A Study Of Inspection Of Cell Surface Defects Based On Machine Vision

As an advanced automatic detection technology, Machine vision can effectively improve the production efficiency and industrial manufacturing level, visual inspection can be applied to the automatic identification of defects of product appearance. The thesis takes button cell as the research object, studies the defects detection methods of the surface appearance of both positive and negative poles.In order to achieve real-time detection of button cells, a new method to the implementation of automatic inspection system based on machine vision is introduced. This thesis analyzes the difficulties of imaging of button cell metal surface defects, studied the principle of imaging, and designs a low-cost hardware system of machine vision. Meanwhile PC software of visual inspection is developed. It consists of functions: system control, image processing and recognition, also includes a human-computer graphical interface which can display the real-time results.A set of machine vision algorithms is the core of defects inspection technology, the thesis focuses on the algorithms: preprocessing, positioning and characters correction, character regional positioning and segmentation, defects analysis and identification.A type of preprocessing algorithm of button cell image is studied, it includes median filtering, contrast limited adaptive histogram equalization, adaptive threshold binarization and morphological processing, and the defect area in image can be significantly enhanced.Positioning algorithm of button cell image is studied. As original image contains multiple circular targets, positioning algorithm can pre-process it and track contour, then calculate the center and diameter of each target using ellipse fitting method. After accurate positioning of targets, each ROI can be cut from original image.Characters correction algorithm of image of cell positive pole is studied. After similarity calculation of ROI with template image set including symbol tilt angle 1°to 360°on each image, the tilt angle value of ROI is obtained. Finally, ROI can complete tilt correction by image rotation. The thesis also studies several tilt correction improved algorithms and compares them.The method of characters removing of positive pole image is studied. First template matching method is used to find the character region, and then the experimental results of template–subtraction and regional segmentation are compared. Experiments show that, Due to limitations of template–subtraction method, regional segmentation method is used to segment the image of positive pole into background and character regions to remove characters.The method of global feature extraction of button cell image is studied, coefficient of image similarity can be taken as the feature value of positive and negative poles respectively, also coefficient of character region template matching in positive pole image can be taken as another feature value of positive pole image.The method of local feature extraction of button cell image is studied. Maximum of geometric feature value of all the connected components can be taken as the feature value of background region in a positive pole image, and average and standard deviation of pixels can be taken as the feature values of character region in a positive pole image. Maximum of geometric feature value of all the connected components can be taken as the feature value of a negative pole image.To combine global features and local features, a multi-classifier cascade algorithm is proposed, it can judge the image of positive and negative poles qualified or not respectively.