) an alarm is given (607). Blocks 605 and 607 return control to line 310 again.

Routines 700 shown in FIG. 7 find the next input pattern sample from file 102, run the active link tree on trial and give control over the output line 311 back to block 307. When block 303 decodes a signal from light pen 105 indicating the "NEXT" operation of list 140, block 711 of routine 710 queries the state of the any parameter. If it is switched off, arrange for it the output 712 the block 713 for reading the next pattern from the file 1Ο2 into the active video buffer. The subroutine 800 then operates the active linking of the pattern, as far as it is set up for it. If the arbitrary parameter is switched on (output 714), block 715 generates an arbitrary number and block 716 skips that number of patterns in file 102 before block 713 is allowed to read the next pattern. Routine 720 is triggered by an "AUTO" signal from the light pen 105 is selected to block 303. Block 721 reads the next

r ^. 209816/1373

Docket RO 970 013

2H9520

Pattern from file 102 and operates the link (800) on the pattern which is the provisional name of the pattern on the, Field 252 of the connection node word 250 encountered in each case. In the last step of each pass of the tree 200 only one connection node 220 occurs. Block 722 gets the actual name of the current pattern, and block 723 compare this with the provisional name. Do the names differ from each other (or is the connection node a rejection rejection node, since field 252 of a rejection node does not match the name of a sample class from list 130), output 724 to line 311 is selected. If the names match the output 725 leads to decision block 726. If no other light pen command interrupts the sequence, output 727 returns control to block 721 to read another pattern. If "AUTO" was interrupted, the returns Output 728 returns to line 311 regardless of the accuracy of the pattern name.

9 shows operations 900 for controlling tape file 102. Block 910 reverses the state of the any parameter used in block 711 (on or off) every time the light pen 105 touches the command 'ANY' on the list 140. The block 920 resets file 102 to the previous pattern and block 930 rewinds file 102.

The "operate link" application routine 800 is carried out by various other routines in program 300 are used. It is not directly addressable from list 140 of port 104. Of the Block 800 initializes the link to the address of the stored word for input node 211 of tree 200. The Block 802 retrieves the appropriate word. Block 803 determines whether the retrieved word is an internal node word 240 or a port node word 250 is of the label type or the rejection type. Is it an internal node (output 804) If this is unpacked on the node (805), the fields 241 and are transformed by the active transformation (806) and the

209816/1373

Docket RO 970 013

2Ί49520

The parity value of the corresponding pattern location is found (807). The parity value and pointer field 244 for the current node suffice for determination of the next word by block 802 over line 808. If an identification node is found output 809 transfers control to block 810 which indicates the decision reached by the link. If the chain parameter is off, output 813, block 814 and output 815 return reject nodes and identification nodes to block 810. The address of the connection node encountered (i.e., pointer and parity to the node) are retained (811) for use in branch operation 440, and the line 812 returns to the routine from which subroutine 800 was called.

Fig. 10 shows various data modification routines 1000. Routines 1020 and 1030 control the modification of the link data in block 806 in Fig. 8 by choosing adjacent transformations of various linear types. The transformation or concatenation parameter is inverted (1020); off to on and vice versa, by touching the "CHAIN" command on list 140 with light pen 105. If X and Y are the original pattern coordinates, the above-mentioned transformations derive the modified coordinates X ¹ and Y ¹ from the relationships with the the general form listed below.

2I

_ ^B 2

This matrix equation represents the most general linear transformations in Cartesian coordinates. B. A = A ₉₉ = X and A = A ₉ . = 0 a shift of χ + B ,, Y + B

if

Docket RO 970 0X3

= 0 and Β _χ =

= 0, the coordinate

09816/1373

2H9520

enlarged in one or both directions. Values other than zero for A _o and A create a slant in the co-

"21
ordinates, and rotation can be created with repeated inclinations. The parameters for the corresponding inverse transforms for block 502 can easily be determined by a known linear algebraic technique. In contrast to the new generation of parameters for each application of each transformation, the parameters for several transformations are stored in an ordered cyclic list. Each transformation can be addressed by a chain number that specifies a specific part of the list. The chain number can be incremented by the routine 1030 each time the light pen 105 the command "CYCL." touched on list 140. Routines 1020 and 1030 both return to line 310.

