Research On Rotation Invariant Off-line Handwritten Digits Recognition

Off-line handwritten digits recognition is an important research subject of pattern recognition, which is be applied in various fields, such as postcode, bank check, industry etc. As the application areas of handwrittern digits recognition is wider and wider, the demand of recognition is more stringent than the past. Thus it is of both theoretical and research significance to carry on recognition of rotation invariant off-line handwritten digits.There are two cases of rotation invariant handwritten digits recognition, including recognition of digits without rotation, where rotation invariant can be guaranteed by slant correction, and free rotation digits recognition. This thesis carries a deep study on those two situations and acquires a series of valuable results which can be summarized into the following aspects:1. Improvement of direction feature on handwritten digits without rotation. In order to reduce errors during direction value extracting, a number of modifications are proposed to the modifed direction feature. During direction feature value extraction process, some adjustment is performed to the conditon of dirction mutation and half dirction is introduced to normalize segments. Two-dimensional array is used to represent the direction value instead of the original one-dimensional to eliminate feature errors. When tested in handwritten digits recognition, the modified feature can obtain higher recognition accuracy, and both lower rejection rate and error rate.2. Modified rotation invariant Pseudo-Zernike feature extraction algorithm. There would be deviation by different values of N during the process of mapping digit image into a unit circle. To address this problem, a fixed N-value is used in this thesis. In addition, a simplified Calculation process of Pseudo-Zernike moment is put forward here to reduce the computational complexity, calculating the Pseudo-Zernike polynomial just on pixels of digit object. The moment feature extracted by modified algorithm has smaller standard deviation rate with less computation time.3. Multistage classification of free rotation handwritten digits. A multistage classifier frame is proposed in this thesis to solve the problems of low recognition accuracy with single classifier and complexity of statistical feature extracting. Serial-parallel structure is applied in this frame, and this first stage uses simple structure feature to carry out rough classification. Statistical features of the rejected digits are extracted and then classified by parallel classifiers in the second stage. The multistage classifier takes advantages of both two kinds of feature and it acquires higher accuracy and lower complexity.