关键词: 表面缺陷定位, 注意力机制, 特征金字塔, 语义分割

Abstract: In the field of steel production, surface defects pose a common challenge that can significantly degrade the quality of the final products. Surface defects in steel exhibit high background similarity and often involve defects of small size. Manual inspection and traditional machine learning methods for defect detection confront issues of insufficient accuracy and low efficiency due to these characteristics. Therefore, innovative techniques are needed to ensure accurate detection. The work of this study lies in the development of an advanced pixel-level detection algorithm called Multi-scale Feature Fusion Network for Steel Surface Defect Localization (MFF-Net) for steel surface defect detection. This algorithm is capable of identifying steel surface defects and accurately localizing these defects, contributing to the enhancement of quality control in the steel industry.
   Aiming at the problem that the classification accuracy still needs to be improved in steel surface defect detection and the localization of small size defects is not good, the proposed MFF-Net algorithm relies on the U-Net architecture, proving effective in small size target image segmentation tasks. By incorporating an attention mechanism, the algorithm enhances the network’s capacity to capture crucial features related to steel surface defects, thereby improving overall performance. To overcome the challenges posed by high background similarity and small size defects, a receptive field enhancement module is integrated into the network. This module facilitates the fusion of multi-scale defect features, contributing to improved accuracy in identifying subtle defects. A joint loss function is utilized during training to continuously adjust network parameters, cleverly addressing the issue of imbalanced defect categories and ensuring the robustness and balance of defect feature learning. Additionally, the algorithm employs data augmentation techniques to expand the dataset, actively enhancing the network’s generalization and overall robustness. The integrated approach implemented in the MFF-Net algorithm solidifies its effectiveness in recognizing and accurately locating steel surface defects in various scenarios. In summary, the MFF-Net algorithm addresses the challenges of category imbalance and insufficient data by incorporating attention mechanisms, receptive field enhancement modules, improved loss functions and data augmentation, thereby enhancing the accuracy of classification and localization of small-scale steel surface defects.
   The experimental results reveal that the proposed algorithm achieves an mIoU (mean Intersection over Union) of 84.53% on the NEU-Seg dataset, improving by 1.29% compared to existing state-of-the-art models. The algorithm demonstrates a high level of accuracy in identifying defect types, positions and shapes, meeting the requirements for steel surface defect detection in real-world scenarios.

Key words: surface defect location, attention mechanism, feature pyramid, semantic segmentation