Recognition Preprocessing Of Visual Document Image

The images are produced with noise, blur and imbalance or distortion because of the different choices of focus and exposure about imaging devices such as digital cameras, document's fold line, besmirch, and change of angles. All these influence the recognition rate, and even some of the images might be beyond recognition. Recognition preprocessing, which includes removal of noise, optical adjustment and geometry adjustment, is proposed to cope with noise, low contrast and blur and distortion of the visual document images by the digital camera. The visual document images can be recognized effectively and the recognition rate will be enhanced after these preprocessing. The main results are described as follows:1. To dispose the salt-pepper noise in gaining, processing and transmitting the visual document images, a new algorithm for removal of salt-pepper noises, based on the character stroke characteristic, is proposed. This algorithm could estimate whether a pixel and its neighboring pixels are salt-pepper noises or not, and differentiate effectively characters from salt-pepper noises. The experimental results show clearly that this algorithm could restrain salt-pepper noises well and truly. At the same time, it avoids the influence for characters stroke and decreases the disturbance for subsequent recognition, compared with those traditional denoise algorithms.2. An approach of eliminating back-to-front interference in visual document images is put forward according to histogram features. When the gray-level histogram of image is bimodal, the selected areas in the image are either enhanced or normalized to eliminate the back-to-front interference. An algorithm based on image background separation is proposed the dispose the images with unimodal gray-level histogram. It has been proven by experiments that these two algorithms achieve favorable results.3. For the fold noise in visual document images, a object enhancement algorithm is proposed to erase the fold noise and enhance the OCR's recognition rate.4. A new method is proposed to eliminate the blurry images effectively, which is to search for the framework of the character. 5. For the problem of images with poor exposure, the cases are as follows:(1) For the low contrast images with poor exposure,, an image object enhancement algorithm is proposed, which can effectively differentiate image object and image background, set various gray values in image background merged into unitary gray value, and enhance image object. The experimental results suggest clearly that this algorithm can effectively dispose the images with poor exposure and enhance OCR recognition rate.(2) For the images of overexposure, the algorithm of double-sided enhancement is proposed to solve the problem.6. The wavelet transform method is proposed for blurry edge images, which decompose the image using wavelet, and expansion or reduces the coefficient which deduce by wavelet, enhance details which interested in and weaken the noise. This method can enhance the recognition rate efficiently.7. For the images with margin lines, neighborhood-following algorithms is used to remove the margin lines and long lines efficiently.8. In the process of adjusting thick documents, the unidirectional extension method, which realizes expansion of writing lines, is proposed, and the partition fitting method is used based on the analysis of the image characteristics of multicolumn documents and enhance OCR recognition rate.