Hybrid Loss Functions for Infrared and Visible Patch Matching using Siamese CNN

Image similarity between cross modal image pairs such as visual and thermal images is a challenging task which is yet to be comprehensively addressed by deep CNNs. The feature representations of images extracted from convolution layers must be generalized enough to differentiate between dissimilar and similar regions of diverse visual and thermal images. One of the training choices, loss functions, has a significant effect in performance of the trained CNN model. We demonstrate the usefulness of combined loss by applying it to find similarity between features learnt from convolution layers of a Siamese CNN model. The combination of loss functions fabricates the convolutional layered architecture/model to learn the various inherent spatial (or regional), correlated and intensity-based features of an image. Our experimentation shows that the performance of model both in terms of training error and in prediction depends on the choice of loss function and significantly improves in the case of hybrid loss function.