Research Of Industrial Automatic Inspection Based On Optical Image Processing

Under the trend of miniaturization,portability and multifunction of intelligent electronic devices,the electronic components in the equipment tend to be miniaturized and dense,and the complexity and integration of the circuit are also increasing,which brings great challenges to the quality detection in the automatic production.As a fast,non-contact and efficient industrial detection method,automatic optical detection is widely used in the quality inspection of industrial production lines such as chip,LCD screen,printed circuit board,precision assembly and so on.Therefore,it is of great academic and practical value to study it.This paper mainly studies optical image acquisition and image processing algorithm in automatic optical inspection technology.Aiming at two detection projects,classification of integrated electronic circuit in TFT-LCD screen and detection of conductive particles in bridging circuit,this paper makes a thorough analysis and research on four aspects:image acquisition and preprocessing,low level feature extraction,high level information extraction,targe t recognition and classification.The main content of this paper are the following:1.Aiming at the integrated circuit images with complex shape components and random morphological defects,this paper classifies the images according to whether they have digital background.First,by using affine transformation,lightness adjustment,and noise introduction,230 integrated circuit images of 2048*2048 sizes are extended to 1840 local images of 136*136 sizes,including 1360 images with manual label.After that,the extended image set is preprocessed by averaging,normalizing,whitening and so on.As it reduces the impact of different light intensity,the correlation between pixels,and the redundancy of data,the accuracy of subsequent classification is increased by about 6%.After that,obtain the best feature extraction and classification network by setting comparative experiments under different structures and parameters.The experiment results show that,the classification accuracy of the image is approximately 98%,while the recall,F-score and Matthews correlation coefficient are about 96%,98%and 97%respectively,when the deep learning method of Stacked Denoising Auto-encoders network(SDAE)and Histogram of Oriented Gradients(HOG)algorithm are used for preliminary feature extraction and deep feature learning.2.For detecting the number,size and location of conductive particles with periodic texture background in the bridging circuit images,a high-accuracy detection method is proposed,which combines the gray value of pixels and the statistical information of gray scales,and solves the problem of particles connected or overlapped in the area with larger particle density.After analyzing the light source and luminance field,a high pass filter is constructed to suppress the non-uniform low frequency illumination field of the image.In order to reduce the interference of the noise and the texture background to the detection,to enhance the details of the particles,and also to significantly increase the contrast between the background and particle,the mean smoothing filter and convolution filter are applied to enhance images.Then,three methods are used to test the number,location and area of conductive particles,including algorithm of local threshold segmentation combined with morphological operation,SIFT feature points clustering method based on support vector machine,and K means clustering method based on Uniform Local Binary Pattern image.The above three methods are used to detect the conductive particles in the 98 bridge circuit images,the missed detection rates are about 20%,12%and 8%respectively,and the second methods have about 6%mistaken detection rates.