Local Brightness Normalization for Image Classification and Object Detection Robust to Illumination Changes

Deep neural networks are typically trained on images taken under controlled illumination. Those networks work well on such images, but not on images that contain severe illumination changes which often occur in practice. To improve the robustness of networks to illumination changes, we propose to use local brightness normalization (LBN) as pre-processing of the input images and to train the network on those normalized images. The LBN can convert images to have similar appearances from various types of illumination changes. Then, we assume that input images of training and testing are more aligned by the LBN. Experimental comparisons of the image classification task and the object detection task show that the proposed LBN-based approach can improve the accuracy with images of uniform and non-uniform illumination changes.