Dual-Path Attention Network for Lightweight Self-Supervised Monocular Depth Estimation

Self-supervised monocular depth estimation realizes training without depth labeling data by mining the geometric consistency of image sequences, which has important application value in fields such as autonomous driving. Traditional methods rely on complex CNN and Transformer hybrid architectures to balance local and global features, but face problems such as a large number of model parameters and low computational efficiency, which severely limit the deployment capability of edge devices. Although the existing lightweight methods reduce the number of parameters through techniques such as depth-separable convolution and channel compression, there are still have problems such as insufficient multi-scale feature fusion, limited interaction ability of global and local context information, and loss of details at the edge of the depth map. To solve these problems, we propose LM-DualNet, a novel architecture with dual-path attention enhancement. Specifically, the encoder integrates a Dynamic Local Context-Aware (DLCA) module for capturing fine-grained local structures, and a Dual-Axis Gated Attention (DAGA) module that constructs two parallel attention paths—spatial and channel—to jointly model positional dependencies and cross-channel correlations. In the decoder, we design a Multi-Scale Depth Enhancement (MSDE) module to refine edge regions and enhance depth continuity. Experiments on the KITTI dataset show that the absolute relative error and squared relative error of LM-DualNet have decreased to 0.106 and 0.731 respectively, and the accuracy has reached 88.8%, which is a good improvement compared with other state-of-the-art algorithms.