基于深度学习的瑕疵检测软件设计与实现

Design and implementation of flaw detection software based on Deep Learning

针对传统瑕疵检测软件对瑕疵识别定位准确率低的问题,设计一个基于改进DenseNet-SSD的瑕疵检测系统。在DenseNet卷积神经网络和SSD目标检测框的基础上,分别加入可变形卷积和Focal Loss机制,得到改进的DenseNet-SSD模型;将此模型应用到瑕疵检测系统进行瑕疵检测测试。应用结果表明,对比于VGG16-SSD和Faster RCNN,本方法的平均检测精度高达93.53%,具备更好的检测性能。对比于原始的DenseNet-SSD,本方法的mAP提高了7.78%,且本方法的IOU仅为0.718,各类瑕疵检测精度的标准差为2.513 9。由此说明,设计的软件具备精准的瑕疵定位能力,能够有效识别各类瑕疵。

Aiming at the low positioning accuracy of traditional defect detection software,a defect detection system based on improved DenseNet-SSD is designed.Based on the DenseNet convolutional neural network and SSD target detection box,deformable convolution and Focal Loss mechanism are added respectively to obtain the improved DenseNet-SSD model;This model is applied to the flaw detection system for flaw detection test.The application results show that the average detection ac-curacy is 93.53%over VGG16-SSD and Faster RCNN.Compared with the original DenseNet-SSD,the mAP of this method improves by 7.78%,and the IOU of this method is only 0.718,and the standard deviation of various defect detection accura-cy is 2.5139.This shows that the designed software has the ability of accurate defect positioning,and can effectively identify all kinds of defects.