Research On Ship Detection Algorithm For Optical Remote Sensing Image Based On Hierarchical Stack Classifier

One-class Support vector machine(OC-SVM) is an important branch of support vector machine(SVM) which has favourable performance in solving the problems of small samples, high dimensional samples and imbalance data sets. With its good generalization properties, it is widely used in the applications of unsupervised clustering, anomaly detection and so on. Ship Target Detection Based on optical remote sensing images has many important applications in some fields, such as civil and military reconnaissance, and it is one of today's hot topic. In this paper, a lot of research about ship detection algorithm and the support vector machine theory is studied. A new Ship Detection Algorithm for Optical Remote Sensing Image Based on Hierarchical Stack Classifier is presented. The main workin this paper includes:1. To solve the problem that based on traditional CFAR detection algorithm to detect targets needs to calculate the mean and variance point by point, We presented an improved CFAR ship detection algorithm which is based on morphological method. First, the top-hat morphological operations is used to weaken the impact of non-uniform light intensity, while accomplishing the goal to make the candidate pixels stand out from the background. Then,the suspected target areas are obtained by the threshold calculated by Otsu's method.Finally, CFAR operation is adopted to get the final ship target candidate regions.2. Gradients of ships in the pictures with medium and low-resolution target have small changes in the local area, but there exist significant differences among the stern, bow and hull. Based on the idea above, an improved HOG feature is presented. First, each candidate ship target is divided into three parts based on its long axis. Then, gradient direction histogram for each part is calculated and normalized separately. Finally, feature vector is obtained. This feature not only can effectively describe the characteristics of ships, but also reduce the characteristic dimension.3. After obtaining the candidate building regions, ship detection model based on hierarchical classification stacking architecture is presented to remove false alarms. First,the OC-SVM is used as the base classifier to solve the problem that the number of shipsand false alarms obtained is extremely imbalanced and the false alarm samples are with uncertain characteristics to be described. Then, due to the differences in requirements, the ships undetected are with the different risks, so the classifier design strategy based on risk constraints is presented. Finally, hierarchical stacking architecture is utilized to over come the shortage that it is hard to simultaneously achieve the purpose of high detection rates and low false alarm rate by a single classifier.On the supports of vast optical remote sensing data, the dissertation analyses and validates the effectiveness of the proposed ship detection method and its key technologies by experimental means for ship detection. As the consequence, the experiments lay the foundation of algorithm's practicality.