Research On DSP Based Image Recognition Algorithms

Image recognition is an important research topic of pattern recognition. This thesis takes printed character recognition algorithms as research object. By analyzing the recognition process, some preprocess algorithms such as image enhancement, binarization, filtering, character segmentation, inclination correction and so on are selected and then implemented. In addition, to reduce image features and computation, the image of each character is normalized and features are abstracted using grid methods.Neural networks are widely used in artificial intelligence, pattern recognition, image processing and many other areas for its excellent nonlinear approximation capability. In this thesis, the fastest version of back propagation algorithms, LMBP (Levenberg-Marquardt Back Propagation), is presented and improved. A matrix compression method is used to decrease the storage requirements. In addition, by deeply analyzing the algorithm, the progress of most computational burden is found. A conjugate gradient method is introduced to accelerate the convergence process. The algorithm performance is improved.At last, these algorithms mentioned above are transplanted to TI TMS320DM642 DSP. A further improvement is taken in the compile option and source code to meet real-time requirements. An acceptable result is obtained.