DFRD-Net: dual-domain feature representation dynamic network for steel surface defect detection

Abstract To address the challenges of complex texture interference and the risk of small defect features being submerged by high-frequency noise in steel surface defect detection, this paper proposes a dynamic detection network with a fusion mechanism of frequency and spatial domains, called DFRD-Net. The network enhances defect feature extraction through the collaborative operation of multiple modules. Specifically, we designed a module called the dual-domain dynamic filter, which first extracts features through spatial domain convolution, then uses a dynamic frequency-domain high-pass filter to adaptively remove low-frequency interference components related to background textures, while preserving the high-frequency features of defects. Then, the frequency domain convolution enhancer promotes cross-channel interaction in the complex domain, preventing noise amplification during the frequency domain processing. The output high-frequency features are further focused on defect areas by combining multi-scale local convolutions and a channel attention mechanism. Additionally, in the image preprocessing stage, a Sobel feature extractor module constructed with a dynamic differentiable Sobel operator and zero-padding max pooling is used to enhance edge detail extraction under complex textures. In the feature extraction stage, a C2f-hybrid attention graph convolution module is introduced to accurately locate the regions of interest and prevent detail loss in deep networks. Finally, the experimental results show that DFRD-Net achieves an accuracy of 83.9% mAP50 on the NEU-DET dataset, with the number of parameters reduced to 2.8 M. Moreover, it achieves 62.5% mAP50 on the GC10-DET validation set, demonstrating superior performance and effectiveness compared to other state-of-the-art models.