When the chain command is on, a node appearing at output 813 of FIG. 8 goes to output 816 of the decision block 814. Block 817 then advances the chain number to select the next transformation in the list. if the new number has completed a full cycle (i.e., when each transformation has already been applied to the link) the output 818 of the block 819 causes control to the block 810. In the opposite case, however, the output 820 causes the re-initialization of the tree by block 801. Thus, if concatenation is on, the appearance of an Rejection node at output 813 results in routine 800 being run through repeatedly, each time with a different one Transformation in block 806 until identification is achieved (output 809) or until each transformation has been unsuccessfully applied was (output 818).

The routine 1010 is actuated by the light pen 105 from one of the four shift commands on the list 140, accordingly the decoding in the decision block 1011 equipped with several outputs 1012. For the selection of the commands "-X", "+ X",

Doocet ro 970 O ₁₃ 209816/1373

2H9520

"-Y" and "+ Y" shift blocks 1013-1016 of the pattern bits in the active buffer by one bit position accordingly to the left, right, up or down. The multiple choice of a shift command shifts the pattern a larger number of positions. The locations corresponding to the locations previously selected by tree 200 Shifted sample bits are then represented on matrix 150 by block 1017. Routine 1010 affects the linkage in no way, just a sample of the current pattern.

Routine 1100 shown in FIG. 11 creates and manipulates the Z us atζ template images shown in FIG denotes the desired name of a pattern by sensing (1111) the coordinates of the light pen 105 on the list 130. Block 1112 reads the actual name of the next pattern in file 102 into the active buffer. If the name matches the desired name, output 1113 of decision block 1114 returns control on line 310. If the names do not match (output 1115) and a specified maximum number of patterns has not yet been read, the output 1116 of block 1117 returns to block 1112 to read another pattern into the active buffer. if exit 1118 terminates routine 1110.

The shift routine 1120 fetches (1122) a pattern from a selected by the light pen coordinates on the line 120 (1121) Video buffer and stores (1124) patterns in a selected (1123) by a subsequent set of light pen coordinates Video buffer.

The pointing process 1130 senses (1131) the highlight pen position on line 120 and fetches (1132) that in the selected video buffer saved patterns. The video routine 1140 fetches (1141) the pattern from active memory. Set these two routines the selected pattern together (1142) into a series of vertical stripes as shown at 152 in FIG. The block

209816/1373

Docket RO 970 013

2U9520

1143 automatically deletes all previous images that were active at that time. The "OLAY" operation 1150 senses the coordinates of the light pen 105 on the dotted line 120 (1151), fetches (1152) the pattern from the selected buffer and assembles it as a plurality of horizontal strips (1153) as shown at 154 in FIG 1 are shown. The "0LAY ^M command does not delete any previously selected patterns. All three routines then display the composite pattern on matrix 150 (1144). Image display is maintained after control is returned to line 310. Delete operation 1160 cancels all of the above mentioned three types of patterns from the matrix 150. The previously mentioned operation "BLNK" 1170 prevents the block 307 from displaying the linking results. When ' ¹ BLNK "is off. the output 1174 of block 1172 returns directly to line 310. If "BLNK" is switched on, the output 1176 is selected in order to block 1177 and thus the link representations by bypassing the block 307 shown in FIG. 3.

