Ship Detection In Optical Remote Sensing Images With Deep Feature

Deep learning has achieved superior performance in object detection with natural images.The optical remote sensing images share a large amount of common visual features in low and middle abstract levels with natural images.In this paper,we transfer learning the deep learning network over object detection,adjust the bounding boxes with the predicted direction and width using deep feature,and compress the deep learning network with Singular Value Decomposition.Firstly,a state-of-the-art object detection model pre-trained from a large number of natural images was transfer learned for ship detection with limited labeled satel ite images.Experiments demonstrated that our method could achieve 87.9% Avarage Precision(AP)at 47 Frame Per Second(FPS)using NVIDIA TITAN X.Secondly,this paper proposed a semi-supervised method for ship direction and width prediction,combining the deep convolutional feature extracted from transfer learned Convolutional Neural Network(CNN).We pre-learn the pseudo-label of ship direction and width using traditional method,and train the prediction model combining the transfer learned deep feature.Experimental results showed that the performance of ship detection is improved with 1% while the bounding boxes are adjusted according to the predicted direction and width.Deep learning networks are computing intensive and storage-intensive.This paper combined the idea of model compression in deep learning,and proposed a Singular Value Decomposition method.It transformed the standard convolution layer into multiple depthwise separableconvolution layers with eltwise,which reduced the number of parameters and multiply-accumulate operations without training data.Experimental results demonstrated that our method can achieve 86.8% AP along with only 1.1% decrease at 57 FPS.