An Efficient Bidirectional Weighted Global Feature Extraction Algorithm for Steel Surface Defects

Focus on the drawbacks of slow detection speed, and poor detection accuracy for small objects in existing strip steel surface defect detection methods, we propose an efficient bidirectional weighted global feature extraction algorithm for steel surface defects. Firstly, the backbone feature extraction network is consisted of BNCSP-TR(BottleNeckCSP with transformer) blocks to enhance global information and optimize the extraction ability of small object features. Secondly, the enhancement in the network's accuracy for detecting small-sized defects is achieved through the design of BiFPN-Light (lightweight bidirectional weighted feature pyramid network) for multi-scale feature fusion. Finally, a more efficient and concise decoupled detection head called Eff-DcHead is designed to solve the contradiction between classification tasks and regression tasks. Extensive experimental results illustrate that the mAP of our algorithm on NEU-DET reaches 82.4%, which is 4.1% higher than the original algorithm. Simultaneously, the algorithm achieves a detection speed of 92.1 FPS, catering to the demands of industrial environments.