A self‐distillation object segmentation method via frequency domain knowledge augmentation

Abstract Most self‐distillation methods need complex auxiliary teacher structures and require lots of training samples in object segmentation task. To solve this challenging, a method via frequency domain knowledge augmentation is proposed. Firstly, an network which efficiently integrates multi‐level features constructed. Secondly, pixel‐wise virtual generation model proposed to drive the transferring through learning, so as improve its generalisation ability. Finally, adaptive augment data, utilise differentiable quantisation operator adjust learnable table dynamically. What's more, we reveal convolutional neural more inclined learn low‐frequency information during train. Experiments on five datasets show that can enhance performance effectively. The boosting our better than recent methods, average F β mIoU are increased by about 1.5% 3.6% compared with typical feature refinement method.