基于特征金字塔的多尺度金属表面缺陷检测

Multi-Scale Metal Surface Defect Detection Based on Feature Pyramid

针对金属表面缺陷检测效率不高的问题,提出一种基于特征金字塔的多尺度缺陷检测方法(MSDD)。首先,构建特征金字塔分类网络模型(FPCN)作为特征提取器并进行分类预训练;其次,在FPCN后连接多尺度回归层并进行微调;最后,利用非极大值抑制将MSDD输出的4165个边框进行筛选得出最终检测结果。在光度立体成像数据集上进行实验,实验结果为该算法在个别类平均精确率(AP)达98%,各类别AP均值(mAP)达90%,召回率Recall平均为88.5%,单张图片检测用时约为13ms。这表明相比于现有多尺度算法SSD和YOLOv3,该算法对缺陷目标特征提取更加精确,同时提高了鲁棒性和检测速度。

Aiming at the low efficiency of metal surface defect detection,a multi-scale defect detection method(MSDD)based on feature pyramid is proposed.Firstly,the feature pyramid classification network model(FPCN)was constructed as the feature extractor and the classification pretraining was carried out.Then the multi-scale regression layer was connected and fine-tuned after FPCN.Finally,non-maximum suppression was used to screen the 4165 borders of MSDD output to obtain the final detection results.Experiments were carried out on the photometric stereoscopic imaging data set,and the experimental results showed that the average accuracy rate(AP)of individual classes reached 98%,the average accuracy rate(mAP)of all classes reached 90%,the average Recall rate was 88.5%,and the detection time of single image was about 13ms.This shows that compared with the existing multi-scale algorithm SSD and YOLOv3,this algorithm is more accurate in feature extraction of defect target,and improves robustness and detection speed.