带钢表面缺陷的RepVGG网络改进及其识别

RepVGG networks improvement of surface defects in strip steel and their identification

对带钢表面缺陷准确快速识别是带钢外观质量评价的一项重要内容,应用深度学习进行带钢表面缺陷识别是一个持续的热点。提出了一种基于改进RepVGG网络的带钢表面缺陷识别方法。首先为提高RepVGG网络在带钢表面缺陷识别中的稳定性和准确率,该方法引入了高效通道注意力(Efficient Channel Attention,ECA)网络,然后为了防止神经网络训练中神经元坏死,使得神经元参数得不到更新,应用了高斯误差线性激活函数。在测试集上,改进RepVGG网络对带钢表面缺陷的识别率达到了99.94%,而其运行速度并没有降低,单张图片的平均检测时间为5.4 ms。

Accurate and fast identification of strip steel surface defects is an important element of strip steel appearance quality evaluation,and the application of deep learning for strip steel surface defect identification is an ongoing hot topic.A strip steel surface defect identification method based on the improved RepVGG network was proposed.The method first introduces an Efficient Channel Attention(ECA)network in order to improve the stability and accuracy of the RepVGG network in strip steel surface defect recognition,and then a Gaussian error linear activation function was applied in order to prevent neuron necrosis during neural network training,which leaves the neuron parameters unupdated.On the test set,the improved RepVGG network achieved a recognition rate of 99.94%for strip steel surface defects without any reduction in its running speed,with a single image detection time of 5.4 ms.