This repository is a comparative study of three most basic models in the field of Computer Vision, i.e., Neural Networks, Convolutional Neural Networks, and Residual Neural Networks.
These models were trained on training data of 50000 images and tested on data 10000 images available as built-in datasets in Keras.
These images were of the size 32x32x3 and they could be mutually exclusively classified into 10 classes. The labeled classes in the datasets are:
- Airplane
- Automobile
- Bird
- Cat
- Deer
- Dog
- Frog
- horse
- Ship
- Truck
The architecture of neural networks consists of 2 hidden layers of 784, and 100 units. A batch normalisation layer is added in between the 2nd and the 3rd hidden layer.
The Model has 2,492,278 parameters out of which 2,490,510 parameters are trainable and rest 1,768 parameters are non-trainable. Atfer training the model over 15 epochs in batch sizes of 128 (391 batches per epoch), it is able to achieve an accuracy of 48.42%.
The architecture of neural networks consists of 3 convolution blocks and a fully connected layer of 100 units.
A convolutional block consists of a convolutional layer followed by a max pooling layer. The number of channels after convolution blocks are 16, 32, and 64 respectively. A batch normalisation layer is added after every convolutional block and the fully connected layer.
The Model 127,942 parameters out of which 127,518 parameters are trainnable and rest 424 parameters are non-trainable. After training the model over 20 epochs in batch sizes of 128 (391 batches per epoch), it is able to achieve an accuracy of 72.899%.
The architecture of neural networks consists of a convolutional layer, 4 convolution blocks and a fully connected layer of 200 units.
The number of channels after convolution blocks are 16, 32, 63, and 128 respectively. A batch normalisation layer is added after every two convolutional block and the fully connected layer.
The Model 540,820 parameters out of which 539,926 parameters are trainnable and rest 894 parameters are non-trainable. After training the model over 20 epochs in batch sizes of 64 (782 batches per epoch), it is able to achieve an accuracy of 77.759%.
| Model | Accuracy | Loss |
|---|---|---|
| Neural Network Model | 48.42% | 1.488 |
| Convolutional Neural Network Model | 72.899% | 1.555 |
| Residual Neural Network Model | 77.759% | 1.289 |
By this comparative study I've observed that by using the concept of residual networks or skip connections, accuracy of a Computer Vision Model can be enhanced.
If you've any suggestions or feedbacks, please feel free to drop comments and make pull requests.