Multiscale Strip Module and Aspect Ratio Sensitive Loss for Remote Sensing Object Detection

The aspect ratio and orientation sensitivity of objects in remote sensing images make object detection under the traditional framework face great challenges. To address these issues, we propose two methods: multiscale strip module (MSM) and aspect ratio sensitive angle loss (ARS Loss). MSM uses parallel strip convolutions with different receptive fields to enhance long-range feature extraction of slender objects, while ARS Loss dynamically adjusts the angle regression penalty according to the target aspect ratio to alleviate the IoU degradation caused by angular bias. Experiments on DIOR-R dataset demonstrate the effectiveness of our proposed method, which achieves 69.75% of the highest mAP under single-scale training. Compared with the baseline model, the proposed framework improves the detection accuracy of slender objects (such as bridges and harbors) by up to 1.57%, while maintaining robustness to near-square objects. This work provides a comprehensive solution for dealing with different geometric targets in remote sensing scenes.