基于改进YOLOv4的地板缺陷检测算法

Floor Defect Detection Algorithm Based on Improved YOLOv4

为解决现有地板表面纹理类缺陷、小目标缺陷检测效果差的问题,提出了一种基于改进-YOLOv4的地板表面缺陷检测方法。修改了YOLOv4的CSPDreknet53特征提取网络输出,增加了104×104的特征层输出,在保持原有的特征层的基础上,增加了对小目标检测的特征层,增加特征层的上采样融合和下采样融合,使特征提取的更加充分,减少信息的丢失,对小目标检测更为精准。对整理好的数据集进行改进后的K-means聚类分析,重新设置了先验框的数量和大小。最后使用改进型-YOLOv4训练出的模型进行mAP计算,实验结果表明,改进型-YOLOv4的mAP达到了89.07%,比改进前提升了5.77%,分类准确率达到了94.62%,能够准确且快速的识别出地板的表面缺陷。

In order to solve the problem of poor detection effect of existing floor surface texture defects and small target defects,a floor surface defect detection method based on an improved-YOLOv4 is proposed.the output of the CSPDreknet53 feature extraction network of YOLOv4 was modified,and the output of the 104x104 feature layer was increased.On the basis of maintaining the original feature layer,the feature layer for small target detection was added,and the up-sampling fusion and down-sampling of the feature layer were added.Sampling fusion makes the feature extraction more fully,reduces the loss of information,and detects small targets more accurately.The improved Kmeans clustering analysis was performed on the sorted data set,and the number and size of a priori boxes were reset.Finally,the model trained by the improved-YOLOv4 was used to calculate the mAP.The experimental results show that the mAP of the improved-YOLOv4 reaches 89.07%,which is 5.77%higher than before the improvement,and the classification accuracy rate reaches 94.62%,which can be accurate and fast.Identify the surface defects of the floor.