Design of Classification Techniques Based on Computational Intelligence for Optical Character Recognition and Image Filtering

• Main Authors: Bae-Muu Chang, 張百畝
• Other Authors: Pao-Ta Yu
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/54889440225505171688

博士 === 國立中正大學 === 資訊工程所 === 97 === The handwritten character recognition plays an important role in the issue of research and development in image processing. Visual recognition process usually begins with the extraction of some features from an input image. Oriented features of different types are commonly employed for the character recognition in OCR. Actually, a character can be regarded as a construction of parts at different orientations, lengths, and positions. Thus, an approach for extracting such oriented parts from an input image is definitely required. Recurrent neural network (RNN) effectively provides a solution to this problem. Fuzzy inference system (FIS) is famous for its 2 characteristics. First, it is able to handle linguistic concepts. Second, it is easily designed to be a universal approximator which performs a nonlinear mapping from an input domain to a corresponding output domain. Also, it easily reflects the human concept of expertise knowledge while designing a recognition system. As the human concept of expertise knowledge can be realized by a set of fuzzy rules, these fuzzy rules can be maintained to obtain optimized recognition system. Hence, the FIS can more precisely recognize handwritten characters and is effective, flexible, and easily realized by readers. Dempster-Shafer (D-S) theory can be considered as a generalization of the Bayesian theory. It can make inferences from incomplete and uncertain knowledge, provided by different knowledge sources. Also, it considers not only individual class, but also unions of possible classes. An advantage of the D-S theory is its ability to deal with missing and ignorance information. Especially, it can provide explicit estimation of imprecision and conflict between information from different knowledge sources. This is particularly useful in pattern recognition problems. Therefore, the D-S theory can offer a more flexible and general approach since it proposes both uncertainty and imprecision. Another main advantage is its robustness in combining bodies of evidence. It effectively achieves the goal of decreasing the uncertainty in information fusion by means of a combination rule which is applied to evidence sources and increasing confidence in the overall hypotheses. Decision tree (DT) is popularly employed in the field of data mining. Especially, it is a very effective method in classification problems. In image enhancement, the major purpose for noise detection is detecting the pixels whether they are corrupted or not. Hence, noise detection will be strongly considered as a classification problem. In the past years, most well-known approaches require the thresholds for classification which are obtained by manual or forced-searching approaches. They are inefficient and time-consuming. Therefore, it is highly essential to employ a systematic approach for solving this problem. In this dissertation, the particle swarm optimization (PSO) is used to optimize the thresholds to obtain the approximate optimized DT and the suboptimal solutions for a set of these parameters. Particle swarm optimization (PSO) derives next generation utilizing the error values in each generation. It is easier to find out the nearer optimal solution employing the PSO. The major reason is that the PSO computes the vector of movement and derives the position for next generation by considering the optimal solution from the first generation to the current generation. Thus, the PSO has high ability with memory. Support vector regression (SVR) is regarded as a learning algorithm based on the statistical learning theory and the structural risk minimization (SRM) principle. As it has the features of the global optimal solution and the structured risk minimization, it is efficiently employed in the pattern recognition and the data mining. This dissertation proposes a recognition system which integrates the RNN and the FIS to recognize handwritten characters. The system employs the RNN to effectively extract oriented features for 8 orientations of a handwritten character and then these features are applied to create the FIS which can powerfully estimate fuzzy similarity ratings between a recognized character and sampling characters in the character database CHDB. Also, another recognition system integrates the simpler RNN and the D-S for the recognition of handwritten characters. The system also utilizes the RNN to powerfully extract oriented features for 8 orientations of a handwritten character and then these features are employed to the D-S which can effectively estimate similarity ratings between a recognized character and all sampling characters. In addition, a novel image filter is used for the recovery of the corrupted pixels by impulsive noise in grey-level images. An adaptive hybrid noise detector which is constructed by integrating 10 impulse noise detectors based on the DT and the PSO is utilized. In this filter, the high performance detector is employed to powerfully detect impulse noise and the modified median filters effectively restore for the corrupted images. Finally, another image filter is presented for the recovery of the corrupted grey-level images. The 2-level adaptive hybrid noise detector which is also constructed by integrating 10 impulse noise detectors based on the DT and the PSO is proposed. In the filter, the robust 2-level detector is employed to powerfully detect impulse noise and the median-type with the support vector regression filter effectively restore for the corrupted images. Experimental results in each research objective respectively demonstrate the system which is presented can achieve a satisfying performance and outperform existing related well-known methods.