Novel context sensitive image thresholding technique

Abstract

Image segmentation is a fundamental task in image processing, video processing and computer vision applications. This is a wide area of research. A lot of research work has been done in this field, still there is not a unique technique to segment each type of image i.e., for each type of images there exist a different technique to segment the image. Histogram based traditional thresholding techniques do not considered spatial contextual information for selecting the optimum threshold and are effective only to identify single threshold. In this thesis we proposed a novel thresholding technique that mitigated both these limitations. First, we proposed an energy function that computes the energy of the image at each gray value by taking into an account the spatial contextual information of the image. The energy value is computed in such a way that the characteristic of the energy curve is similar to histogram of the image. Thus, by using the energy curve instead of using histogram, we incorporated spatial contextual information in threshold selection process. Second to mitigate multiple thresholds selection problem, here we exploited genetic algorithm. The fitness function of the genetic algorithm is modeled by extending the criterion proposed in [17]. We improved Kapur’s method [17] results by using the energy curve generated by our spatial contextual information method and compared the results on the basis of DB index with Kapur’s original method. To find the thresholds values is an optimization problem. So we used genetic algorithms to find the multiple thresholds values. Thresholds result shows that genetic algorithm is very promising in this field. Results show that this spatial contextual information which represented as the energy curve for the image is very effective for the better segmentation of the image.