The Research Of Algorithms For Document Image Processing

Document image processing has always been a hot branch of image processing. It is used widely in image retrieval, image analysis, intelligent transport and automatic processing of cheque and identity card. It includes image filter, correction, text location, skew detection, binarisation,character recognition and so on. Text location and skew detection have been paid more attention to than others. And many corresponding algorithms have been proposed. Four novel algorithms on text location and skew detection are proposed in this thesis on the basis of existing method.The first algorithm proposed in this thesis is page segmentation using mathematical morphology. It contains two stages: coarse page segmentation and post processing. Candidate text regions which also include some picture regions are located using morphological operations in stage of coarse segmentation. Then morphological operations are called again to remove the remaining picture regions. Experimental results show that this algorithm can segment document images whose layouts are diverse in an acceptable accuracy.For document image whose background is complex, text regions can not be located accurately only by morphological operations. To solve this problem, wavelet transform is introduced. Wavelet coefficients in regions which contain only background are much smaller than those in regions which contain both background and text. According to this significant difference, text can be located accurately after binarisation and morphological constriction. Wavelet transform is also used in skew detection in this thesis. First, the skewed document image is decomposed using wavelet, producing four sub-bands: low-low, low-high, high-low and high-high. Then the project profile analysis is introduced to the high-low sub-band. Portion of non-text regions in the whole image has less effect on the detection accuracy because of the introduction of wavelet. And the computation is also reduced significantly because of down-sample operation in wavelet transform.Skew detection using particle swarm optimization (PSO) is also proposed in our thesis. Through the global searching ability of PSO, parameter value that maximizes the fitness value which is considered as the skew angle in our algorithm can be found. Experimental results show that using this method, any angle in the range of [ ?90 ,90 ]can be detected in high accuracy for many documents.Document images with various layouts and font sizes are used in our experiments to evaluate the effectiveness of the algorithms proposed in this thesis. And the results show that the proposed algorithms are robust and fast.