CHC-DETR: An End-to-End Efficient Defect Detection Network

To improve the performance of defect detection for electric power fittings while maintaining detection efficiency, this paper proposes an innovative end-to-end defect detection model named CHC-DETR. First, the ConvNeXtv2 architecture is adopted as the backbone network, integrating GRN to optimize the model’s generalization capacity and stability. Second, a HiLo-Aware Feature Interaction module (HAFI) is designed to enhance computational efficiency and multi-scale feature extraction by separating high-frequency and low-frequency information processing. Finally, a CSP-Refined Wavelet Convolution Module (CWCM) is introduced to strengthen the encoder’s ability to extract complex defect features and leverage wavelet transform to enhance low-frequency features, improving the recognition of tiny or irregular defects. On the self-constructed CU-DET dataset, the proposed model reaches a mAP of 94.9% and an FPS of 56.8. Experimental results reveal that the proposed algorithm provides an productive and high-performance solution for industrial defect detection.