Vehicle detection in high resolution satellite images with joint-layer deep convolutional neural networks

Vehicle detection can provide volumes of useful data for city planning and transport management. It has always been a challenging task because of various complicated backgrounds and the relatively small sizes of targets, especially in high resolution satellite images. A novel model called joint-layer deep convolutional neural networks (JLDCNNs), which joins features in the higher layers and the lower layers of deep convolutional neural networks (DCNNs), is proposed in this paper. JLDCNNs aim to cover different scales and detect vehicles from complex satellite images rapidly by overcoming the insufficient feature extraction of traditional DCNNs. The model is evaluated and compared with traditional DCNNs and other methods on our challenging dataset which includes 20 high resolution satellite images (including over 2400 vehicles) collected from Google Earth. JLDCNNs significantly improve the precision rate by 16% and the recall rate by 6% compared with traditional DCNNs, let alone outperform other traditional methods.