Multi-scale Semantic Information Refinement and Hybrid Dual Attention Module based Object Detection in Remote Sensing Images

The high-resolution remote sensing images are of high research value in a lot of areas of life as remote sensing technology continues to develop. However, when compared to ordinary images, the more complicated scenes of remote sensing images, characterized by large scale variations of their target objects and complex backgrounds, which lead to the detection of these types of images still confront numerous challenges. Therefore, in response to the above challenges, this paper proposes a framework for object detection in remote sensing images with multi-scale semantic information refinement (MSSIR) and hybrid dual attention module (HDA). In the first place, MSSIR is designed in the feature pyramid section. This module is a multi-scale module with multi-level jump connections and it refines the semantic information of each feature layer in both top-down and bottom-up ways, while using jump connections to increase the connection between the deeper and shallower networks as a way to improve the extraction of feature information from each layer of the network. Meanwhile, transposed convolution is applied in the upsampling process of the feature pyramid, which allows the upsampling process to learn to achieve optimal results through the network. Secondly, the HDA module, which combines the dual benefits of channel attention and spatial attention, is designed to guide the network to concentrate more on feature information. For the last, the addition of the high level in the multi-scale module allows the network to learn further to higher level semantic information. The experimental results on three open datasets demonstrate that the proposed algorithm has remarkably improved the average accuracy with respect to the other algorithms.