Visual Inspection Of Product Quality Using Deep Learning

At present,traditional visual inspection algorithms and machine learning algorithms mainly used in product quality visual inspection greatly rely on manual feature extraction.It is difficult for these algorithms limited by feature engineering to meet the requirements of diverse defect detection tasks.The visual detection algorithm based on deep learning can extract sample features by itself and obtain high recognition accuracy and strong generalization ability.However,in the process of collecting raw data,there are often problems of samples insufficient and high cost of sample labeling,which is contrary to deep learning training.Therefore,this paper takes the real samples of commutator defects as the detection object,and proposes a multi-task defect detection framework based on deep learning.The details are as follows:Aiming at industrial defect detection,a deep learning-based product quality visual inspection framework with segmentation and decision is proposed.The fra mework includes a backbone network,a segmentation network and a decision network.The three have different divisions of labor.The backbone network extracts abstract feature expressions from the input image.The segmentation network identifies the specifi c location of the defect.The decision network predicts the class of the sample based the feature and the segmentation result.Different training methods are explored for the proposed defect inspection framework with segmentation and decision.First,through full supervision training,the feasibility and practicability of the framework on small data sets are verified.Then a weakly supervised objective function is proposed to solve the problem of high cost of sample labeling.The weakly supervised objective function only train with image-level labels,which can segment the location of the defect,and the classification performance is comparable to the full-supervised training result.Aiming at the problem of poor accuracy of the segmentation result of the we akly supervised method,different processing methods are explored.First,the graph-based segmentation method is introduced as a post-processing of the segmentation results.Then the graph segmentation method is combined with the deep learning training pro cess,and a new algorithm Deep Cut Plus is proposed to train the segmentation network.Deep Cut Plus obtains better segmentation accuracy than the correlation graph segmentation method without changing the network structure and increasing the detection time.Using the real commutator data set as an experimental sample,a comprehensive experiment was conducted to verify the advantages of our proposed defect inspection framework.Under supervised training conditions and weakly supervised training conditions,our detection framework has demonstrated excellent detection performance.Under the proposed weakly supervised objective function and Deep Cut Plus algorithm,excellent segmentation results are obtained with image-level labels,which is superior to the existing graph segmentation methods and weakly supervised segmentation results.In summary,this paper proposes a deep learning-based product quality visual inspection framework and corresponding training methods for industrial defects,and experiments verify its feasibility.