Arbitrary-Oriented Ship Detection Based on Feature Filter and KL Loss
Chen Li, Lei Wang, Bin Kang, Mingkai Chen, Jingwu Cui, Baoyu Zheng · 2020
Although deep learning has been dominant in the field of target detection, there are still some challenges in the field of detection of ships: horizontal boundary box contains too much redundancy and noise in the scene of large scale and dense arrangement of ships, and there are many interferences in remote sensing optical images that affect the bounding box regression. Recent neural network for ship detection is generally used to extract the pixel information in the image to achieve rapid positioning and boundary box regression, or to process the dataset, so as to train a more robust network. However, these methods do not focus on the original feature map and loss functions of the target, which directly lead to the effects of ship positioning and regression. In this paper, we propose a ship detection network based on feature filter and Kullback-Leibler (KL) divergence loss function. In this paper, we propose to use mask filters on the output of the region of interest (ROI) network to remove the noise around the rotating bounding box, which is conducive to the regression of the rotating bounding box and angle in the second stage. In order to get a better bounding box, we propose to use KL loss function for regression of bounding box parameters. With the optimization of KL loss, the bounding box can better surround the ship. We carried out experiments on our own remote sensing ship image dataset for ship detection, it contains 5,126 remote sensing satellite images and more than 23800 ships in 6 categories. The dataset contains a variety of scenarios and challenges. Experiments show that the method is accurate and effective, and the bounding box has good wrapping property.