Adaptive restoration of text images containing touching and broken characters
Abstract
For document processing systems, automated data entry is generally performed by optical character recognition (OCR) systems. To make these systems practical, reliable OCR systems are essential. However, distortions in document images cause character recognition errors, thereby, reducing the accuracy of OCR systems. In document images, most OCR errors are caused by broken and touching characters. This thesis presents an adaptive system to restore text images distorted by touching and broken characters. The adaptive system uses the distorted text image and the output from an OCR system to generate the training character image. Using the training image and the distorted image, the system trains an adaptive restoration filter and then uses the trained filter to restore the distorted text image. To demonstrate the performance of this technique, it was applied to several distorted images containing touching or broken characters. The results show that this technique can improve both pixel and OCR accuracy of distorted text images containing touching or broken characters.