Research On Segmentation Of Oil Contamination Region On Silicon Steel Surface Based On Superpixels

The steel plate surface defect detection system is a real-time online monitoring of the surface quality of steel plate, which has great significance for guiding the production and reviewing the quality. The accuracy and real-time performance of the segmenting defect area on the surface of the steel plate directly influences the effect and the efficiency of subsequent image processing. Due to large amounts of interference texture, low contrast between oil droplets and background region and generous time-consuming information of color image, traditional segmentation methods are unable to accurately segment the oil drop region on the surface of silicon steel images. To solving these problems, this thesis researches on oil droplet region segmentation for silicon steel plate surface color image.The main Achievements of this thesis are as follows:(1) after analyzing the characteristics of the surface of steel images with oil droplets, which are the texture interference, low contrast etc., this thesis proposes a segmentation method using SLIC superpixels segementation for low contrast regions. SLIC algorithm can not only avoid the interference of the surface texture of silicon steel plate, but also adsorbed effectively onto oil drop region edges which provides premise for the oil drop region segmentation. On this basis, this paper proposes a pixel polymerization method, with the LAB color space of the LABXY distance, namely color distance and spatial distance as similarity measure parameters of superpixels. The method is so simple, rapid that can meet two great importance of steel plate surface quality inspection system, that is the real-time and accuracy.(2) The amount of information of color image is 3 times larger than that of gray image of the same size, so that the color image has the details which the gray image does not, what’s more, some image processing needs this information to achieve segmentation. But due to the large amount of information, time-consuming processing limits the application of color image. The proposed approach is so simple that provides ideas and good application prospects for the segmentation of color image processing.(3) This thesis studies the irregular distribution of superpixels. Based on the rule of eight pixel neighborhood, this thesis presents a named L neighborhood (L as the default length of super pixel) method to measure the neighborhood super pixels. The L neighborhood metric adds appropriate tags to superpixels, furthermore, superpixels aggregation will be success.(4) This thesis studies the cause factors of threshold, then finds it largely influenced by the different brightness. To make this method has stronger adaptability, BP neural network to fit the adaptive threshold method is proposed. With the average values of brightness of the maximum and minimum arithmetic, global value and the average background values as input parameters, and the corresponding image threshold as output, adaptive threshold BP neural network fitting out on the oil-drop a good image.