Inshore ship detection in remote sensing images based on deep features

Abstract

Inshore ship detection in remote sensing images is a challenging task because of the connectivity and similarity between ships and backgrounds. The usual shape feature is not always applicable because sometimes it is hard to be extracted. In this paper, deep features extracted from a convolutional neural network (CNN) are used for inshore ship detection. In order to feed the CNN with exclusively positive and negative samples, a novel parallelogram image cropping (PIC) method is proposed. The traditional cropping method can only generate rectangle image samples along a horizontal or vertical orientation, thus ships and docks often coexist in these samples. The proposed PIC method can generate parallelogram samples that only contain single ship or dock. For a 2-class classification task, the number of neurons in fully-connected layers of the widely used CNN models is too large. Reducing the number from 4096 to 512 can improve training and detection speeds while maintaining accuracies. Experimental results on Google Earth images demonstrate that the proposed method can accurately and robustly detect inshore ships in remote sensing images.