Shallow multiplexing and multiscale dilation convolution combined attention based oriented object detection in remote sensing images

Remote sensing images are becoming increasingly important in many areas of life because of the valuable information they provide. However, detecting objects in these images remains a difficult task due to their complex and variable characteristics, such as size, scale, and orientation. Moreover, there is a growing demand for efficient and speedy detection methods in practical applications. Therefore, in this paper, we propose a framework for oriented object detection in remote sensing images based on shallow multiplexing and multiscale dilation convolution combined attention. To achieve a lightweight network structure, we utilize ResNet18 as the backbone network. First, a shallow multiplexing module (SM) is designed to improve the utilization of detailed information in the shallow layer of the network. It enhances the interaction between the shallow and deep layers, resulting in a richer representation of network features. Second, a multiscale dilation convolution combined attention module (MDCA) is proposed to prioritize contextual information by using convolution with different dilation rates. This guides the network to focus more on the object information in remote sensing images. Then, the dilated encoder (DE) is employed at the feature fusion stage to enhance the semantic information of the context and produce a feature map with multiple receptive fields. Finally, the log₂ loss function is applied to improve the training results. The experiments are being conducted on three publicly available remote sensing image datasets, and the results demonstrate that the proposed algorithm outperforms other algorithms in terms of detection performance on these datasets. Code is available at https://github.com/sbsfsum/SM-and-MDCA.