SCFNet: Semantic correction and focus network for remote sensing image object detection

In high-resolution remote sensing images, the problems of large scale variation, large intra-class variance of background, small variability and irregularity of arrangement between different targets always exist in remote sensing images, making the modeling between targets and background more difficult and the target detection task more difficult. However, general target detection methods mainly use convolutional layers of different scales to enhance the target's perceptual domain and fuse different scale features to solve the scale variation problem, without considering the other two problems prevalent in remote sensing scenes of earth observation. In order to solve the above two problems, this paper proposes a semantic correction and focusing network (SCFNet) from the perspective of modeling the relationship between background and target and target to target. The network consists of two core modules the Local Correction Module (LCM) calculates the similarity of local features through the global features of the image to correct the local features and exclude the non-relevant The Non-local Focus Module (NLFM) enhances the recognition of target features by obtaining the non-local dependencies and the corrected local features from the LCM. To demonstrate the effectiveness and robustness of our proposed method, we conducted extensive experiments on two publicly popular large remote sensing multi-target detection datasets, namely DIOR and DOTA. the experimental results show that our SCFNet achieves best-in-class performance and significant accuracy improvement on the datasets.