GM-APD LiDAR-based hierarchical noise suppression and range reconstruction for directional sub-pixel line-shaped weak targets with deep learning semantic guidance

Reconstructing directional sub-pixel line-shaped (DSPLS) weak targets under low signal-to-background ratio (SBR) conditions remains a critical challenge for Geiger-mode avalanche photodiode (GM-APD) Light Detection and Ranging (LiDAR) imaging. Such targets, exemplified by power lines, typically occupy less than one pixel in width and may appear with arbitrary orientations, making their range reconstruction highly vulnerable to background noise. To address these challenges, this paper proposes a hierarchical noise suppression and range reconstruction method for DSPLS weak targets in array GM-APD LiDAR systems. The proposed method contains three progressive stages. A hybrid GM-APD semantic segmentation network, referred to as HGSNet, is developed to discriminate sub-pixel targets from background noise by jointly exploiting multi-scale and spatiotemporal features, providing high-confidence semantic target priors with a segmentation accuracy exceeding 91%. Building upon these priors, an adaptive signal-level noise suppression strategy is employed to stabilize line-shaped range estimation under extremely low SBR conditions, followed by a structure-oriented continuity enhancement scheme based on improved DBSCAN clustering and directional dilation to preserve the structural continuity and integrity of line-shaped targets. Experiments on real GM-APD LiDAR power line datasets demonstrate that the proposed method reliably reconstructs a line target with a diameter of 4 cm at a distance of 210.15 m with an SBR as low as 0.0227, with the effective target width spanning only 0.5 pixels (angular resolution of 0.028°). Quantitative results show pixel accuracy exceeding 88% and an improvement of approximately 10% in structural similarity compared with state-of-the-art methods, validating the effectiveness of the proposed hierarchical framework in preserving directional sub-pixel structural continuity for long-range, photon-limited LiDAR applications.