面向星载平台的光学遥感图像港口船检测

Port ship detection of optical remote sensing images for space-borne platforms is an important branch in field of remote sensing. On-board hardware resource limitations make existing detection methods difficult to meet real-time processing requirements: Although the knowledge-driven method represented by manual features has certain detection capabilities, its robustness is low. Because manual features are difficult to fully characterize and model the ship's prior knowledge and are susceptible to complex background interference. However, the existing data-driven deep learning methods tend to have high computational complexity, and the computing resources of the onboard platform are difficult to support the existing deep learning methods. In view of the above problems, this paper proposes a lightweight deep learning port ship detection method for space-borne platforms. First, we use a dense connection mechanism and a two-way convolution to design a lightweight backbone network with lightweight and strong gradient streams. Secondly, in order to solve the problem that the ship target has large scale difference, we design the feature pyramid structure with context feature fusion to improve the detection capability of multi-scale ships.At the same time, to improve the detection performance of the rotate object, the method used the quadrilateral anchor regression mechanism to accurately locate the port ship with the rotating direction. Compared with the current classical deep learning object detection method in many complex port scenarios, the experimental results show that the proposed method can achieve high detection accuracy while reducing the computational complexity by 30 times.