Ship candidates extraction for optical color imagery

Ship detection is of great significance lowing to its wide applications. In most existing approaches, some predetection methods are often used to extract ship candidates since applying an accurate algorithm throughout the whole image will be time-consuming and could even cause a lot of false alarms. In addition, most related work focuses on panchromatic imagery but only a little attention has been paid to color imagery. Color images contain more discriminative information of ships than panchromatic images, so it will be easier to extract ships in color images. Further, more information also means more potential to implement image enhancement techniques to solve the problem caused by poor illumination, which is very common in optical images. In this paper, with respect to optical color imagery, we propose a new predetection approach to extract ship candidates preliminarily and rapidly using color information. Firstly, an image enhancement algorithm is employed to improve the quality of input images. Then, we regard the color image as a hyperspectral image and extract ship candidates using a hyperspectral algorithm based on spectral signature model. This hyperspectral algorithm, in essence, utilizes the color information of ships, but the color information is processed in a hyperspectral manner. Unlike the commonly used color segment algorithms which focus on the thresholds in color space, this hyperspectral algorithm concerns more on the patterns of color vectors. Experimental results on real dataset indicate that this image enhancement algorithm is quite suitable for remote sensing images and its performance is better than histogram equalization based techniques. In addition, the hyperspectral algorithm also shows good performance in extracting ship candidates in color images, especially for small ships. As a whole, large areas of background can be removed and most ships can be detected. Although some false alarms still remain, the mount of false alarms is decreased greatly.