Rapid Digital Ball Detection And Recognition

Pattern recognition is an important research in the field of artificial intelligence, and character recognition is an active area of pattern recognition. The traditional character recognition methods are mainly based on planes. There is a considerable deformation on the curved surface, and it's also difficult to determine right direction for characters on curved surface. Thus, it brings great challenges in the field of character recognition on curved surface. This thesis do research and experiments on detection and recognition of digital ball.The second chapter describes the detection process of digital ball and three-dimensional reconstruction of the digital ball. In this thesis, we use Hough transform method and constructing circumscribed circle method to detect and locate the digital ball. According to the results of camera calibration and the geometric constraints, we establish the relation between two-dimensional projection images and three-dimensional spherical surface, then we achieve the three-dimensional reconstruction of the digital ball.The third chapter explains the detail feature extraction process of digital ball. First, we describe an ellipse detection algorithm to extract interest area of digital ball, the corresponding spherical interest area is established through the three-dimensional reconstruction. Second, we make the camera towards the center of the corresponding spherical interest area according the three-dimensional coordinate transformation. Then we re-project the digital ball. The projection results adjust the interest area whose size, shape and direction are confused formerly to a unified scale. Finally we use the form of polar coordinates to describe the results, which transform the rotation problem into one-dimensional translation problem.The fourth chapter discusses three kinds of recognition methods, including recognition methods based on template matching, PCA and sparse representation. We improve the recognition accuracy through multi-sampling of the same object, and Bayesian theory is also applied in the process of recognition method based on sparse representation. In our experiment, we achieve a 100% recognition rate using multi-sampling recognition method based on sparse representation and Bayesian theory.The fifth chapter discuss the general ball recognition whose surface are common pattern. In this chapter, we introduce the spherical harmonic functions and spherical correlation method. We propose a method to synthesis spherical surface automatically. Then, we describe the procedure of pose estimation and ball recognition.