Research And Application In Unsupervised Defect Detection Based On Reconstruction

The defect detection method based on deep learning has been widely used in industrial production.Compared with the traditional supervised defect detection method,the unsupervised defect detection method has the advantages of low cost,stronger versatility,short training time,etc.,and is more suitable for the needs of actual industrial production environment.However,current unsupervised defect detection methods have shortcomings in detection accuracy,visualization effect and robustness.It is a challenging problem to improve the detection accuracy when there is no or only a few labeled data sets.In addition,the visualization effect of unsupervised defect detection method needs to be further optimized in order to enable inspectors to understand the test results more intuitively.In terms of robustness,unsupervised defect detection methods need to be able to adapt to different production environments and defect types,and be robust to noise,interference,lighting changes and other factors to ensure stability and reliability in actual production.In this paper,based on the reconstruction idea,combined with the classical convolutional autoencoder network and the discriminant model in recent years,two methods of unsupervised defect detection with high accuracy,good visualization and strong robustness are improved.The improved method has higher accuracy and better detection effect,and has been tested in practical application scenarios.The experimental results show that the improved unsupervised defect detection method in this paper has good practical effect.The main research contents are as follows:(1)Aiming at the problems of different area sizes of abnormal defects and background interference,an unsupervised anomaly detection method based on residual self-coding machine is proposed to detect and segment abnormal defects on object surfaces.The proposed model is designed and the residual pool module is used to increase the reverse reconstruction ability of the traditional model for anomalies,which makes the watershed between normal and abnormal more obvious and solves the problem of incomplete segmentation of large area defects.Gaussian smoothing function is introduced in the abnormal scoring stage,which makes the model robust and reduces the interference of background on the model.In the simulation industry of MVTEC AD data set,the accuracy of image level detection reaches95.6%,pixel level detection accuracy reaches 96.5%,and region level detection accuracy reaches 91.7%.(2)Aiming at the problem that there is still a gap between the real picture and the reconstructed picture in the model with more high-frequency information,an anomaly detection method based on the air-frequency domain fusion model is proposed.The proposed method introduces frequency-domain loss into the reconstruction subnetwork to solve the problem that the abnormal reconstruction high level noise is difficult to deal with in the denoising coding machine,and also cleans the reconstruction embedding set.The frequency domain channel attention module is introduced into the segmented subnetwork to realize the information fusion between the spatial domain and the frequency domain,improve the information richness of the segmented network,and thus improve the segmentation effect of the anomaly graph.This method improves the anti-interference ability of the model against high-frequency information,enhances the robustness of the model,and further improves the quality of the reconstructed image.The accuracy of picture level and pixel level in MVTec AD data set are both 98.3%,which proves the effectiveness and superiority of the model.(3)A mechanical parts defect detection system is developed based on Py Qt,and the air-frequency fusion model algorithm proposed in this paper is applied to the industrial scene.The air-frequency domain fusion model proposed in Chapter 4 is built into the system as the core algorithm.Through demand analysis,appropriate industrial cameras,light sources,lenses,power supplies and other equipment are selected.Use QTdesign to make the program user interaction interface,implement the specific functions of each control by Python,and read the camera information by calling Ailled Vision industrial camera.The system has reasonable hardware structure,high security and reasonable shooting environment.The software system can not only input parts pictures in the form of data sets for off-line defect detection,but also realize real-time and online abnormal detection of hydraulic control valves by virtue of its hardware part.In addition,the system can realize the model training function,and with the My SQL database.According to the experimental results,all the functional modules of the system can run normally,and the system can realize the defect detection efficiently.