Dense-Cascade Neural Network for Thermal and Visible Image Registration

Abstract

The cross-modality image registration task is becoming more and more important in the field of image fusion, thermal and visible image fusion can reduce the impact of environmental factors while maintaining the image’s texture and detail. However, the common registration methods always have the problem of overfitting and poor generalization when facing large modality span task. To register the thermal and visible images, a simple yet effective registration model called Dense-Cascade network is presented in this paper. Our network uses two independent branches to extract the cross-modality features respectively. In order to further reduce the regression error, we use a cascade structure by stacking the networks and using STN layer to achieve gradient backpropagation. Finally, to validate the robustness of our model, we add gaussian noise severity from 0-4 on thermal images to test the performance of our models, and the results show that the registration performances of our models are better than others.