Squeeze and Excitation Rank Faster R-CNN for Ship Detection in SAR Images

Abstract

Synthetic aperture radar (SAR) deliver detection is an essential a part of marine monitoring. With the improvement in pc vision, deep gaining knowledge of has been used for deliver detection in SAR pics including the quicker area-primarily based totally convolutional neural community (R-CNN), single-shot multibox detector, and densely related community. In SAR deliver detection field, deep gaining knowledge of has plenty higher detection overall performance than conventional strategies on near shore areas. This is due to the fact conventional strategies want sea–land segmentation earlier than detection, and inaccurate sea–land masks decreases its detection overall performance. Though cutting-edge deep gaining knowledge of SAR deliver detection strategies nonetheless have many fake detections in land areas, and a few ships are missed in sea areas. In this letter, a brand new community structure primarily based totally at the quicker R-CNN is proposed to in addition enhance the detection overall performance with the aid of using the use of squeeze and excitation mechanism. In order to enhance overall performance, first, the function maps are extracted and concatenated to achieve multistage function maps with Image Net retrained VGG community. After area of interest pooling, an encoding scale vector which has values among zero and 1 is generated from sub feature maps.