CDS-YOLO: Industrial Surface Defect Detection Based on an Improved YOLO11 Framework

Abstract

Industrial surface defect detection is simultaneously constrained by weak-contrast small targets, strong textural backgrounds, and directional structures. Although existing real-time YOLO methods offer high speed, neck enhancements often lack scale-specific functional partitioning, edge information is easily entangled with semantic features, and strip-oriented contextual cues are rarely modeled explicitly. Using YOLO11 as the baseline, this paper introduces a content-adaptive soft decomposition and interaction module (CSDIM), a dual-stream edge residual module (DSER), and a strip-pooling cross-attention module (SPTCA) in the neck: CSDIM strengthens P3 fine-grained representations, while DSER and SPTCA sequentially perform gated edge injection and direction-aware refinement on P4; the detection loss remains unchanged by default to isolate structural contributions. On the GC10-DET test set, relative to YOLO11n, mAP50 / mAP50-95 improve from 0.658 / 0.331 to 0.695 / 0.353. The model retains nano-level complexity and real-time inference.