| Abstract [eng] |
OCR (optical character recognition) - is raster image digitalizing method which is widely used nowadays and definitely will be used in the future. There are still many documents, books, articles and other papers which is still not digitalized this way. Digitalizing and using OCR improves documents saving, storing, using, searching, processing capabilities. However, various noise types can prevent accurate optical recognition. These noises may be consequence of scanning, processing tasks or defects in the document itself. These noises must be removed to get satisfactory optical character recognition results. One of the ways to achieve that is by using genetic algorithms. Genetic algorithm in this case can be used in various ways: creation of new filters, combinating filters, or by using direct image transformation. However, the main problems which may occur when using these methods for filtering images in the practice are these: the original image without noises is needed or the type and level of noise must be known. The automated method of image filtering by using genetic algorithms is suggested in this thesis. The main goal of this thesis is to implement suggested method and evaluate filtering results by applying this method. The suggested denoising method has these steps. At first noises types and levels are evaluated in the input image by artificial neural network. According to these results, corresponding filter is applied, which had been constructed by using genetic algorithm. These filters consists of some other simple filters, and vary from each other by different level of filter application and the order of application. This thesis consists of these parts. At first used filters and noises are defined. The second part describes genetic and artificial neural network algorithms. Related works are reviewed in the following part. The suggested method and its implementation is described in the next part. Finally, results of the experiment are stated and evaluated. The results show that filtering according to this method decreases level of noise in the image and increases optical recognition accuracy in synthetic images. However, to get similar results with real images, method must be improved. |
