Rejecting Character Recognition Errors Using CNN Based Confidence Estimation

Abstract

Although Optical character recognition (OCR) technology has achieved huge progress in recent years, character misrecognition is inevitable. In order to realize high fidelity content of document digitalization, we propose a new Convolutional neural networks (CNN) based confidence estimation method. We detect the misrecognized characters through comparing the confidence value with a preset threshold, so as to leave the recognition errors as embedded images in the output digital documents. We adopted sofmax as the estimation of posteriori probability, overlap pooling and maxout with dropout technologies in CNN architecture design. Experimental results show that our method has achieved an explicit improvement compared to baseline system.