一种基于改进Faster RCNN的金属材料工件表面缺陷检测与实现研究

目的 针对传统检测算法在工件表面缺陷检测上的局限性，以及检测精度不高、准确率较低、检测过程繁琐等问题，提出了一种基于改进RCNN的金属材料工件表面缺陷检测算法。方法 图像预处理过程中，运用了图像缺陷定位标注与图像数据的增强处理的方法。模型训练时为了避免某些分类数据不足，防止因数据集过小导致系统测试模型出现过拟合现象，使用了对原图像进行数据扩增处理。检测网络模型设计时，采用非极大值抑制算法对缺陷图像进行候选区域筛选，构建了区域建议网络，实现网络多层特征的复用和融合，在减少候选区域冗余的基础上提高系统的检测精度。引入多级ROI池化层结构设计算法，消除ROI池化取整而产生的系统偏差，实现高效并准确检测零件表面缺陷的目的。基于ROI-Align算法的原图位置坐标改进，利用双线性插值法获得原图的位置坐标，克服了基于最近邻插值法的ROI-Pooling设计算法带来的像素位置偏移和检测不匹配（misalignment）的问题。结果 设计的检测方法在测试集上，金属材料工件表面目标缺陷检测速度达22 帧/s，准确率达97.36%，召回率达 95.62%。结论 与传统的工件表面检测方法相比，改进的FasterRCNN方法对目标识别与定位处理具有较快的速度与较高的准确度，能在复杂场景条件下，提升工件表面缺陷的检测性能。

Surface Defect Detection and Realization of Metal Workpiece Based on Improved Faster RCNN

The work aims to propose a new algorithm in surface defect detection of metal workpiece based on an improved RCNN method in view of the limitation of traditional detection algorithm in the detection of workpiece surface defects, as well as the problems of low precision, low accuracy and tedious detection process. In the process of image pre-processing, the methods of image defect location and annotation and image data enhancement were used. While in model training, in order to avoid the shortage of some classification data and avoid the over fitting phenomenon of system test model caused by too small data set, the original image was processed through data amplification. In the design of detection network model, the non-maximum suppression algorithm was used to filter candidate regions of defect image so that a regional suggestion network was constructed, which realized the reuse and fusion of multi-layer network features and improved the detection precision of the system on the basis of reducing the redundancy of candidate. A multi-level ROI pool layer structure design algorithm was introduced to eliminate the system deviation caused by ROI pooling and rounding, which could effectively and accurately detect the surface defects of parts. The position coordinate of original drawing based on ROI-Align algorithm was improved and the position coordinate of original drawing was obtained by bilinear interpolation method, which overcame the problem of pixel position deviation and detection misalignment caused by ROI-Pooling design algorithm based on nearest neighbor interpolation method. The detection method proposed in this paper proved that in the test set, the detection speed of the target defects on the surface of metal workpiece was 22 fps, the accuracy rate was 97.36%, and the recall rate was 95.62%. Compared with the traditional workpiece surface detection method, the improved Faster RCNN method has faster speed and higher accuracy for target identification and positioning processing, which can improve the detection performance of workpiece surface defects under complex environment.