Combining Multi-Threshold Saliency with Transfer Learning for Ship Detection and Information Extraction From Optical Satellite Images
Thai Nguyen Truong, Tuan Do Ngoc, Bang Nguyen Quang, Su Le Tran · 2019
In terms of both the military and civilian industries, ship detection and information extraction in satellite optical images are critical. Nevertheless, with dynamic surroundings, such as tides, small islands, clouds and the variety of ships of different locations, forms and sizes, the problem became extremely difficult. To solve these issues, this paper explored a combination of multi-threshold saliency and transfer learning to detect and extract ship information. Unlike traditional methods, the proposed method is general, simple and designed for various types of images with different weather conditions and regions. Visual saliency that concentrates on the emphasis on the outstanding signal from scenes combined with the multi-threshold to extract candidate regions. Then these candidates were discriminated against and classified using the CNN model that was built by the transfer learning technique. For those ships that have been detected, we extracted ship rectangle bounding boxes to measure ships by converting the rectangle size to the actual size. Finally, we determined the ship heading by using the ship head discrimination CNN model. Experimental results on numerous satellite images showed our method's good performance compared with state of the art approaches.