Adversarial learning based intermediate feature refinement for semantic segmentation

Image semantic segmentation is a meaningful task that requires both accuracy and efficiency in computer vision. At present, most current deep learning based semantic segmentation methods needs extensive computational resources, and knowledge distillation may reduce such a computational burden due to its model compression ability. In this paper, different from previous knowledge distillation methods that directly transfer the knowledge of the teacher network to the student network, we propose a novel intermediate feature refinement method for semantic segmentation based on adversarial learning, which reduces the error and redundant information contained in the teacher network in the process of knowledge distillation, enhances the correct information contained in the teacher network and transfers it to the student network. Then we improve the conventional discriminator in adversarial learning to help the student network align more correct intermediate features in the teacher network. Our method can make the feature distribution of the student network closer to that of the teacher network, and finally improve the segmentation performance of the student network. Finally, we conducted experiments on three popular benchmarks to verify the effectiveness of our proposed method, including Pascal VOC, Cityscapes and CamVid. Compared with the competitive baseline, our proposed method can improve the performance of the student network by up to 1.43% (the mIOU increases from 67.14% to 68.57% on the Cityscapes val set).