Linear micro-damage defect detection network: integrating similarity mapping and dual dynamic mechanisms

Despite notable progress in surface defect detection for industrial metals in recent years, identifying linear defects (e.g., scratches and weld seams) remains a significant challenge. This stems from the scarcity of relevant data and the unique morphological characteristics of linear defects, which are often obscured by background noise. To address these issues, this study integrates linear defect data from five publicly available datasets—NEU-DET, GC10-DET, APSPC, Kolektor, and Severstal—and augments the samples to increase diversity. On this basis, we propose a novel detection algorithm, termed SDD-Net, incorporating three key innovations: (1) a Similarity Map Attention (SMA) module that enhances critical feature representation by computing similarity between input and template features; (2) a Dysample module that dynamically optimizes the spatial distribution of sampling points, significantly improving linear edge representation; and (3) a Deformable Convolutional Network v2 (DCNv2) module that dynamically adjusts convolutional sampling positions, enabling the network to adapt to diverse shapes, scales, and geometric deformations. Compared to 40 state-of-the-art algorithms, SDD-Net achieves superior parameter efficiency and computational complexity, with the highest detection accuracy of 78.0%. Moreover, it exhibits excellent generalization performance on the NEU-DET, GC10-DET, and APSPC datasets.