This repo aims at ternary classification of images using siamese neural network.
The dataset consists of images of three tpyes of coal-
Class 1 has 30 images.
Class 2 has 23 images.
Class 3 has 35 images.
original data- https://drive.google.com/open?id=15bkqBI1rsjN21Yx6jSk_97PqF4sqinNG
Since the images are very few in number, I used different augmentation techniques and in the final augmented dataset-
Class 1 has 390 images.
Class 2 has 299 images.
Class 3 has 465 images.
augmented data- https://drive.google.com/open?id=1eM4k49cKBuQw3GqmU4q4abSb5iyP4zGP
Thus there are a total of 1154 labelled images.
I tried using CNN and resnet for this classification problem and the results achieved were as follows-
CNN - Train accuracy 98%, Test accuracy 83%
resnet - Train accuracy 98%, Test accuracy 87%
Siamese Network-
Training pairs - I paired a single image with 10 different images. If the images were same I labelled the pair as 1 else I labelled it as 0. I did this for a total of 500 images and thus I had a total of 5000 training pairs. This is one of the major advantages of siamese neural network, we can generate a large number of trainig pairs using relatively smaller number of training images.
Base Network - I used CNN as the base network. The last layer used was sigmoid with 64 units. Thus the base network converts an image to embeddings (of size 64).
L1_distance was used to get the difference between the two embeddings. And finally a dense layer with softmax activation was used to get the output as 0 or 1 depending on whether the two images matched or not.
Testing - The remaing 654 images were tested by pairing them with the training images.
Accuracy-
Training accuracy - 98%
Testing accuracy - 90%
Thus, the siamese network gave a much better training accuracy as compared to CNN and resnet.