? Π 9 8 1 6/1 3 7 3

Docket RO 9 70 013

Claims (46)

2U9520 PATENT CLAIMS Device for generating a linking scheme for the recognition of a number of input pattern classes, characterized in that the following are provided: an input device (102; Fig. 1) for the input pattern scans , a display device (104) having a matrix (150) of variables representing a number of predetermined parts of the input pattern scans, the display device selectively displaying a visible image which generates states associated with a number of variables for a current scan, an operating device (140; 105) for the selection of a series of variables, their states form the image and for the specification of a name from a number of names of the input pattern classes and a computing device (101) connected to the input device, the display device and the operating device is connected, and the one program for accepting the present pattern scan from the input device, for the transfer of the states of the selected variables to the display device, to store the data that represent the named variables and their states in a recognition logic circuit with a predetermined format and for storing the specified Name in the recognition combination circuit Docket RO 97Q 013 2U9520 executes. 2. Device according to claim 1, characterized in that the parts of the pattern areas within the pattern scans accordingly, where the areas are determined by a number of coordinates and where the states represent a certain intensity of the visible image at the relevant areas. 3. Device according to claim 1 and / or 2, characterized in that that the recognition combination circuit a Contains sequence of the selected variables and their assigned states in the order in which they were selected. 4. Device according to one or more of claims 1 to 3, characterized in that the recognition combination circuit has the format of a decision tree made up of a number of nodes and branches. 5. Device according to one or more of claims 1 to 4, characterized in that the decision tree from consists of the following elements: a number of nodal words, each of which is a representation of the coordinates of an area an identifier for at least one predetermined state of this one area and finally at least has a pointer and a number of leaf words, each having a representation of one of a number of names, said leaf words being addressable by at least one of said pointers. 6. Device according to one or more of claims 1 to 5, characterized in that the input device 2 nights 9 8 1 6/1 3 7 3
Docket RO 970 013 2U9520- also provides a number of actual names each associated with particular sample scans and that the computing device is also used to compare an actual name of the present pattern with a specified name and to inhibit the recording of the specified name in the recognition association circuit, if the specified name does not match the actual name, provided is. 7. Device according to one or more of claims 1 to 6, characterized in that the computing device is a Outputs an error signal if the specified name does not match the actual name. 8. Device according to one or more of claims 1 to 7, characterized in that the display device is a provides a visible list of all names and that the computing device is at the selected position of the name list in the actuator to identify the specified name. 9. Device according to one or more of claims 1 to 8, characterized in that the computing device further generates an output of a record of the recognition linkage, 10. Device according to claim 9, characterized in that the actuating device sends a signal for the control of the output of the recording. 11. Device according to one or more of claims 1 to 10, characterized in that the input device has a file with a number of sample scans, that further the computing device further pattern scans from the file in a specified sequence. t RO 970 013 209316/1373 2H952Q. takes and to the display device the states at least transmits some of the previously selected variables and that finally the display device provides a visible picture of the states of the previously selected variables. 12. Device according to claim 11, characterized in that the actuating device is one of a number of possible Sequences are specified and given to the computing device as said predetermined sequence. 13. Device according to claim 12, characterized in that the outgoing sequence is any sequence of pattern scans is in the file. 14. Device according to claim 12, characterized in that the file further contains a number of actual names, the very specific pattern scans and that furthermore the computing device continues the recognition linkage for a preliminary sequence of sample scans builds up preliminary names for the scans The provisional sequence also developed the provisional names with the actual names of the samples the provisional sequence and that it finally compares a predetermined sequence of those samples whose actual and tentative names do not match. 15. Device according to claim 12, characterized in that the computing device carries out a next sample scan the specified sequence after performing a specified number of selections from the series of variables accepts. 16. Device according to claim 12, characterized in that the computing device stores the data in the Docket RO 970 O ₁₃ 209816/1373 2H9520 Recognition logic circuit for those selected variables which have a first state in the present sample scan if corresponding to the Variables have a predetermined state in at least a subset of previous samples in the predetermined Had a consequence. 17. Device according to claim 16, characterized in that the predetermined state is the first state. 