Height-aware attention for industrial defect detection: co-designing imaging and deep networks for rubber-metal...

Defect detection in rubber-metal bushings is difficult because many faults appear as small surface-height changes rather than intensity changes. Here, we present a height-information inspection method that converts 16-bit laser profile measurements into three-channel height-color images for detector input. Paired grayscale images are used only as a baseline for modality comparison. The model replaces the standard attention block in a real-time one-stage detector with a height-aware attention module that combines feature-group channel attention, multi-scale spatial attention, and adaptive feature fusion. This design lets the network emphasize local height transitions while it retains contextual surface information. On 11,000 bushing images, height-color input alone increased recall by 45.3% over grayscale input. With the proposed attention module, the detector reached 92.9% precision, 93.5% recall, and 96.7% mean average precision at 0.5 intersection over union. Compared with representative transformer, graph-based, and multimodal detectors, the method improved mean average precision by 4.9-18.8 percentage points and reduced computation by 40-82%. These results show that joint design of image representation and network architecture improves industrial bushing inspection.