ASV-Net: a lightweight network for high-precision steel surface defect detection

To address the prevalent issues of limited accuracy and excessive parameter sizes in the current steel surface defect detection methods, we propose a high-precision lightweight network named ASV-Net. First, we introduce the adaptive hybrid attention (AHA) module, which enhances traditional channel and spatial feature capture mechanisms by providing a more refined and adaptive focus on critical information. The integration of AHA into the backbone of ASV-Net significantly improves feature extraction efficiency and accuracy. Second, the design of the slim depth attention convolution module optimizes the trade-off between lightweight architecture and high-precision detection. It is seamlessly integrated into SDAFlow-Detect to achieve more efficient feature extraction and information fusion. Finally, the VoV-GSCSP-GS3 module is incorporated into the neck part of ASV-Net, in which multiple GSConv convolutions are used to improve and redesign the original VoV-GSCSP structure. This modification effectively reduces redundant computational operations, resulting in enhanced computational efficiency and network lightweighting while maintaining high detection precision. Extensive experiments conducted on the GC10-DET dataset demonstrate that ASV-Net achieves a remarkable 4.0% improvement in average precision compared with the baseline model, along with a 33.98% reduction in parameters and a 41.98% reduction in giga floating-point operations per second. When evaluated against 19 state-of-the-art algorithms, ASV-Net exhibits outstanding performance and demonstrates robust generalization capabilities across the NEU-DET and APSPC datasets.