Objective Assessment of Nonlinear Segmentation Approaches to Gray Level Underwater Images

Automatic target recognition is a challenging task with a wide variety of potential applications in the industrial, military and medical fields. Region segmentation is a crucial step towards automatic segmentation of images. Under some severe conditions of improper illumination and unexpected disturbances, the blurring images make it more difficult for target recognition, which results in the necessity of segmentation. For instance, the underwater images are generally captured under water dispersing and atmospheric fluctuation. Segmentation is thus needed to clarify feature ambiguity against stochastic disturbances. Region segmentation splits images into regions based on similarity measures, such as pixel intensities, locations and textures or combinations. It categorizes an image into separate parts, which correlates with objects involved. Both K-means segmentation and watershed segmentation can be applied. Segmentation by K-means clustering refers to grouping similar data points into the clusters. It requires that the number of clusters be specified whose distance metrics should be defined to quantify orientation closeness of objects. The winner-take-all algorithm can thus be selected to update the cluster centers. It has the capability of simplifying computation and accelerating convergence. Another typical methodology is watershed segmentation. It is based on the gradient magnitude of images, which can classify diverse objects automatically, where watershed lines separate catchment basins. The erosion and dilation operations are essential procedures involved in watershed segmentation. Also to avoid over segmentation, the foreground and background markers should be selected accordingly. To evaluate the actual role of nonlinear image region segmentation, quantitative statistical measures have been proposed, such as the gray level energy, discrete entropy, relative entropy, mutual information and information redundancy. The assessment measures will further quantify the impact from image segmentation. The objective assessment approach has the potential to solve other image processing issues.