基于轻量化深度学习网络的工业环境小目标缺陷检测

工业环境下表面缺陷检测是质量管理的重要一环,具有重要的研究价值.通用检测网络(如YOLOv4)已被证实在多种数据集检测方面是有效的,但是在工业环境的缺陷检测仍需要解决两个问题:一是缺陷实例在表面占比过小,属于典型的小目标检测问题;二是通用检测网络结构复杂,很难部署在移动设备上.针对上述问题,提出一种基于轻量化深度学习网络的工业环境小目标缺陷检测方法.应用GhostNet替代YOLOv4主干特征提取网络,提高网络特征提取能力及降低算法复杂度,并通过改进式PANet结构增加YOLO预测头中高维特征图比例以实现更好的性能.以发动机金属表面缺陷检测为例进行实验分析,结果表明该模型在检测精度(mAP)提升5.83%的同时将网络模型参数量降低83.5%,检测速度提升2倍,同时满足缺陷检测的精度和实时性要求.

Small-scale defect detection in industrial environment based on lightweight deep learning network

Automated surface defect inspection in industrial environments is an important aspect of quality management and is of significant research value. Generic detection networks, such as YOLOv4, have proven to be effective in the detection of a wide range of datasets. However, defect detection in industrial environments still needs to address two issues: one is that the percentage of defect instances on the inspected surface is too small, which is a typical small-scale object detection problem; the other is that the structure of generic detection networks is complex and difficult to deploy on mobile devices. To address these problems, this paper proposes a small-scale defect detection method in the industrial environment based on the lightweight deep learning network. Firstly, we replace the YOLOv4 backbone feature extraction network with the GhostNet to improve the feature extraction capability and reduce the complexity of the algorithm. Secondly, the proportion of high-dimensional feature maps in the YOLO head is increased by the improved PANet structure to achieve better performance. The experimental results show that the model can improve the detection accuracy(mAP) by 5.83% while reducing the number of network parameters by 83.5% and improving the detection speed by 2 times, which meets the requirements of accurate and real-time detection.