Rotation invariant object detection assisted ship identification from high-resolution remote sensing imagery

In recent years, with the development of high-resolution remote sensing images and ship intelligence, the detection and identification of ship targets through remote sensing technology in a wide range has played an important role in the field of marine supervision and safety, etc. Considering that the human visual circuit system has a strong directional selectivity towards specific targets in the outside world, drawing on the directional selectivity mechanism of vision will help to improve the performance of the detection and identification task of ships. We simulate the direction-selective mechanism of vision in three ways: we use the Gabor convolutional kernel decomposition of the convolutional layer to simulate the directionality of visual circuits, so as to make the deep convolutional neural network directionally invariant; we simulate the direction-selective mechanism by estimating the main direction of ship targets through directional regression; and we combine the directional target with the direction-selective target to improve the performance of ship detection and identification task. The results of the tests showed that compared with Faster R-CNN（Faster Region Convolution Neural Network）, SSD（Single Shot multibox Detector） and ORN（Oriented Response Network） methods, this method can achieve good results and show potential advantages, and the mAP（mean Average Precision） can reach about 98%.