Research On Some Key Technologies For Detection Of Irregular Parts Based On Machine Vision

Machine vision plays an irreplaceable role in automatic product inspection due to its characteristics of high speed, high precision and non-contaction. This paper took advantage of the research results of experts in the field of machine vision, combined theory with practice, designed the algorithm of the irregular size measurement and surface defect detection, built the testing platform and developed the irregular clamp detection system.The main contents of this thesis include:(1) Shadow research. In order to solve the shadow interference and improve the robustness of machine vision, shadow is one of the research emphases of this paper. The paper proposed a new shadow detection algorithm based on YCb Cr color space and bilateral filter. In this paper, several kinds of algorithm were analyzed and finally verified the validity and feasibility of the proposed algorithm. In the test, we can see that algorithm can not only remove the shadow but also better preserve the details in the original image.(2) Registration research. In this paper, we systematically studied the existing mature image registration method and similarity measure function which are often used in image registration, and constructed a measure function based on maximization mutual information according to the image gray scale and space structure. The culture algorithm was used to optimize measure function. Cultural algorithm provides a more evolutionary process calculation model. From the view of optimization calculation, any evolutionary algorithm can be embedded into the cultural algorithm as an operation of evolutionary population space. This paper combined the cultural algorithm and particle swarm optimization algorithm as a new algorithm to direct the image registration,the improved algorithm speeded up the convergence speed, thus overcame the problems, such as large amount of calculation and slow search speed.(3) Segmentation research. Neighboring neurons of the traditional PCNN space can fire at the same time, but for the region growing image, pixel space adjacent, which will not lead to division,will appear error segmentation. Neurons in the improved PCNN allow fire only once, if be fired, the output will remain the same. The improved PCNN model neurons receive compensation from the edges and provide the basis for the late accurate segmentation. The improved PCNN model does not need a series of train but only a few times of iteration to extract the information of the edge and region.The experiment results showed that the improved PCNN segmentation effect was superior to other traditional algorithms and had high segmentation accuracy for our test objects.(4) According to the randomness and convergence, we analyzed its orbits respectively using algebraic model, analytical model and state space model. Parameters are the key of the performance and efficiency of algorithm. The main five parameters in the PSO algorithm are discussed in detailed studies. Analyses the convergence of PSO algorithm by using linear discrete-time system, and gives the corresponding parameters. After segmentation, we fit curve based on the segmentation information to get the information such as area, perimeter, eccentricity and spherical feature, which provides reference for the system. In this paper, we fit curve based on least absolute deviation algorithm and the fitting method is particle swarm optimization algorithm. This paper mainly analyzed the convergence of traditional swarm particles optimization algorithm from different aspects,and proved its convergence from the view of mathematics. Then, the improved PSO is used to guide curve fitting with least absolute deviation and ultimately complete size detection. According to the irregularity of the ear, the randomized Hough transform is used to test the angle of the ear and gets good results.(5) Design and built detection system. Field test demonstrated the reliability of the design.Moreover, it was applied successfully in practical production, saved costs and improved the detection accuracy. In this paper, various defect detection algorithms were studied based on the architecture of the detection system and a clamp detection system was implied based on machine vision, and was used in industrial field.The function of the entire system was verified in practical engineering environment, which included: the effectiveness of the system, anti-jamming ability of the system and diagnostic ability of the system.The test results showed that the designed system can satisfy the detection speed: seven clamps per second; detection error was control in 0.3 mm; the lowest detection accuracy is 97%.