基于改进YOLOv3算法的钢板缺陷检测

Defect detection of steel plate based on improved YOLOv3 algorithm

针对目前钢板缺陷检测精度和速度的不足,提出了一种改进的YOLOv3检测算法。首先使用小波-中值滤波处理缺陷图像,清除图像里的噪声使图像更平滑。然后在原有网络中的密集连接网络(Darknet-53)上增加一个尺度输出增强算法对小目标缺陷的识别能力。最后为了增强算法模型的准确性对算法原有的损失函数进行优化,得到改进版的YOLOv3算法模型。改进的算法在测试集上的mAP值可以达到64.31,比原有的YOLOv3网络提高了7.9,结果表明了改进算法在钢板缺陷上具有较好的检测效果。

The steel industry is the supporting industry of social development. In order to improve the level of industrial automation and effectively detect the surface defects of steel plates, an improved YOLOv3(you only look once) detection algorithm was proposed. Firstly, wavelet-median filter is used to improve the image contrast. Then, a scale output is added on the Darknet-53 network to enhance the algorithm′s ability to recognize small target defects. Finally, in order to enhance the accuracy of the algorithm model, the original loss function of the algorithm is optimized and the improved YOLOv3 algorithm model is obtained. The mAP value of the improved network can reach 64.31 on the test set, which is 7.9 higher than that of the original YOLOv3 network, which has a better application prospect in plate surface defect detection.