Research On Defect Detection Technology Of Phone Components Based On Multi-task Deep Convolution Network

In the process of industrial production,due to various factors such as processing technology,facility conditions and technical personnels,various defects will appear on the surface of the product.In order to reduce the impact of defective products on industrial production,a variety of defect detection methods have been applied to production with the development of economy and technology,and detection technology is also undergoing rapid changes.From the initial low detection efficiency and high labor cost manual detection,to the machine vision that semi-automatically completes defect recognition by manually setting and extracting features,and then to the use of this article,which can automatically complete the phone component surface feature extraction work,through deep learning technology,and can realize end-to-end intelligent classification,detection and recognition.In order to solve the problem of low efficiency of defect detection by traditional methods in industrial production,and the difficulty of detecting multi-scale and small-scale defects,this thesis studies the recognition technology of phone component surface defects and defect detection technology.The research content includes:(1)Aiming at the problem of lacking sufficient amount of phone component defect images that can be obtained in the industry,from the perspective of model and data,this article uses a pre-trained model that has been pre-trained on a large number of data sets to initialize the model parameters through transfer learning,and uses data augmentation to enrich and expand the data set when a large amount of new data is not available;For the problem of ineffective recognition ability of traditional models in recognising multi-scale and small-scale defects,this thesis uses the methods of multi-scale feature fusion,layer jump connection and hybrid pooling to enhance the generalization ability of the model,and carries out the comparison experiments between improved model and multiple traditional models.The recognition effect is verified,and a phone component defect recognition model is obtained.The result shows that the proposed classification method has an accuracy rate of 97.81% in the test set compared with the existing defect classification methods,which has a higher recognition accuracy.(2)Aiming at the problem of feature extraction of multi-scale defects,especially small-scale defects in the defect detection task,as the convolution depth progresses,the feature information is gradually reduced.This thesis integrates the astrous convolution into the defect detection backbone network,by expanding the receptive field,increasing the amount of information extracted from defect features,and carrys out comparison experiments before and after the improvement.The detection effect is verified,and a phone component defect detection model based on a multi-task deep convolutional network is obtained.The average precision rate on the test set is 76.7%.Experiment shows that the proposed detection method improves the model’s ability to detect multi-scale defects compared with the existing detection methods.(3)The thesis has realized the setup of a phone component defect data collection platform,has completed the research on the recognition technology of phone component defect and defect detection technology,and has enhanced the generalization performance of multi-scale defect recognition based on original model,improving the robustness of the models.It provides a reference for applying deep learning to defect detection areas and introducing it into real industrial producting sites,and has significant reference value for introducing deep learning into other fields.