Research On Texture Surface Defect Inspection Algorithm Based On Deep Convolutional Autoencoder

Many products in industrial production are textured surfaces.However,due to the complex manufacturing process of products,it is easy to produce various defects in the production process.In order to ensure the quality of products,it is very important to detect the defects online in the production process.Machine vision detection technology is an important basis for intelligent manufacturing of products.Texture surface defect detection is a key theoretical method in machine vision detection.At present,the existing texture surface defect detection algorithms can only detect specific types of texture surfaces or specific defects.However,due to the variety of defects,the inspection algorithm is prone to overinspection or misinspection,which is difficult to apply to inspect various defects on various texture surfaces and unable to meet product quality testing standards.This paper proposes a texture surface defect inspection algorithm based on deep learning and level set,which can inspect multiple defects on various texture surfaces,and has been successfully applied in the procuction line of a new type of display device TFT-LCD.The detailed works are as follows:(1)In order to solve the problems such as low inspection accuracy of regular texture surface defects,difficulty in collecting defect samples and difficulty in manual labeling,this paper proposes a defect detection algorithm based on hidden space adversarial network(HSAN).The algorithm only needs to collect the defect-free images of the production line.The artificially defective images are generated through the defect-free images,and then the adversarial learning is carried out in the hidden space.Then the defect-free texture background images are reconstructed.Finally,the pixel-level segmentation of the defect is achieved by residuals.The experiments show that the average inspection accuracy of F1-measure in the ten data sets of Kylberg Texture Dataset,KTH-TIPS2,DAGM,MVTec and TFT-LCD reached 0.009,0.382,0.724,0.341,0.309,0.753,0.559,0.864,0.764 and 0.176,respectively,which improved the inspection accuracy on the regular texture background surface compared with the current mainstream advanced algorithms.(2)In order to solve the problem of low precision of defect inspection of the irregular texture surfaces,this paper proposes an anomaly-feature-editing-based adversarial network for texture defect visual inspection(AFEAN).This algorithm utilizes the method of central-constraint-based clustering method to detect abnormal features.Next,a novel global context feature editing module(GCFEM)is proposed to edit the detected anomaly features to normal features to restrain the reconstruction of defects and reconstruct the texture background.the average inspection accuracy of F1-measure in the ten data sets of Kylberg Texture Dataset,KTH-TIPS2,DAGM,MVTec and TFT-LCD reached0.841,0.624,0.734,0.852,0.527,0.760,0.735,0.880,0.867 and 0.513,which greatly improved the inspection accuracy on the irregular texture background surface compared with the current mainstream advanced algorithms.(3)In order to solve the problem that the detection algorithm based on deep learning is not accurate for irregular and low contrast defects,this paper proposes an accurate segmentation algorithm based on edge enhancedment level set(EELS).The algorithm utilizes Fourier transform to delete texture.Then the detection result of deep learning is used as the initial contour.On the basis of the regional level set,the edge enhancement energy term is added to enhance the low contrast edge information,and the precise contour of the defect is segmented by energy functional evolution.Compared with the mainstream level set method,the proposed EELS improves the segmentation accuracy of low contrast defects.In this paper,the proposed algorithms are applied on the semi-automatic optical inspection equipment.The 100 defect images with 1536×2560 pixels were collected for testing.The average Precision and Recall of HSAN algorithm were 0.807 and 0.786,respectively.The average Precision and Recall of AFEAN algorithm were 0.836 and 0.865,respectively.EELS algorithm takes the initial contour of AFEAN algorithm as an example to improve the average Precision and Recall to 0.937 and 0.898 respectively,which can meet the requirements of high precision segmentation of various defects.