Container-surface-guided dual-task learning for defect detection in complex port scenes

Abstract

Automatic container defect inspection in complex port environments is hindered by substantial scale variation and defect-like interference from non-container regions. This article presents CSG-YOLO, a container-surface-guided dual-task framework that jointly learns five-class defect detection and container-surface segmentation. Its scale-tuned feature fusion module adapts multi-level neck features to defects of different spatial extents, while mask-guided cross-task interaction bidirectionally exchanges the learned surface prior and defect cues between the task heads. Center-to-container consistency and edge-aware mask refinement losses respectively encourage defect-center coverage and surface-boundary consistency. On a single-terminal dataset of 4,000 distinct port images containing 4,480 defect instances, CSG-YOLO-N achieves mean scores of 58.1 average precision (AP), 93.0 AP50, and 65.0 AP75 across three independent training runs, with 16.3 GFLOPs. Compared with YOLOv12-N, it improves these metrics by 2.4, 3.2, and 4.1 points, respectively, while reducing mean overall missed detection rate from 6.4% to 5.7%, false positives per image from 0.19 to 0.11, and outside-region-of-interest (ROI) detection ratio from 1.8% to 0.2%. It also obtains a mean Surface intersection over union of 76.7%. These results support the effectiveness of explicit container-surface modeling for defect detection and suppression of spatially invalid detections on the evaluated dataset. Generalization to other port environments remains to be established.

The authors' abstract, as published at the source. Intelligence & Robotics, 2026 · DOI ↗