| Modern image processing is a synthetic technology including computer image processing, optical image processing, and photoelectric processing. The real-time processing and recognizing of 3-D object can not be accomplished without organic combination of optical and digital processing. It makes use of the optical processing method to quickly realize high efficient operation such as transform of image and correlation operation, while the computer to accomplish the flexible and precise control, analysis and judgment. The focus of research nowadays is pointing to real-time image processing, optics correlation pattern recognizing, and engineering accomplishing fields.The well-known Vander Lugt optical correlator, appeared in 1964, indicated optical pattern recognition (OPR) technology came into being. With 40 years of research and development, OPR has become an important branch of information optics. With the rapid development of computer technology and spatial light modulation (SLM) technology, image recognition technology, realized by optical methods, has gradually changed from the traditional system composed of pure optical components to the system that combined optics with computer hardware and software and has been widely applied. However, traditional optical patternrecognition is based on two-dimensional (2-D) image correlation; witch has difficulties in 3-D object recognition.The primary study of this paper is 3-D object recognition based on structured light projection. The characteristic of the method is to construct a new recognizable plural function by structured illumination, in witch the height distributing information of the object is encoded as the phase of the plural function. So the new method has the characteristic of intrinsic 3-D recognition. Two methods of 3-D object recognition are proposed in the paper. One is based on structured light projection and traditional Vander Lugt correlator. Another is based on structured light projection and neural network (NN). Principles, calculation expressing, recognition system frames and experiment results are proposed. Original work in this paper includes:1. A. new method for rotation-invariant three-dimensional (3-D) object recognition is proposed. It's based on multi-channel filter. Firstly, encoding the depth information by projecting structured light onto reference object in different directions, the deformed fringe images are acquired by CCD camera. After taking the intensity of the first diffraction order of the composite image that processing by computer, the conjugate of the phase distribution that we got is the filter we proposed. Secondly, we got the intensity of the first diffraction order of the test object in the same way. Finally, the intensity of correlation peaks will be detected in the output plane by matching the intensity of the first diffraction order of deformed fringe pattern of the test object and the filter. Different objects can be distinguished from magnitude of the correlation peaks. By using Cubic spline data interpolation for experiment data, the rotational angle of the object can be estimated. Simulated and experimental results show the utility of the proposed method and its ability to estimate the rotational angle of the objects.2. An automatic method for multi-task three-dimensional (3-D) object recognition with rotation invariant is proposed. The method is based on the structured illumination and B-P neural network. A standard B-P neural networkhas been used. Training samples are the first diffraction order of deformed fringe pattern of test object obtained by structured illumination. The samples are taken every 45 °. Each kind of object has 8 samples. Multi-task three-dimensional (3-D) object recognition with rotation invariant achieved by entering the first diffraction order of deformed fringe pattern of test object into trained network. Simulated results show the utility of the proposed method.
|
PDF Full Text Request