18. Device according to claim 17, characterized in that the first state is the background of the pattern scans represents. 19. Device according to claim 17, characterized in that the predetermined state of the said first state opposite state is. 20. Device according to one or more of claims 1 to 19, characterized in that the actuating device also generates a removal signal and that the computing device for removing at least a part the detection link as a function of this elimination signal eliminated. 21. Device according to claim 20, characterized in that the recognition link is in the format of a decision tree and that the actuator has a number of elimination signals for selective elimination specified nodes and esters of the tree are generated. 22. Device according to one or more of claims 1 to 21, characterized in that the actuating device generates an indication signal that further the computing device provides a complete pattern to the display device transmits, the pattern to at least one of the Docket ro 970 013 209816/1373 2U9520 One of the input pattern classes and finally the display device creates a visible image of the entire pattern on the display matrix (150; Fig. 1). 23. Device according to claim 22, characterized in that that complete characters is the sample scan at hand. 24. Device according to claim 23, characterized in that the actuating device selectively comprises at least one
a number of display signals representing the name
of the input patterns and that the computing device further selectively transmits to the display device at least one of the complete patterns, a pattern belonging to a class identified by a display signal
was specified. 25. Apparatus according to claim 24, characterized in that the computing device further selectively transmits another one of the complete patterns to the display device
in dependence on a further display signal and that the display device simultaneously images of the
one and the other complete pattern is generated. 26. The device according to claim 25, characterized in that the display device further simultaneously contrasts images Representations generated. 27. The device according to claim 26, characterized in that the input device has a file which contains a number of input pattern scans and a number of actual names which are each assigned to the corresponding pattern scans and that the computing device searches the file for at least one of the pattern scans, who have an actual name,
which corresponds to the class indicated by an indication signal Docket ro 970 013 209816/1373 was specified. 28. Device according to one or more of claims 1 to 27, characterized in that the actuating device further generates a transform signal and that the computing device takes the shape of one of the input pattern samples changes according to a predetermined affine transformation as a function of the transformation signal. 29. Device according to claim 28, characterized in that the input pattern scan is the present pattern scan is. 30. Device according to claim 29, characterized in that the predetermined transformation is a shift of the present one Pattern scanning is effected linearly in a specified direction. 31. Device according to claim 30, characterized in that the predetermined transformation is the change in the size of the current pattern scan relative to the matrix causes. 32. Device according to claim 30, characterized in that the predetermined transformation is the rotation of the current present pattern scan is effected relative to a predetermined direction. 33. Device according to one or more of claims 1 to 32, characterized in that for the input pattern, which belong to a predetermined set of categories, a pattern classifier is provided which has the following: a receiving device for a number of scans * Docket RO 970 013 2098 16/1373 2U9520. stungen the input patterns, which belong to different categories belong, an organizer for organizing the pattern scans into arrays of a number of terms, these terms having designations corresponding to the coordinate and intensity values of a number of points of the sample scans correspond, a receiving device for receiving the designations of certain members, a transmission device for selective transmission of representations of certain terms for the display of the intensity values in a matrix that contains a Has number of positions that correspond to the coordinate values, a modifier responsive to the transmitted representations and the structure of the pattern classifier modified, a receiving device for receiving the designations of estimated category names of the sample scans and to generate verifications of the correctness of the estimated category names, a verification device, responsive to the verifications, for displaying representations of the categories being estimated to be used at specified places in the pattern classifier and an output device for outputting a representation of the pattern classifier. Docket RO 970 013 709816/1373 2H9520 34. Device according to claim 33, characterized by an extraction device for extracting a number of coordinate values of the arrangement members from the pattern classifier and a transmission device for the selective transmission of representations of the intensity values of the extracted arrangement elements of the current present sample for display in the matrix. 35. Device according to claim 34, characterized in that the pattern classifier is made from a logical decision sbaum exists. 36. Device according to claim 35, characterized in that the receiving device for the scans also has a Number of actually existing categories receives names associated with the sample scans, respectively and that further the designation receiving means provide the actual category names with the estimated category name compares and that the device is blocked for the insertion of the representation if actual and estimated category names do not match. 