Research On Defect Detection Algorithms Based On Deep Learning

China is a big country in the manufacturing,with many manufacturing companies.China is a big country in the manufacturing and textile industries,with many manufacturing and manufacturing companies.For now,most manufacturers still use traditional machine vision methods(ie,by thresholding,filtering,color statistics,etc.)and even manually detecting defects.However,as China's production level continues to increase,the traditional machine vision processing method can not meet the needs of the manufacturing industry in terms of accuracy,and it is also difficult to solve the target detection problem of defects.At the same time,deep learning has flourished,and deep learning has achieved amazing results in some areas.So people begin to link defect detection with deep learning.Among the defect detection problems,the most basic problem is the defect classification problem.This thesis proposes three improved neural networks by studying three classical neural network models: VGGNet,Res Net,and Goog Le Net: VGGNet-ch,Res Net-ch,and Goog Le Net-ch.These three improved networks adjust the size of the input and output,change the number of convolution layers and the size of the convolution kernel,increase or delete the fully connected layer and the number of neurons in it,and finally construct a network more suitable for extracting the characteristics of chopsticks.And when applied to the problem of chopsticks defect classification,the three neural networks trained have a significant improvement in accuracy compared with the original network.At the same time,in order to reduce the frequency of artificial parameter adjustment in neural network construction and reduce the difficulty of solving the defect classification problem,two neural network structure search algorithms are proposed in this thesis: neural network depth self-growth algorithm and neural network structure search algorithm based on simple hill climbing.By using these two neural network structure search algorithms,the neural network can continuously increase the number of convolutional layers or downsampling layers in training,so that the depth and width of the neural network are wider.Finally,a neural network suitable for the current defect data set can be trained.And on the current data set,the neural network can achieve classification accuracy similar to that of the manually adjusted neural network.There are more than one defect classification problem in defect detection,and there are also defect target detection problems.In this thesis,we focus on two classic target detection algorithms: Faster-RCNN and SSD,and propose a new defect target detection model SSDch.SSD-ch combines SSD with VGGNet-ch proposed in this paper,which reduces the number of parameters in the neural network and makes the initial feature extraction network more suitable for extracting the characteristics of chopsticks.When applied to the target detection of chopsticks,SSD-ch can significantly improve the speed of the network while maintaining a high m AP.In addition,this thesis also addresses the problem of defect location marking for textured textiles to propose an improved denoising autoencoder algorithm and a defect target detection algorithm based on denoising autoencoder and convolutional neural network for the problem of defect location marking of textured textiles.These two algorithms are capable of restoring defective images by constructing a set of shallower,less-parameterized neural networks.Then,by performing the operations such as the difference and the threshold processing on the restored image and the original image,the pixel-level position of the defect in the image to be detected can be obtained more accurately.