Ship Detection in Visible Remote Sensing Image Based on Saliency Extraction and Modified Channel Features

Ship detection in visible remote sensing (VRS) images has been widely used the military and civil fields. However, various backgrounds variable scale orientation bring great difficulties to effective detection. In this paper, we propose a novel ship target scheme based on small training samples. The contains two main stages: candidate region extraction identification. first stage, visual saliency model difference covariance statistical characteristics quickly locate potential ships. Moreover, multi-scale fusion for is designed overcome problem of variation. second three-channel aggregate feature, which combines rotation-invariant histogram oriented gradient circular frequency feature. feature can identify well by avoiding impact its rotation shift. Finally, VRS dataset that more realistic scenes. results demonstrate achieves best AUC value with 0.9476, overall better performance 65.37% terms [email protected]:0.95, basically meets need tasks.