Ship Detection Method based on Scale Matched R3Det

Abstract

Given their high detection rates and low false alarm rates, object detection neural networks based on deep learning have been widely used in ship detection. However, the detection in real-world scenario with complex back ground remains a challenge in marine dynamic ship detection, whose performance is limited by the scale of training datasets, where training a model with high-performance usually requires a large number of multi-scale datasets. However, it is difficult to obtain a large-scale dataset in such cases. In addition, R3Det solves the problem that the vertical and horizontal ratio of the object to be detected is large, the objects to be detected are densely arranged, and category asymmetry of objects to be detected have been widely concerned. However, R3Det uses the nearest neighbor interpolation to up-sampling the image, which leads to a blocky effect of the image with a certain probability, which affects the object detection. In order to alleviate these problems, we propose a new model called “Refined Single-Stage Detector with Feature Refinement for Rotating Object based on Scale-match”. The new pre-training strategy of scale match and improved feature pyramid network (IFPN) were introduced. The method not only expands the training data sample set, but also improves the clarity of training pictures, and improves the ship detection rate and reduce the false alarm rate. Experiments with DOTAv1.5 and high-resolution datasets showed that the ship detection rate and false alarm rate are better than baseline methods.