37. Device according to claim 35, characterized in that a device for generating a sequence for the Control of the order in which the sample scans are received by the receiving device is provided is. 38. Device according to claim 35, characterized in that the following is also provided: a comparison device for comparing the intensity values at corresponding coordinate values a number of specific members belonging to different pattern scans and Docket RO 970 013 2 0 9 8 16/1373 2U9520 a locking device for locking the device for inserting representations for those Members whose intensity values have a predetermined relationship to one another. 39. Device according to claim 35, characterized in that the following are also provided: a receiving device for receiving further designations of category names, a selection device for the selection of patterns which have names corresponding to the further drawings correspond and a transmission device for selective transmission of representations of the latter patterns for display in the matrix. 40. Device according to claim 35, characterized in that there are further provided: a receiving device for receiving designations a transformation and a changing device for changing the coordinate values of a number of points of the pattern in the arrangement, the change being made accordingly to the transformation. 41. Device according to one or more of claims 1 to 40, characterized in that for the modification of the recognition, the recognition of predetermined areas of an input pattern processed for identification, with the input pattern as a member of a class in a Is a group of possible classes and where the next Docket RO 970 013 209816/1373 2U9520 The following algorithm is processed: Receipt of a sample sample belonging to one of the classes Carrying out the recognition with the above-mentioned pattern scanning as the input pattern, Generating a visual display of data indicative of at least some of the predetermined parts and is processed by the recognition, Receiving an indication of a selected portion of the sample scan, Adding data showing the selected part to the visible image, Modification of recognition depending on data indicative of the selected part and Implementation of the link establishment by adding another Pattern scanning used as the input pattern will. 42. Device according to claim 41, characterized by: the receipt of an actual name, that of the sample scan assigned, Receiving an indication of a tentative name for said sample scan , Comparison of the actual name with the provisional name and Docket ro 970 013 209816/1373 2U9520 Blocking of the modification of the recognition depending on the data of the selected part, if the actual name is different from tentative name. 43. Device according to claim 42, characterized in that it is further provided: the receipt of an indication of the class from the group of classes and Generating a visual display of data similar to the previously mentioned display of an image of a given Link of the latter class is superimposed. 44. Device according to claim 42, characterized by: the selection of a first part of the input pattern, Testing of the part against a number of predefined possible states. Identification of the input pattern when the first part has a predetermined state and when one the specified recognition criterion is met, Selecting a second part of the input pattern if the first part is a first of several possible Has states Selecting a third part of the input pattern if the second part is a second of several possible States and Repetition of the above for other specified Docket RO 970 013 209816/1373 2H9520 Parts. 45. Device according to one or more of claims 1 to 44, characterized in that the computing device with which a source of linkage patterns, a source of Control data and a graphic display device are connected, is controlled by a program which causes that said patterns and control signals are accepted as input data in order to produce a record as output data to generate a logical decision tree for the identification of the pattern, with the program from tangible representations of the following components: a first routine for the sequential storage of a number of said patterns in a work area, so that a number of parts of the stored patterns are individually addressable, a second routine, responsive to a first type of control data, for selecting selected pattern parts and data representing the areas of the selected parts in the patterns to the display device transferred to, a third routine responsive to the control data; to compile data in the memory record, which is indicative for at least some selected parts and for the order of their selection is and a fourth routine that acts on the data in the memory record responds to further selected sample parts and data representing the areas of further Parts within the pattern represent to be transmitted to the display device. Docket RO 970 013 2 0 9 8 1 U / 13 7 3 2H9520 46. Device according to claim 45, characterized in that the pattern source includes a source of actual names associated with particular patterns, where the program also causes the actual names to be accepted as input data, with the following Elements provided are: a fifth routine which is responsive to a second type of control data for the actual names to compare with the second type of tax data and a sixth routine for the collation in the memory record of data relevant to the actual Names are indicative when the actual names match the second type of control data. Docket ro 970 013 2 0 9 8 16/1373 Blank page