Research And Implementation Of Surface Defect Detection Technology Based On MSSD Network

Automatic detection of surface scratch defects has a great impact on improving the efficiency of industrial automation production and improving product quality,and the research method based on machine vision has strong flexibility and efficiency in the field of defect detection,and has become a process in industrial production and manufacturing.the part that can not be lost.The application of surface defect detection technology can enable technicians to find existing problems in time and make improvements to reduce economic losses.The detection system deployed in each production link plays an important role in reducing production costs,improving production quality and production efficiency.To this end,this article takes the product surface defects as the research object.Using the advantages of deep learning model self-learning and feature self-extraction,research on defect learning detection algorithms based on deep learning is carried out,focusing on solving surface defect detection problems.The specific work content is as follows:1.This paper compares different target detection algorithms,and finally chooses SSD target detection algorithm as the research basis.A defect target detection method suitable for the interference of complex texture background information is proposed.The improved SSD algorithm meets the detection performance requirements of small targets such as scratches in the production of industrial products by fusing feature information of different scales with local receptive fields and designing the initial ratio of the default frame to achieve surface defect feature extraction and recognition.It is very robust to defects of different sizes,types and forms.In order to solve the problem that the embedded platform has poor computing power and cannot perform a lot of convolution calculations,the lightweight mobile convolution network MobileNet is used as a frontend feature extractor to replace the VGG16 network to obtain a more efficient defect detection network.By training and testing the DAGM 2007 data set,the experimental results show that the average accuracy and detection speed are better than the original detector.2.Aiming at the problem of high automation level in product manufacturing and production,low production rate of defective products,and failure to collect defect sample data sets.A defect image generation algorithm based on generating an adversarial network is designed to expand the sample data set.In order to reduce the difficulty of network training and learning,a residual module is added to the traditional adversarial network.The model only learns residual feature maps,which can improve the generation of defective images.quality.In order to grasp the size,shape and location of defects and increase the diversity of samples,this paper takes the mask binary map of the defect location as the input of the generation network,and controls the defect generation area by controlling the position and shape of the mask area.In order to make the fusion effect of the defect area and the background area better,this paper divides the discriminator model into a local discriminator and a global discriminator,focusing on the local and overall image,respectively,to make the generated defect samples more natural and better quality.3.There are many frameworks for deep learning target detection,and there are a variety of implementation methods,and there is no unified standard.The code quality is uneven,in order to build the network,you need to repeatedly recreate the framework code under the convolutional neural network.For industrial production,the significance of defect detection technology lies in the optimization of the accuracy and detection speed of specific data scenarios in the differentiated field.In order to deal with the above problems,this article builds a defect target detection system based on the Tensor Flow framework.Using the image classification library Slim tool under the Google open source project,it can be used with various convolutional neural networks,such as: VGG16,Inception and other models to serve as feature extraction.Convert the sample data set to TFRecord format.Then configure relevant information in pipline.config,use the API to call various networks for training and test data,and finally convert the well-trained model to a.pb file to deploy to the desired location,making the target detection model easy to build,train,and deploy.