FlexShip: A More Flexible Network for Small Target Detection on Marine Ships

Abstract

Remote sensing ship detection is of great significance to the fields of maritime traffic management and marine resource monitoring. Although the latest ship target detection models have shown excellent results, they still face many challenges. We summarize the three key challenges that ship target detection still faces: 1) Due to the resolution limitation of remote sensing images and the small size of ships, the targets show weak edges in the images and are difficult to accurately identify. 2) The sea surface environment is changeable, often accompanied by interference information such as ripples, clouds, and wake waves, which can easily cause false detection and missed detection. 3) The length-width ratio of ships is very different, and the directions are different. Conventional rectangular frames are difficult to accurately envelop targets, resulting in inaccurate positioning and background redundancy. To address the abovementioned problems, this article proposes a new method called FlexShip. Specifically, first, the self-correcting convolution mechanism is applied to remote sensing image target detection. The network adaptively modifies the feature response strength of different scales and contrast regions through dynamic convolution kernel parameter calibration, enhancing the expression ability of weak ship edge structures. Second, as a basic element of the feature enhancement method, we also create a feature-guided attention module to guide the network to focus on the key texture and contour data of the ship area. Finally, polar coordinates are used to move the prediction box. Compared with the rectangular coordinate system, the polar coordinate system has more adjustable rotation angles and fewer parameters. Extensive experiments on multiple public datasets show that the proposed FlexShip achieves state-of-the-art performance in ship target detection.