Research On Surface Defect Recognition Methods Of Wheel Hub Based On Convolutional Neural Network

As one of the leading industries of the national economy,the automobile industry has strongly driven the economic growth of our country,and wheel hubs are important driving parts of the automobile.Therefore,the quality of the wheel directly affects safety of the automobile,and restricts the brand quality and market competitiveness.As a result,an intelligent detection method is urgently needed because the means and methods of manual detection lead to a large number of false detection and missed detection.Industrial vision technology based on deep learning rapidly replaces traditional technology with features such as automatic feature acquisition,wide application range,strong robustness and high efficiency.However,in the production of automobile wheels,the recognition task is complex,such as the defects show complex inter-class similarity and intra-class diversity,and the number distribution of various defects is extremely unbalanced,which brings great challenges to the current recognition technologies base on deep learning,therefore,it puts forward higher requirements on the performance of the recognition algorithm.In this dissertation,theoretical research and experimental research about deep learning were conducted,through which to realize innovation on improve defect image recognition performance,so as to expand the deep learning-based recognition technology,and meet the engineering needs of wheel defect recognition,The main research works and achievements in the dissertation are as follows:1.The wheel hub defect(WHD)database was built aiming at the problem of lack of training data in practical recognition task.The images in the database came from domestic well-known wheel hub manufacturing enterprises,and quality samples were selected,labeled and augmented.Three subsets were defined : WHD-4,WHD-9 and WHD-12,which can meet the requirements of experimental tasks in different chapters.Different classification rules were set for the database based on different recognition tasks and methods to compare with the current mainstream models.2.In order to solve the problems of domain disparity,small object and multi-scale in multi-defect recognition task,the Faster R-CNN* model was proposed.Four improvements on the baseline model Faster R-CNN were completed: adding a small-size anchor in the Regional Proposal Network(RPN),implementing hard negative sample mining,using feature cascade and multi-scale training to improve the overall recognition rate of defects.The self-built dataset WHD-4 was used to complete the experiment,and the experimental results showed that the detection accuracy was significantly improved.The success of the proposed model verified that hyperparameters changing and using training techniques can improve the detection accuracy of CNN model.3.Aiming at the problem of balancing detection accuracy and detection speed in hub defect recognition task,a fast hub defect recognition model Single Shot Multi-Box Detector and Fusion+Reasoning(SSD(F+R))was built.Resnet-50 was used to replace VGG to meet the detection speed requirements.Feature fusion strategy was introduced to make each layer contain rich detail information and semantic information.In addition,the visual reasoning module was added to improve the recognition accuracy of stacking defects and fuzzy defects.Dataset WHD-4 was used to complete the experiment,the balance of detection accuracy and detection speed was improved obviously.The application of the model proves that the choice of base network and the embedding of function modules can improve the network performance.4.To solve the problem that the model performance drops sharply due to the uneven distribution of wheel hub defect data sets,A Meta Faster Region-Convolutional Neural Network(MFRC)was developed.The model combined the meta-learning network with the object detection network,which was equipped with a predictor reconstruction network.The meta-learner can guide the learner to update parameters quickly to adapt to the new class of recognition task.Meanwhile,predictor reconstruction network made the new class of defects attract attention.Dataset WHD-12 was used to complete the experiment,the performance of few-shot recognition was significantly improved.Experimental results show the effectiveness of the few-shot learning method based on meta-learning for uneven data set.5.In order to solve the problem of invalidation of traditional CNN model due to the lack of defect training samples,a three-step method(Deep Embedding Network and Matching,Adaption,Calibration,DEM+MAC)was applied.Firstly,structural information in the data set was made full use of;Secondly,domain adaptation was performed on the test samples of unseen classes.Finally,the classification score of seen defects in generalized zero-sample recognition was calibrated.Dataset WHD-9 was used to complete the experiment,the three-step operation alleviated the hubness problem,the projection domain shift problem and the recognition bias problem to some extent.Experimental results show that the application of generalized zero-shot recognition for wheel hub defects is effective.This dissertation is supported by the project of the National Natural Science Foundation of China: “Research on the visual autonomous identification,high precision positioning and compliance control method of intelligent assembly robot”,and carries out a series of studies on key issues related to “visual recognition”.This topic can be extended to different fields,so it lays a foundation for the future research on deep learning-based correlation recognition.