Extraction of Character Regions through Machine Learning and Filtering

Abstract

Characters in images are able to provide main information of the image. Therefore, it is important to analyze various kinds of image data and accurately extract the characters in images. This study proposes a new method of excluding background regions and accurately detecting character regions from input images with the uses of MCT features and Adaboost algorithm. The proposed method first extracts candidate character regions from input images with the uses of MCT features and Adaboost algorithm. It then excludes non-character regions and detects real character regions from the extracted candidate regions with the use of geometrical features. In the experiment of this study, the proposed method more robustly detected character regions from various input color images than a conventional method. For performance comparison, this study compared the method based on existing texture analysis and the proposed method. In this study, to qualitatively evaluate the performance of the proposed method of extracting license plate regions, the accuracy measure was defined. The measure is used to show the ratio of the accurately extracted character regions to all character regions of an image. The conventional method using the frequency factor-based texture information had many errors of character region detection, since it failed to execute binarization of background and character regions properly. On contrary, the proposed method made use of MCT features and Adaboost algorithm, effectively filtered candidate regions with the use of geometrical features, so that it detected character regions more accurately. The proposed character detection method is expected to be usefully applied to the fields of pattern recognition and image processing, such as store sign recognition and license plate recognition.