基于改进Mask R-CNN的钢铁表面缺陷检测与分割

Segmentation of metal surface defect based on improved Mask R-CNN

摘要:

针对金属表面缺陷存在的问题,提出一种基于改进Mask R-CNN的表面缺陷检测与分割方法。使用先进的ConvNeXt-T替换ResNet-50以改进用于特征提取的骨干网络,在特征金字塔部分添加交错稀疏自注意力模块增强模型的全局建模能力,同时通过多级区域特征融合以加强模型的上下文信息表达能力。在钢铁表面缺陷数据集上开展了对比和消验验证,结果显示骨干网络改进的效果最明显,其mAP_bbox指标和mAP_mask指标分别提升了8.2%和6.3%,相较于对比方法,所提方法对钢铁表面缺陷的检测和分割精度最高,mAP_bbox指标和mAP_mask指标分别达到了0.690和0.662。

Abstract:

A surface defect detection and segmentation method based on improved Mask R-CNN is proposed to address the problems of metal surface defects. Replace ResNet-50 with advanced ConvNeXt-T to improve the backbone network for feature extraction, add interleaved sparse self attention modules in the feature pyramid section to enhance the global modeling ability of the model, and enhance the contextual information representation ability of the model through multi-level regional feature fusion. Comparison and validation were conducted on the dataset of steel surface defects, and the results showed that the improvement of the backbone network was the most significant, with an increase of 8.2% and 6.3% in the mAP-bbox and mAP_mask indicators, respectively. Compared with the comparison method, the proposed method has the highest detection and segmentation accuracy for steel surface defects, The mAP_bbox index and mAP_mask index reached 0.690 and 0.662, respectively.