Researches On Geometric Graphic Analysis And Process Issues

Most objects involving Computer Vision and Image Processing can be decomposed into several geometric primitives, which can be expressed by analytical functions, such as points, line segments, circles, rectangles, etc. Particularly in the computer vision inspection system and control system, the geometric elements, which composed by geometric primitives, are used to describe objects. The geometrical properties of the objects are very helpful to improve the accuracy and efficiency of computer vision algorithm.The paper focus on the three key technologies of geometric elements image processing: image segmentation, geometric elements shape recognition and geometric parameter measurements. Based on some typical algorithms, introducing pulse coupled neural networks (PCNN) and particle swarm optimization (PSO), the automatic image segmentation algorithm, shape recognition algorithm and geometric parameters detection algorithm are constructed respectively. These algorithms are applied to the LED die geometric parameters detection and measurements system and calibration of servo motion control system. The related research work and main context are presented as follows:(1) Utilizing PCNN for automatic image segmentation.PCNN is novel neural models proposed by emulating the mechanism of small mammal's visual cortex. By its nice characteristic of spreading pulse the PCNNs couple term and synchronous pulse burst property, Image segmentation is processing according to the similarity of the pixel gray and correlation of spatial distribution, and is favorable toward preserving the image detail. The image segmentation process is completely dependent on the natural properties. Because the difficult problems for applications of multitudinous PCNN parameters and complex structure, The article proposes automatic image segmentation PCNN (PSO-PCNN) based on PSO algorithm: firstly, based on the simplified model Unitlink-PCNN image segmentation algorithm, Lena and the LED die image segmentation experiment verifies the feasibility of the method, it also compares the results and the best number of iterations under different segmentation criterion. Secondly, the design of fitness function is employed to reflect segmentation quality and operating efficiency. The PCNN model parameters are automatically set using improved particle swarm optimization algorithm. Simulation results show that the PSO-PCNN can suppress the weak edge of the background noise effectively, and the accuracy of pixel detection is higher than the PCNN and OTUS automatic thresholding segmentation algorithm.(2) Shape recognition algorithm for geometric primitives. For the feature, feature descriptor and feature similarity measure have potentially huge impact on shape recognition, classification and efficiency of the algorithm, this article describes the several common descriptors, focusing on the applications and performance of invariant moments especially Hu moments in shape recognition. The contour moment is introduced to solve computation problems of large amount of data for regional moment. Experimental results show that computation efficiency has been greatly improved by contour moment compared to regional moment. It also verifies the contour moment has a good translation, scaling, rotation invariance in the identification.(3) The geometric parameters of geometric primitive detection algorithm.Existing geometric primitive detection algorithm has defects of computational complexity, detection of low accuracy, sensitive to background and noise. In this paper, randomized Hough transform based (RHT) of geometric primitives detection is focused based on the standard Hough transform and analysis of several typical Hough transform. The RHT and the least squares method combination of geometric parameters of geometric primitives detection algorithm method is proposed (RHT-LSM). The algorithm selects the target pixel set of edge points from the noisy data set, reduces the fitting error of the edge point set using the least square optimization (LSM). LSM is used to solve issues of RHT algorithm for low precision, the large memory consumption and peak proliferation. Finally this paper shows the lines, rectangles and circles detection algorithm based on RHT-LSM. Probabilistic model of random sampling in multi-target detection is also analyzed. Methods are approached by summary of designating sub-region, combined with prior knowledge, excluding Invalid non-target point to reduce the probability. Experiment results show that the RHT-LSM algorithm has good performance on object detection and higher precision(4) LED die application of visual detection and measurements systemThis paper describes elements of the LED die detection and measurements system, basic working processes and visual inspection of LED die platform. Application experiment consists of two parts: First, LED die microscopic image is obtained by the experiment platform, the binary segmentation image of LED die electrode is extracted by PSO-PCNN automatic image segmentation algorithm. After connected-area scan and similarity measure analysis of binary electrode image, achieve shape recognition of LED die electrodes. The parameters of LED die electrode are calculated by RHT-LSM geometric primitive inspection algorithm. Second, with the calipers image, calibration experiment of servo motor system is completed. The experimental results show that the analysis and processing algorithm of geometric primitives has high detection accuracy and good robustness.