Lightweight Semantic Segmentation Network Based on Linear Attention and Supervised Contrastive Learning

This paper proposes a lightweight semantic segmentation network tailored for medical images, integrating a linear attention mechanism (KernelLinear Attention) and an embedded supervised contrastive learning module. The linear attention module approximates global self-attention through kernel-based decomposition, effectively reducing the computational complexity from O(n2) to O(n) while preserving long-range context modeling. The supervised contrastive module enforces semantic consistency in intermediate feature spaces by constructing pixel-level positive and negative pairs based on label guidance, thereby enhancing boundary delineation and small-object discrimination. To evaluate the proposed method, we conduct experiments on the Kvasir-SEG dataset and benchmark against the standard U-Net model. Results show a performance improvement of +4.26% mIoU with minimal additional parameters and inference overhead, demonstrating the effectiveness and practicality of the proposed architecture. This study provides a promising path for balancing segmentation accuracy and efficiency, contributing to the development of high-performance, deployable models for medical images