Identifying Semantic Category of Distorted Images along with Text Recognition

Scene recognition provides visual information from the level of objects and the relationship between them. The main objective of scene recognition is to reduce semantic gap between human beings and computers on scene understanding. For example, recognize the context of an input image and categorize it into scenes (forest, seashore, building etc). Some of the applications of scene recognition are object recognition, object detection, video text detection etc. Different methods are there for scene recognition and understanding. All are having positive and negative aspects. One of the main difficulties is to increase the accuracy. The negative aspects which affect the improvement in accuracy are, intrinsic relationship across different scales of the input image are not analyzed and impact of redundant features. This paper develops a framework to overcome these limitations and provide better understanding of input image. The suggested framework includes reconstruction of distorted image and text recognition, if any, along with scene recognition to get a clear idea about the input image. Image reconstruction is done by using Non Local Mean Image Denoising method. To detect and describe local features of an image SIFT method is used. Finally classification is done by SVM classifier. Text recognition is achieved with the help of OCR Methodology.