A Similarity Constraint Divergent Activation Method for Weakly Supervised Object Detection in Remote Sensing Images.

With the development of remote sensing technology and object detection technology, many fully-supervised convolutional neural networks (CNN) methods based on object labeling information such as bounding box have achieved good results in remote sensing image object detection. However, due to the wide detection range of remote sensing images, diversity of objects, and the complexity of background, it is very difficult to manually label large-scale remote sensing images. Therefore, in recent years, more and more attention has been paid to the weakly supervision method using only image-level labels in object detection. Class activation mapping (CAM) method based on weakly supervision works well for object detection in natural scene images, but it has the problem when it is used in remote sensing images: a large number of small objects are lost. In this paper, we propose an object detection method for remote sensing image based on similarity constraint divergent activation (SCDA). The divergent activation (DA) module in SCDA improves the response intensity of the low response regions in the shallow layer feature map. According to the similarity between the objects, the similarity constraint module (SCM) is used to further improve the feature distribution and suppress background noise. By fusing DA and SCM, the missed rate of small objects can be reduced. Comprehensive experiments and comparisons with state-of-the-art methods on WSADD and DIOR data sets demonstrate the superiority of our proposed method.