Few shot object detection in remote sensing images

Abstract

Object detection in remote sensing images has far-reaching significance. However, compared with object detection tasks in the field of natural images, there are still some challenge problems that need to be improved in remote sensing, due to the similarity and unbalanced scale between objects in remote sensing images. Currently, object detection algorithms based on deep learning, i.e. Faster-RCNN, YOLO, SSD, etc., have reached an incredible grade with the concerted efforts of researchers, and have been used in various aspects such as remote sensing, face recognition, pedestrian detection and so on. However, these methods always need a lot of labeled data, while collecting a dataset is labor intensive and time consuming. In order to overcome these difficulties, we introduce a few shot learning algorithm with attention mechanism for object detection in remote sensing images, aiming to detect objects of unknown classes with only a few labeled remote sensing images. We use the structure of Siamese-network to extract the features of the target from support images and query images, and then use the features of support images as a kernel to do a depth-wise convolution on the feature map of query image. By this way, we can enhance the characteristics of the target category and weaken the characteristics of other categories and the background. For unbalanced scale problem, our method takes care of various objects of different scales in the process of extracting features. We integrate the information of multi-layer feature maps and make predictions on three feature maps of different scales. Experiments on HRSC-2016 dataset validate the effectiveness of our method.