Lightweight Semantic Segmentation Based on Detail-Semantic Alignment in Large Receptive Field

Abstract

In the lightweight semantic segmentation network, a large receptive field and accurate feature fusion are crucial. However, the traditional dilated convolution is sparse for pixel category inference, and there is a feature mismatch when fusing different levels of features. To address these issues, we proposed a novel method called Detail-Semantic Alignment Network (DSANet), which integrates a Large-Kernel Aggregated Convolutional Block (LACB) and Detail Semantic Flow Alignment (DSFA) module into a unified framework. First, the LACB uses large-kernel depthwise convolutions to capture dense pixel dependencies. Then, the large kernel results are aggregated by $3 \times 3$ convolution to further expand the receptive field, and LACB through the parallel architecture design to construct dense pixel dependencies at a low computational cost. Second, the DSFA module uses a new alignment method: Gated Dual Spatial Flow Alignment Module (GD-SFAM). It encourages channel-wise fine-grained features in adjacent detail feature maps to align with their semantic categories. Finally, we verify the effectiveness of our proposed DSANet on three benchmark datasets: Cityscapes, CamVid, and BDD100 K. Experiments show that our method balances high accuracy and low computational cost effectively.