CNN-ViT fusion with feature enhancement for defect detection of diamond tool segments

Existing deep learning-based industrial defect detection methods primarily focus on workpieces with regular textures and uniform imaging conditions, while insufficient attention has been paid to highly reflective alloy components with complex surface reflections and textural interference. Owing to the presence of diamond particles and metal powders, their surfaces exhibit strong specular reflections and random highlights, causing scratches, holes, edge defects, and other imperfections to be highly entangled with background noise and thus significantly increasing detection difficulty. To address this challenge, a cross-modal dynamic fusion framework, termed DCVNet, is proposed for defect detection on complex, highly reflective composite material surfaces.A local-global feature decoupling mechanism is constructed to separate defect-related information from reflection-induced background interference. A multi-stage defect-enhanced clustering algorithm is designed to achieve physical prior separation of background and defects. Furthermore, a progressive feature fusion module is introduced to realize deep cross-scale feature fusion between convolutional neural network (CNN) and vision transformer (ViT). A dedicated surface defect image dataset of diamond tool segments was constructed to train the model. Persuasive experiments were conducted by comparing DCVNet with GoogLeNet, ResNet50, ResNet101, ViT-L16, MobileNetV2, CRAD, PNI, SuperSimpleNet, and MAML models. The results demonstrate that the DCVNet model achieves a detection accuracy of 0.841 and a recall rate of 0.866, outperforming the comparison models. The proposed DCVNet model exhibits strong robustness and high detection performance for defects on complex, highly reflective composite material surfaces, providing an effective solution for industrial defect inspection scenarios.