YOLO11n-Glass: An Industrial Glass Defect-Detection Framework Integrating Dual-Illumination Line-Scan Imaging and Candidate-Slice Inference

Accurate and efficient detection of industrial glass defects poses persistent practical challenges. Small, low-contrast defects are difficult to distinguish, while background redundancy in high-resolution line-scan images increases computational cost. This study proposes a lightweight detection framework, YOLO11n-Glass, that combines dual-illumination imaging with an improved YOLO11n detector. Using time-division strobing, a single camera acquires coaxial and backlight images during one scan, and these are used as independent training samples. For the detection network, targeted structural optimizations are introduced to YOLO11n to address the characteristics of glass defects. Selected neck downsampling layers are replaced with ADown modules to reduce complexity while preserving local and contextual features. A Multi-Scale Direction-Aware Interaction (MDAI) module is introduced into C2PSA to enhance elongated, low-contrast defect features through directional spatial interaction and local–global channel recalibration. Shape-IoU is adopted in place of CIoU for scale- and shape-aware bounding box regression, while GA-SAHI filters background-dominated slices before inference. The proposed model achieves 0.956 [email protected], 0.935 precision, and 0.937 recall. These values exceed those of YOLO11n by 2.0, 1.8, and 2.5 percentage points, respectively. The parameter count decreases from 3.7 to 3.2 million, and computational cost decreases from 8.2 to 7.3 GFLOPS. On a representative 15,163 × 8750-pixel image, GA-SAHI reduces inferred slices from 209 to 166 and end-to-end processing time from 7.3 to 6.9 s while preserving detection performance. The framework improves glass defect detection using a compact model and fewer redundant slice evaluations.