Open-Set Remote Sensing Object Detection Using Edge Information Extraction

Typically, remote sensing object detection is limited to a closed-set detection environment. A problem is that when detecting the new untrained but valuable objects, they are incorrectly classified as known classes or background. In this paper, a method is proposed for open-set detection of remote sensing objects, which assigns robust pseudo-labels to unknown classes and trains the network to recognize new classes. The main idea is to combine the feature information of remote sensing objects and select regions in the image that are likely to contain unknown classes to form pseudo-labels. Through supervised learning, the network can distinguish and detect both known and unknown classes. Remote sensing objects are observed from an Earth observation perspective, with rich edge information. Pseudo-labels for unknown classes are obtained using image convolution features and object edge information. Experimental results show that this method outperforms existing methods in open-set target detection for remote sensing images.