In the past few years, the digital imaging devices, such as cameras become more and more popular. Cameras have advantages over scanners on capturing images of texts, such as more portable, more convenient, non-contact etc. But the difference on imaging mechanism between cameras and scanners, and the more complex objects they are handling make the traditional OCR software cannot be used in the texts captured by cameras unchangingly. Some factors are not considered or seldom considered in the traditional scanner-based OCR technology. They seriously affect the OCR performance in camera-based text images. This dissertation takes a study on camera-based character recognition technologies, the major contributions are: 1. We propose an integrated perspective distortion correction method for small-square document images. It uses the characters of small-square documents: small area, fewer words and complicate layout. The boundaries of small-square document region as well as the texts information in the document are used to correct the distortion. In the boundary detection part, multiple features are used to detect them correctly. 2. We propose a text curves’ fitting and one-image based method to correct this curved distortion in bound document images. In this kind of correction method, it is the crucial step to locate the curved text lines automatically and accurately. We propose a graph based locally optimized text curves detection method. This algorithm is robust in document distortion type and the curved extent of text lines. After the text curves are detected, the local continuation of text curves are proposed to use as filtering strategy to revise the detected curves. Afterwards, image warping methods are used to correct the image. 3. We propose a text string detection method based on text distribution information. This algorithm utilizes the distribution information, which is the crucial difference between text strings and other objects. A group of stripe features are proposed, which represent the horizontal distribution information of text strings. And in consideration of the task of text detection, one positive data biased AdaBoost algorithm is proposed as feature selection and classifier construction mechanism. In the later process of detection, text strings’ vertical distribution information is used to further filter out non-text regions and make localization more precise. 4. New methods to evaluate the performance of text localization algorithms in different usage conditions are proposed. It includes: 1) One set of metrics in recognition usage are proposed, this set of metrics gives a description of the detection method. 2) One set of metrics in detection usage are proposed, this set of metrics gives a recall and precision evaluation on the object level, and the scores are independent of the sizes of the ground truths. And to describe the text detection method in detail, the sizes of false alarm regions are also considered.
修改评论