Delving into deep representations for remote sensing image retrieval

Deep features extracted from Convolutional Neural Networks (CNNs) have proved to have strong ability to transfer for various visual recognition tasks, such as image classification, object detection, fine-grained recognition, and image instance-level retrieval. In this paper, we focus on the issue of high-resolution remote sensing (HRRS) image retrieval and take advantage of CNN representations to solve this significant problem. In order to explore how domain-specific fine-tuning would impact the property of CNNs, we retrain several representative off-the-shelf CNN models with a public HRRS dataset RSSCN7, which contains limited annotated satellite images, and perform retrieval on other two standard HRRS datasets, RS19 and UCM. Then we introduce multi-scale concatenation and multi-patch pooling methods for further performance improvement. Our experimental results indicate that fine-tuning is effective to make progress on CNN transferability and provide remarkable accuracy that outperforms previous state-of-the-art methods.