A Study And Application On Deep-learning-based Texture Surface Defects Vision Inspection Algorithms

Texture defect vision inspection is an important research issues in machine vision and industrial automation fileds,and is widely used in industrial manufacturing,quality inspection and other fields.A texture defect is a local area whose texture color or structure is different from the surrounding.After a textured surface image of the product is captured by an industrial camera,the defect region can be evaluated in the target texture image by the texture defect inspection algorithm.Existing texture inspection methods are poorly adaptable:a texture defect inspection method is suitable for only one material.This paper proposed a texture defect inspection methods based on deep learning framework,which can inspect many kinds of materials texture surface defects simultaneously.The specific research work is as follows:(1)Aiming at the problems of out of focus blur,noise,nonlinear distortion and nonlinear illumination in barcode recognition,this paper proposed an Adaptive Edge Detection and a Mapping Model(AEDM)algorithm.The edge position of bars and spaces in the barcode are initially obtained based on the gradient-direction and the block-average based scan line.The nonlinear distortion,noise and out of foucs blur is mathematically modeled.Then the edge position of bars and spaces are iteratively adjusted.The recognition rate of proposed AEDM method reachs 94.8%on the standard barcode image dataset created by Wachenfeld,and the recognition speed reaches 100ms on 640×480 pixels images.The accuracy and speed are better than the best methods like Orazio,Wachenfeld algorithm and ZXing,DataSymbol commercial decoding.system.(2)Aiming at the problems of different shapes,scales,low contrast and sample scarcity of texture surface defects,this paper proposed a Multi-Scale Feature-Clustering-based Fully Convolutional Autoencoder for texture surface defects inspection(MS-FCAE),which uses multiple Fully Convolutional Autoencoders(FCAE)at different scale to reconstruct texture background images.Feature clustering modules is used to improve the discrimination of features,which improves the reconstruction accuracy of the texture background.This paper establish eight datasets selected from Kylberg Texture Dataset,Kylberg Sintom Rotation,KTH-TIPS2,DAGM,Thin Film Transistor Liquid Crystal Display(TFT-LCD).The accuracy of MS-FCAE reaches 0.780,0.810,0.360,0.533,0.524,0.854,0.742,0.830.On 512×512 pixels images,the speed reaches 16.8ms,which is superior to the best methods like LCA(Lowpass Coefficients Analysis),PHOT(Phase Only Transform),TEXEMS(Texture Exemplars),ACAE(CAutoencoderonvolutional Autoencoder for Anormaly Detection),RCAE(Robust Convolutional),MSCDAE(Muti-Scale Convolutional Denosing Autoencoder)texture defect inspection algorithm in accuracy and speed.(3)Aiming at the over inepsection problem caused by the generate-model-based texture defect inspection algorithm,this paper proposed an edge-guide-based image inpainting for texture defect inspection algorithm(EGI).The partial-convolution-based edge inpainting network extracts the edges of the texture image,and predicts the missing edge regions simultaneously.Then the complete edge image is used as a priori information to guide the texture generation in the partial-convolution-based image completion network for high-precision textures background inpainting.This paper establish eight datasets selected from Kylberg Texture Dataset,Kylberg Sintorn Rotation,KTH-TIPS2,DAGM,TFT-LCD.The accuracy of EGI reaches 0.824,0.869,0.782,0.388,0.700,0.917,0.839,0.920,which significantly improves the precision of MS-FCAE resultsThis paper not only study on separate performance tests on each algorithm,but also study related application tests on the TFT-LCD automatic optical inspection equipment(AOI).In 1920×1080 pixels images,the inspection accuracy of MS-FCAE reaches 0.95,the speed reached 82ms,which meets the requirements of online inspection applications.The EGI accuracy reaches 0.97,which meeting the requirements of high-precision inspection applications.