Dectection Of Surface Defects On Steel Balls Using Machine Vision

The quality of steel ball's surface is very important for precision, sports performance, and life-span of roll bearing, but manual method of surface defect detection and recognition on steel ball which has large workload and poor reliability is widely used in domestic manufactories. As a result, it is necessary that an efficient, low-cost technology is appliyed to detect the appearance of the ball. Machine vision is a new technology advanced along with the development of computer technology, and it is applied into automated surface defects detection on steel balls in this paper. According to the surface defect image's traits, the paper abstracts the defects feature from the steel ball surface defect image, and makes classification of them according to their effective features, to achieve the ball automatically detect defects on the surface.The defects of steel balls'surface form different types of diversity. The basic framework of the processing and recognition is designed according to the characteristics of familiar surface defect and hardware platform is established. Aiming at imaging characteristics of the metal steel ball with the strong reflection surface, the illuminate system and lighting are designed in order to solve the problem that the light can not uniform irradiation and get clear images of the ball's surface in this paper.The foremost vision detection algorithm towards surface defect image is focused in this paper. Traditional smoothness filters include median filter, Gauss filter, and anisotropic diffusion filter and so on. On basis of comparing each traditional smoothness filters, the paper designs an assembled filter combined median filter and eight to anisotropic diffusion filter. This filter is effective in eliminating the Gaussian noise and impulse noise and protects the edge. And then with a threshold of the gradient operator is applied to image sharpening processing, the iterative algorithm for the best threshold is adopted to achieve the binary division. Afterwards, recursive connected component labeling algorithm, simply connected domain and multi-connected domain contour tracking technology are applied to count the defects regional chain code. Geometry perimeters of every connected region are calculated for recognition. Geometry features which always reflect surface defects shape features include perimeter, area, center of gravity, compactness, rectangularity and eolngatedness. Hu moment and the shape of the moment features are also extracted which have invariance at the time of the region changing in the size and angle. According to the experimental, compactness, rectangularity and eolngatedness and the shape of the moment foregoing 4 features are selected out to identify defects on the surface as an effective classification.The research on the ball surface defect classification is based on the design of BP neural network classifiers to meet the surface defects identification requirements. Experiments prove that BP neural network to identify defects on the surface than the template matching method has a higher accuracy rate, and the average recognition rate reached 87.5 percent.The paper also analyzes the possible causes of making mistake.After a series of testing to the system, the conclusion is that the system can offer effective detection because of having the characteristics of high-speed, efficiency and good stability and have a great potential to popularize it.