As a popular research subject in the field of computer vision, knowledge distillation (KD) is widely used in semantic segmentation (SS). However, based on the learning paradigm of the teacher-student model, the poor quality of teacher network feature knowledge still hinders the development of KD technology. In this paper, we investigate the output features of the teacher-student network and propose a feature condensation-based KD network (FCKDNet), which reduces pseudo-knowledge transfer in the teacher-student network. First, combined with the pixel information entropy calculation rule, we design a feature condensation method to separate the foreground feature knowledge from the background noise of the teacher network outputs. Then, the obtained feature condensation matrix is applied to the original outputs of the teacher and student networks to improve the feature representation capability. In addition, after performing feature condensation on the teacher network, we propose a soft enhancement method of features based on spatial and channel dimensions to improve the dependency of pixels in the feature maps. Finally, we divide the outputs of the teacher network into spatial condensation features and channel condensation features and perform distillation loss calculation with the student network separately to assist the student network to converge faster. Extensive experiments on the public datasets Pascal VOC and Cityscapes demonstrate that our proposed method improves the baseline by 3.16% and 2.98% in terms of mAcc, and 2.03% and 2.30% in terms of mIoU, respectively, and has better segmentation performance and robustness than the mainstream methods.
Keywords: feature condensation; feature soft enhancement; knowledge distillation; prediction information entropy; semantic segmentation.