Wavelet Knowledge Distillation via Decoupled Target for Scene Text Detection

In this paper, we investigate the knowledge distillation strategy for training a compact student model for scene text detection, using a cumbersome teacher model that is too computational to apply on resource-constrained devices. We observed that the frequency domain information of the response map is different between the teacher and student models obviously, which can effectively guide the student model to learn more effective knowledge. Furtherly, we propose a wavelet knowledge distillation method via decoupled target for training accurate compact scene text detection networks. Specifically, we first use discrete wavelet transformation to decompose the probability map into different frequency bands which contain different characteristic components, transferring knowledge in the high-frequency band and low-frequency band respectively. In addition, we decouple the target to enhance the distillation effect of the corresponding region, by separating text and background regions through the ground truth mask. Extensive experiments demonstrate that our method consistently improves the F-measure of the student model and outperforms the other mainstream distillation methods.