Research On Mobile Phone Screen Defect Detection Method Based On A Few Samples

With the rapid development of the Internet,mobile phones are everywhere.As one of the main parts of mobile phones,the quality of the mobile phone screens seriously affects the user's sense of experience.In the production process,quality inspection is a particularly important part.With the rapid development of computer vision in the industrial scene,defect detection based on computer vision has gradually become one of the research topics.However,in the production process of mobile phone screens,the mobile phone screen defect images are limited in both types and quantity,that directly affects the performance of big data-driven detection networks.Therefore,how to use a few samples to complete the defect detection becomes an emphasis of this thesis.In general,the research advances of defect detection methods are introduced in this thesis.On this basis,the relevant theoretical bases are summarized and analyzed.Then,combined with the characteristics of mobile phone screen defects,the defect detection methods based on a few samples are proposed including defect classification and defect segmentation.Finally,the two networks are evaluated on the mobile phone screen defect image dataset.Specific work is as follows:In this thesis,the Attention-Relation Network is proposed to classify the mobile phone screen defects.Based on Relation Network,the skip connection and dilated convolution are used to obtain the feature representation of defects of different sizes.In the feature metric module,the convolution network is used to learn the correlation between features,and the position attention and channel attention are introduced to capture the dependency between any two positions and any two channels respectively,to further emphasize the correlation between the same goals.Then the defect classification is completed by softmax.The experimental results show that the classification accuracy of the Attention-Relation Network is 0.9478 under the 5-way 1-shot training strategy,and0.9043 under the 5-way 5-shot setting.Compared with other classification networks,the classification performance of the Attention-Relation Network is better.In this thesis,A segmentation method called Co-Attention Segmentation Network(Co-ASNet)is proposed for the mobile phone screen defect segmentation.Combining with the few-shot segmentation network architecture,the Co-ASNet uses the feature extraction module to obtain the feature representation of the support images and query images,and uses the conditional branch to guide the segmentation process of the query image together with the feature information of the support images and its mask image.Besides,in order to further improve the segmentation effect,the co-attention method is applied to enhance the defect feature information interaction between the same defect target in the support image and query image,and then reinforce the defect feature representation.Compared with Similarity Guidance Network(SG-ONE),the Mean Intersection over Union(MIo U)value of the proposed segmentation network is increased by 11%,reaching 0.5771,which has a better segmentation effect.