基于YOLO网络系统的材料缺陷目标检测方法研究

A Material Defect Object Detection Method Based on YOLO Network System

基于图像的传统材料缺陷目标检测技术存在检测精度低,检测速度慢等问题。卷积神经网络(CNN)的出现很大程度上改善了上述问题,但是大多数的CNN都是基于候选区域方法来对目标进行定位,这样虽然能够提高检测的精度,但是对于实时检测系统而言,受硬件条件限制,检测速度难以满足工业实时检测要求。针对这一问题,本文提出了一种基于YOLO网络系统的材料缺陷目标检测算法来提高检测速度,利用YOLO网络把整张图像作为输入,直接在输出层回归目标边界框的位置和其类别,不再需要候选区域生成步骤,但是这样会对精度有所损失,所以最后本文对YOLO网络系统进行优化,利用DenseNet网络的优点,结合神经网络前面特征层的信息,在不影响检测速度的同时,保证材料缺陷目标检测的精度。

Traditional material defect object detection techniques have some well-known problems,such as low detection accuracy and slow detection speed. The emergence of convolution neural networks( CNN) has greatly improved the above problems,most CNNs are based on region proposal methods to locate targets,which can improve the accuracy of detection,but for real-time detection systems,the detection speed is not up to standard. Aiming at this problem,this paper proposes a material defect detection algorithm based on YOLO network. The advantage of YOLO network is that the whole image is used as input,and the position of the target bounding box and its category are directly returned to the output layer,which means that the region proposal generation step is no longer needed. However this will damage the accuracy,so in the end,the YOLO network is optimized by utilizing the advantages of the DenseNet network,that is,the information of the front layer of the neural network is combined to improve the accuracy of the material defect detection while guaranteeing real-time detection speed.