Deep Residual Networks for Image Recognition

Abstract

Training deep neural networks is more challenging. We introduce a residual learning framework to simplify the training process for networks that are significantly deeper than those previously used. We specifically redesign the layers to learn residual functions based on the layer inputs, rather than learning functions without any reference. We present thorough empirical evidence indicating that residual networks are easier to optimize and can achieve higher accuracy with significantly increased depth. On the ImageNet dataset, we tested residual networks up to 152 layers deep, which are 8 times deeper than VGG networks but with lower complexity. An ensemble of these networks achieved a 3.57% error rate on the ImageNet test set, securing 1st place in the ILSVRC 2015 classification task. We also analyze networks with 100 and 1000 layers on CIFAR-10. The depth of representations is crucial for various visual recognition tasks. By utilizing extremely deep representations, we achieved a 28% relative improvement on the COCO object detection dataset. Deep residual networks were the basis of our submissions to the ILSVRC and COCO 2015 competitions, where we secured 1st place in ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation tasks.