ADRNet-S*: Asymmetric depth registration network via contrastive knowledge distillation for RGB-D mirror segmentation

Abstract

Mirrors are frequently encountered in computer vision segmentation scenes and can considerably impact salient object detection (SOD) accuracy owing to their complex and variable appearances. In recent studies, several methods have achieved good performance. However, most of their structures are complex and heavy, while lightweight structures fall short in terms of accuracy. In this study, a novel asymmetric depth registration network student model trained with distilled knowledge (ADRNet-S*) is proposed to overcome these limitations. First, to filter redundant red–green–blue information from the image features, a depth maximal guidance module is introduced that orchestrates multimodal feature fusion. Next, a novel pixel-relation graph convolution module establishes global contextual relationships to further expand image understanding. Inspired by contrastive learning and knowledge distillation (KD), a contrastive KD framework is devised that enables the ADRNet-student (ADRNet-S) to significantly enhance its segmentation capability under the guidance of an experienced ADRNet teacher (ADRNet-T) to produce the ADRNet-S*. The proposed design compresses the number of parameters and floating-point operations from 25.93 M and 85.83 G for the ADRNet-T to 2.46 M and 7.32 G for the ADRNet-S, while maintaining 96.7 % accuracy. Finally, extensive experimentation demonstrates the robustness and effectiveness of the proposed method and its novel SOD framework.78.01