Research On Online Self-Learning Vision Detection System Of Contact Appearance

As an important industrial electrical component, contact component decides electric equipments'life. The surface appearance features of contact component play an important role in contact performance. The detection methods depending on manual stay low level, using computer vision technique to replace human vision in quality detection has great practical value. Aimed to different varieties of contact appearance features, the online vision detection system achieving self-learning function is researched in this thesis.Firstly, the structure and self-learning process of detection system are designed through the detection requirements analysis. The function parts and their selection methods are discussed in detail. The function modules and workflow of software system are designed.In image processing, according to real-time and accuracy requirements, the Gaussian filter template is studied as a quick and efficient image pre-processing method. An adaptive threshold segmentation method based on improved minimum point algorithm and multi-graph median algorithm is proposed. And the 5 major categories of image feature library is built, which includes 45 features.Through the analysis of major factors that influence the multi-category pattern recognition, the feature pre-processing method and optimization method based on canonical variables are put forward. To realize multi-category feature selection, a two-category floating search and multi-category backward selection algorithm is designed. Combined these algorithms with SVM multi-level binary tree classification strategy, the system realizes the multi-category quick self-learning and classification detection.Based on the work mentioned above, the experimental platform is developed. The experimental system is tested through several groups of contact components and gets good results. The future research direction is proposed after the error analysis.