ASFM: a multi-scale feature-focused steel defect inspection model

Abstract Currently, in the steel surface defect detection process, challenges such as acquisition noise, lighting variations, and other factors significantly complicate the identification of small defects. To address these issues, a multi-scale feature focusing steel defect detection model, termed the Adaptive Shape Focusing Model (ASFM), is proposed. Built upon the YOLO11s framework, the ASFM incorporates several key innovations to enhance defect recognition capabilities. First, the Adaptive Downsampling (ADown) module replaces some of the convolutions in the backbone network, reducing the number of parameters and computational complexity while expanding the model’s receptive field to improve its ability to recognize defect information. Second, the Feature Focused Diffusion Pyramid Network (FDPN) is introduced to replace the necking network, enabling individual scale features to connect with contextual information and enhancing the network’s ability to aggregate and acquire feature information. Third, the Dynamic Upsampling (DySample) module is employed to mitigate the problem of missing feature information during the upsampling process. Fourth, the Pinwheel-shaped Convolution (PSConv) module is utilized as the downsampling module to expand the network’s receptive field and improve the extraction of small defects. Finally, the Mixed Aggregation Network with Faster Convolutional Gated Linear Unit (MANet_FasterCGLU) is integrated to enhance the model’s information gradient, thereby improving the identification and capture of edge defects. Experiments were conducted using the publicly available NEU-DET dataset from Northeastern University. The dataset was preprocessed with Gaussian filtering and noise enhancement. After preprocessing, the ASFM model was compared with the YOLO11s model in terms of Mean Average Precision (mAP), Precision (P), Recall (R), and mean Average Precision @0.5:0.95 (mAP@0.5:0.95). The results demonstrated improvements of 5.1%, 5.9%, 8.1%, and 9.1%, respectively, showcasing superior recognition accuracy compared to other state-of-the-art models. The ASFM model is capable of detecting steel defects in most industrial scenarios.