Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (14): 201-208.DOI: 10.3778/j.issn.1002-8331.2203-0414

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Ceramic Tile Surface Defect Detection by Improved Faster RCNN

ZHAO Chu, DUAN Xianhua, SU Junkai   

  1. School of Computer Science, Jiangsu University of Science and Technology, Zhenjiang, Jiangsu 212100, China
  • Online:2023-07-15 Published:2023-07-15

改进Faster RCNN的瓷砖表面瑕疵检测研究

赵楚,段先华,苏俊楷   

  1. 江苏科技大学 计算机学院,江苏 镇江 212100

Abstract: Aiming at the problems of minimal defect target, large difference of defect shape, easy missing detection and low accuracy in ceramic tile surface defects, an improved ceramic tile surface defect detection algorithm based on Faster RCNN is proposed. Firstly, based on the original Faster RCNN, resnet101 is selected as the feature extraction network, and deformable convolution networks is introduced in the last three stages of resnet101 to adaptively learn the defect features. Secondly, the regional proposals network is optimized, and the anchor generation parameters are improved through the analysis of ceramic tile data set, so that the generated anchors are more consistent with the target scale and the positioning is more accurate. Finally, the loss function is optimized and Rank & Sort Loss is introduced to reduce the number of super parameters and improve the performance of the model, making it more robust to the class imbalance problem in training. Experimental results show that the average detection accuracy of the improved Faster RCNN is 76.3%, which is 17.9 percentage points higher than that of Faster RCNN. It can better detect small target defects and meet the requirements of ceramic tile surface defect detection.

Key words: target detection, ceramic tile surface defect, Faster RCNN, Rank &, Sort Loss, deformable convolution networks

摘要: 针对瓷砖表面瑕疵中存在极小瑕疵目标,瑕疵形态差异较大,易出现漏检、准确率低等问题,提出了一种改进Faster RCNN的瓷砖表面瑕疵检测算法。在Faster RCNN的特征提取网络resnet101的后三个阶段引入可变形卷积,自适应地学习瑕疵特征。优化区域建议网络,通过对瓷砖数据集的分析,改进锚点生成参数,使得生成的锚框更加契合目标尺度,定位更加准确。优化损失函数,引入Rank & Sort Loss,减少超参数数量,提高模型性能,使其对训练中类别不平衡问题更加鲁棒。实验结果表明,改进后算法的mAP为76.3%,比原始Faster RCNN算法提高了17.9个百分点,可以更好地检测小目标瑕疵,满足瓷砖表面瑕疵检测的要求。

关键词: 目标检测, 瓷砖表面瑕疵, Faster RCNN, Rank &, Sort Loss, 可变形卷积