Study On The Key Technology Of Surface Defects Inspection On Highly Reflective Sphere Based On Vision

The surface quality of a steel ball, which is the key components of a bearing, directlyafect its precision, dynamic performance and service life. The requirements for the defecton the surface of steel balls are becoming increasingly strict that bring challenges for thetraditional manual inspection method, because the requirements for the quality of bearingare daily increased. Computer vision has the advantages of high precision, fast speed,non-contact and easy to automated,which is adapted to the surface defect detection ofsteel balls as a novel inspection technology. A series of research and discussion of defectinspection on surface of steel balls are carried out in this paper based on vision method.Furthermore, an automated machine on defect detection on the surface of steel balls aredeveloped and achieved. The main content and contributions are as follows:A new illumination of uniform scattered light was proposed and a light source wasdesigned for the defect image acquisition problem of the highly reflective shiny metal ballsurface. An optical model for the surface of shiny metal balls was built after a discussionof some illumination methods. The distribution of intensity is simulated to verify theillumination uniformity of the source. A difuse light source is designed based on thephysics analysis of light reflection.A completely new sphere surface unfolding method was presented for adequatelyspherical inspection. This approach can achieve unfolding using two CCD cameras whenthe rolling balls go under the cameras. It has only one-dimensional mechanical move-ment instead of traditionally two-dimensional motion, which avoids a complex structure,easy to wear and randomly unfolding wheel system. Experiment verified this method isefective. Moreover, an international invention patent was applied and PCT believes thatit’s novelty, inventiveness and practicality.Two-steps location method was presented for objects to find which object defectsbelong to in real-time image processing. Specifically, gray projection was served as askeleton location for circular objects; what’s more, improved canny algorithm based onsegmented arcs combined with least square method is used for accurate location. Difer-ence image method and Otsu segmentation were employed in extracting defect regions.We discussed the boundary that the measured part is qualified or not and proposed rigidjudge principle for metal spheres.21statistical features that combination of geometric and texture features were ex-tracted as efective characteristics for classification of the five types of defects. KNN classification algorithm was used to analyze the shape characteristics, the moment fea-tures, statistical texture features and texture features based on GLCM, which were as theclassification described characteristics. Finally, feature selection was carried out withPCA principal analysis.Artificial neural network (ANN) and support vector machine (SVM) were improvedto complete classifiers design for steel balls defects. We investigated algorithm to improvestandard BP network such as pre-processing method based on the nearest circle algorithmfor the original data, multi-dimensional geometric method to select the initial value ofthe network and method of variable step size with momentum term for training. A Cross-validation algorithm was used among the building of SVM model. Experiments show thatSVM with C-SVC model can get higher classification recognition rate than ANN and hasbetter generalization ability.