I tried to identify similar images for a given image using CNN. Especially, I used pre-trained resent50 because of a lack of time resources. The reason why I chose resent50 is that It doesn't hurt performance much, because of skip connections, even though resnet50 is one of the heavy models.
On top of the trained resnet with imagenet which does not include the original top, I stacked two more layers. One is for feature extraction named embedding. The next is a softmax layer for training new data. And I did not train resnet layers. I only train added two layers. Thus I set the trainable variable of resnet model to False. And the named embedding layer use ReLU as a activation function because ReLU is the most prominent one and the de facto standard one.
model = Sequential()
resnet = ResNet50(include_top=False, pooling='avg', weights='imagenet')
model.add(resnet)
model.add(Dense(32, activation='relu', name="embedding"))
model.add(Dense(NUM_CLASSES, activation='softmax'))
model.layers[0].trainable = FalseThere are lots of gradient descent optimization algorithms such as Stochastic gradient descent, Adam, RMSprop, etc. Some people say that adam is the best. But others say RMSprop is better than the others. However, from my experience, There is no big difference in performance. Thus I chose the Stochastic gradient descent.
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])I split all the data into train and validation. In general, train, validation and test dataset are required to estimate model performance or to compare other models
However, I don't care about measuring model performance among models. So I only split the dataset into train and validation.
cb_checkpointer = ModelCheckpoint(filepath=self.trained_weights_path, monitor='val_loss', save_best_only=True, mode='auto')
fit_history = model.fit_generator(
self.train_generator,
steps_per_epoch=100,
epochs=NUM_EPOCHS,
validation_data=self.validation_generator,
validation_steps=100,
callbacks=[cb_checkpointer]
)After training, I built a new model from the previous model to extract image embeddings
base_model = load_model(self.trained_weights_path)
intermediate_layer_model = Model(inputs=base_model.inputs, outputs=base_model.layers[1].output)And then, I extracted embeddings from all of the images
img = image.load_img(image_path, target_size=(IMAGE_RESIZE, IMAGE_RESIZE))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
intermediate_output = model.predict(x)Finally, I builded a inverted index; Key is file path and value is a sorted list by cosine similarities. The reason why I chose cosine distance instand of euclidean distance is that cosine distance is a better choice in high dimension space, because of curse of dimensionality.
for key in tqdm(self.image_embedding.keys()):
embedding = self.image_embedding[key]
for target in self.image_embedding.keys():
if key == target:
continue
target_embedding = self.image_embedding[target]
embedding = np.array(embedding).reshape(1, 32)
target_embedding = np.array(target_embedding).reshape(1, len(target_embedding))
similarities[target] = cosine_similarity(embedding, target_embedding)[0][0]
self.inverted_index[key] = sorted(similarities.items(), key=lambda x: x[1], reverse=True)[0:top_k]Computing the cosine similarity takes too long in order to calculate all the distance of images because its time complexity is big o of n squared. So I also implemented spotify’s annoy version to compute nearest neighbors. This Algorithm is one of the popular the Approximate Nearest Neighbors algorithms. Usually, the ANN Algorithms are much faster than general Nearest neighbor search because the time complexities is almost constant time.
self.annoyIndex = AnnoyIndex(EMBEDDING_SIZE, 'angular')
filenames = {}
for idx, key in tqdm(enumerate(self.image_embedding.keys())):
self.annoyIndex.add_item(idx, self.image_embedding[key])
filenames[idx] = [key, 0]
self.annoyIndex.build(-1)
self.inverted_index = {}
print(len(self.image_embedding))
for idx, key in tqdm(enumerate(self.image_embedding.keys())):
indexes = self.annoyIndex.get_nns_by_item(idx, top_k)
values = []
for target in indexes:
if idx == target:
continue
values.append(filenames[target])
self.inverted_index[key] = values$ python main.pydef main():
imageSimilarity = ImageSimilarity()
# download train data
imageSimilarity.download_file()
# split data into train and validation
imageSimilarity.split_data()
# train model
imageSimilarity.train()
# extract embeddings
imageSimilarity.create_all_vec()
# create inverted index with cosine similarities for searching
imageSimilarity.calac_knn(10)
# create inverted index with ANN for searching
imageSimilarity.calc_knn_annoy(10)
# check top 10 neighbors from inverted index
imageSimilarity.get_knn("./geological_similarity/marble/PDF9R.jpg", 10)
imageSimilarity.get_knn("./geological_similarity/gneiss/1OK58.jpg", 10)
imageSimilarity.get_knn("./geological_similarity/andesite/0JDL9.jpg", 10)
# get similar images from a unseen image
imageSimilarity.get_similar_image("./data/valid/quartzite/2P42M.jpg", 10)- main.py
- source code
- /working/best.hdfs5
- trained weights
- embeddings.json
- embeddings of all images
- inverted-index.json
- inverted index for searching
