基于卷积神经网络的带钢表面缺陷检测算法

Strip Surface Defect Detection Algorithm Based on Convolutional Neural Network

针对现有带钢表面缺陷检测方法准确率低、特征泛化性不强、参数多、识别速度慢等缺陷,基于卷积神经网络,采用DenseNet网络的密集连接算法解决梯度消失和梯度爆炸问题,堆叠式空洞卷积扩大卷积核感受野,深度可分离卷积减少网络参数量,提出一种用于带钢表面陷检测的深度神经网络模型Ds-DenseNet算法。以NEU带钢表面缺陷数据集为基础缺陷样本,加入正样本,并对其进行数据增强操作,创建AUG-NEU数据集,本算法在AUG-NEU数据集上的测试精度高达99.38%,参数量为117958,仅占DenseNet121和ResNet50参数量的1.7%和0.5%,识别速度高达1.3ms/frame,分别是DenseNet121、ResNet50识别速度的2.3倍和2倍,完全可以满足带钢生产线实时检测的需求。

In view of the defects of low accuracy,low generalization,high parameters and slow recognition speed of existing strip surface defect identification methods,this paper uses the dense connection algorithm of DenseNet network to solve gradient disappearance and gradient explosion problem based on convolutional neural network.Stacked dilated convolutions expands convolution kernel receptive field,depthwise separable convolution reduces network parameter quantity,and proposes a deep neural network model Ds-DenseNet algorithm for strip surface defect detection.Based on the NEU strip surface defect dataset,the positive samples are added,and the data enhancement operation is performed to create the AUG-NEU dataset.The accuracy of the algorithm on the AUG-NEU dataset is as high as 99.38%.The quantity is 117958,which only accounts for 1.7%and 0.5%of the parameters of DenseNet121 and Resnet50,and the recognition speed is up to 1.3ms/frame,which is 2.3 times and 2 times of the recognition speed of DenseNet121 and Resnet50 respectively,which can fully meet the requirements of real-time detection of strip production line.