Channel Distribution Distillation-Guided Advanced Lightweight Network for Referring Image Segmentation

Referring image segmentation (RIS) is a difficult task that covers complicated scenes by learning strong pixel-level feature representations. It has been demonstrated that knowledge distillation is an easy and efficient method for training picture segmentation algorithms. Previous methods rely on spatial knowledge distillation to recognize target objects or categories, and fail to capture the fine-grained features, resulting in enormous computation and model size and limiting their real-time processing capabilities on resource-constrained devices. To overcome this challenge, this paper proposes a Channel Distribution Distillation-guided Advanced Lightweight Network (CDD-ALNet) for referring im-age segmentation. Technologically, we propose a bidirectional relation inference network, which predict differentiable direction vector within an end-to-end framework. The channel-wise knowledge distillation mechanism is introduced to enable bidirectional interactions between linguistic and visual features, which strengthens the discriminative ability to bootstrap fine-grained knowledge between different modalities. In particular, we utilize the Kullback-Leibler (KL) divergence to compare the channel-wise probability maps of the teacher and student networks. This approach allows the student network to concentrate on the most significant regions, thereby improving its overall performance. The experimental results showcase the efficacy of the CDD-ALNet, underscoring quicker segmentation times and a more streamlined method for tackling referring image segmentation tasks.