Light Siamese Neural Network Architecture for Image Comparison

Deep neural networks typically require large training sets and powerful processing hardware, thus making their use difficult in training data or processing resource-limited systems. Siamese neural networks use few-shot learning to address this problem. However, their twins are typically implemented with standard multilayered perceptrons (MLP) or convolutional neural networks (CNN), which still require large training sets and powerful processing hardware. We propose a simplified Siamese model that uses a single Inception module for the twins instead of a full Inception CNN, for 20 times less parameters to learn and an easier implementation. Our validation experiments to assess whether the image of a molecular structure bears similarity with a reference shows 95.6% average accuracy for three different reference molecules, using training sets of just 25 randomly selected pairs of positive and negative examples for the twins. The training took 43 minutes to complete on an Intel I5-based PC station with no GPU acceleration.