TGeoYOLO: Leveraging Multi-Scale Features and Enhanced Loss for Remote Sensing Detection

As remote sensing technology advances, precise target detection using satellite imagery has become a critical area of research in military and industrial applications. Traditional target detection methods face unique challenges when processing remote sensing images, such as high resolution, complex backgrounds, and detection of small objects. In this paper, we propose TGeoYOLO, which is built upon modifications to the YOLOv8 model. First, we introduce a Multiscale Perceptive (MP) module. Second, we enhance the algorithm with an Enhanced Focal Loss (EFL) function to improve the accuracy and robustness of detecting small objects in remote sensing images. Specifically, the MP module effectively utilizes multiscale features to provide appropriate receptive fields for targets of various sizes. The EFL function optimizes for the localization errors and small object issues prevalent in remote sensing imagery. Experimental results demonstrate that TGeoYOLO performs exceptionally well on the NWPU and VisDrone datasets, achieving improvements of $1.3 \%$ and $1.6 \%$ in detection accuracy over the original YOLOv8 model.