To solve problems in the complex geometry scene, dense object distribution, and the large range of object size variations in high-resolution remote sensing object detection and to address the limitations of model resources in application scenarios, a lightweight remote sensing object detector based on YOLOX-Tiny is proposed. A multi-scale prediction method is used to enhance the detection capability of dense objects. Moreover, a coordinate attention module is introduced to improve the attention of important characteristics while suppressing background noise. The key prediction convolution layer is replaced by deformable convolution to strengthen the spatial modeling capability. Finally, the loss function is optimized to increase the localization accuracy of remote sensing objects. The effectiveness of the proposed algorithm is evaluated on the public remote sensing image target detection dataset DIOR. The experimental results show that compared with the benchmark algorithm (YOLOX-Tiny), the proposed algorithm improves the average precision (AP) and AP50 indexes by 4.1 percentage points and 4.42 percentage points respectively; on the premise of maintaining high accuracy, the number of detection frames per second (FPS) reaches 46, which can meet the needs of real-time detection and is superior to other advanced algorithms.