Research On Surface Defects Detection And Recognition Of Touch-screen Images Using Sparse Features

With the development of information technology and flat panel display technology, theglobal market demand of touch-screen increased as the rate of47%. China is one of maincountries for the manufacture of cell phone parts. Scale and stable manufacture abilitybecomes the core competitiveness of the enterprises in China. Also, the quality requirementsof touch-screen become stricter since the update of cell phone functions and appearance.Defects inspection of touch-screen is one of the key issues for quality control in themanufacture process. Up to now, in some of the enterprises of China, defect inspection oftouch-screen still depends on manual inspection. It not only wastes human resources, but theinspection accuracy also relies on the experiences of the inspectors significantly. In this thesis,the structure and related manufactural processes are analyzed to understand why the defectsoccur during the manufactural processes. Based on the structural properties of different layersof touch-screen, image processing approaches are proposed for touch-screen surface defectsinspection, segmentation and recognition under different camera resolutions and variedilluminations. The main innovations of this thesis are as follows:(1) In order to detect and recognize the defect types of cell phone cover glasses, a defectdetection and recognition algorithm based on Princinple Components Analysis(PCA) isproposed. In this algorithm, firstly the cover glass images are pre-processed, which includesthe contrast enhancement and binaryzation, connected component labeling and de-noising.Then the binary images after pre-processing are recognized by PCA algorithm. In the PCAalgorithm,“Eigen defect” of each image is extracted as the natural feature for defectrecognition. The “Eigen defects” not only can reduce the image demention, but also they canspan a subspace—“defect space”. All possible images can project into this subspace and theprojection results can be compared to achieve the detection of defects and recognition ofdefect types. Finally, boundary following algorithm is utilized for edge extraction.Experimental results showed that under the condition that when there are90images in thetraining set, the defect type recognition ratio is over90%. For defect detection, the False Negative Rate and False Positive Rate are12%and6%respectively for a test set with50defect-free and50defective cover glass images.(2) A defect inspection algorithm based on the sparse representation for the touch-screenimages is proposed. In this algorithm, an atom candidate set is constructed by using the defectfree images. Then an optimal set of atoms was selected from the candidate set to form aredundancy dictionary. According to thel1-minimization algorithm, the coefficients for linearrepresentation of a testing image under the redundancy dictionary can be obtained. Thus, thedefect detection problem can be transferred to the problem that if an image can be sparselyrepresented under the redundancy dictionary or not. For the optimal selection of the atomsfrom the candidate set and the linear representation of the testing image, a vector similaritybased orthogonal matching pursuit algorithm is proposed. Also, a sparsity ratio function isproposed as the sparsity measurement of the linear representation coefficients of the testimage. Under various illumination conditions, our proposed approach can efficiently andquickly detect the touch-screen defects for low resolution images and different defect types.In our experiments with100defect-free and100defective images, False Negative Rate andFalse Positive Rate are7%and11%respectively.(3) A defect extraction algorithm for structural texture products based on the low-rankmatrix reconstruction is proposed. According to the property that there are periodical,repetitive horizontal and vertical lines in the touch-screen images, the background of thetouch-screen image can be treated as a low-rank matrix and the foreground of the image canbe modeled as a sparse matrix. Thus, the defect area extraction problem can be boiled down toa low-rank matrix reconstruction problem and it can be solved by convex optimizationalgorithms. This achieves a quick and accuracy extraction of the defect area of thetouch-screen images. In addition, this algorithm can be extended to the surface defectextraction of any industrial products with low-rank structural textures. Experimental resultsshowed that our algorithm obtained good performance for different defective product imageswith structural textures.In this thesis, the sparse property of the touch-screen images was sufficiently explored and utilized for defect inspection, recognition and extraction. For the low resolution industrialproduct images with varied illumination, defect types and manufacture processes, PCA,low-rank matrix reconstruction and redundancy dictionary algorithms are improved andemployed for the defect inspection. Experimental results showed that our proposed defectinspection and recognition algorithms are robust and fast. Also, they achieved goodperformances for different industrial product surface images. These provide efficientalgorithms and references of surface defect automatic inspection for real industrialmanufacture processes.