Kernel-mask knowledge distillation for efficient and accurate arbitrary-shaped text detection

Abstract

Recently, segmentation-based approaches have been proposed to tackle arbitrary-shaped text detection. The trade-off between speed and accuracy is still a challenge that hinders its deployment in practical applications. Previous methods adopt complex pipelines to improve accuracy while ignoring inference speed. Moreover, the performance of most efficient scene text detectors often suffers from weak feature extraction when equipping lightweight networks. In this paper, we propose a novel distillation method for efficient and accurate arbitrary-shaped text detection, termed kernel-mask knowledge distillation. Our approach equips a low computational-cost visual transformer module (VTM) and a feature adaptation layer to make full use of feature-based and response-based knowledge in distillation. More specifically, first, the text features are obtained by aggregating the multi-level information extracted in the respective backbones of the teacher and student networks. Second, the text features are respectively sent to the VTM to enhance the feature representation ability. Then, we distill the feature-based and response-based kernel knowledge of the teacher network to obtain an efficient and accurate arbitrary-shaped text detection model. Extensive experiments on publicly available datasets demonstrate the state-of-the-art performance of our method. It is worth noting that our method can achieve a competitive F-measure of 86.92% at 34.5 FPS on Total-text. Code is available at https://github.com/giganticpower/KKDnet.