Tri-FRNet: a three-path feature fusion network for high-precision small object detection

Abstract In the task of UAV aerial image and remote sensing image analysis, the detection of small targets is very important. Although the existing target detection algorithms perform well on natural images, the detection accuracy is still challenging when dealing with dense small targets in UAV images and sparse small targets in remote sensing images. In view of the above problems, this paper proposes a three-way feature fusion and enhanced small target detection model-Tri-FRNet. Firstly, the backbone network based on the encoding-decoding architecture is combined with the newly designed feature reconstruction network. Through the reconstructed high-resolution feature fusion and the improved model architecture, the model’s ability to perceive small targets is significantly enhanced, and the number of model parameters is greatly reduced. Secondly, a multi-scale attention module—Multi-Scale Attention Module is proposed, which combines depthwise separable convolution and attention fusion to guide the model to focus on key feature areas and effectively improve the overall detection performance. Finally, an upsampling module based on pixel rearrangement (Pixel Shuffle Upsample) is designed to avoid the sawtooth distortion caused by traditional nearest neighbor interpolation through feature map rearrangement. Extensive experiments on the VisDrone2019 and AI-TOD datasets show that compared with the benchmark model YOLOv11n, the Tri-FRNet proposed in this paper increases by 3.7% and 2.4% on mAP@50:95, respectively. At the same time, the amount of parameters of the model is reduced by 59%, which fully verifies the efficiency and effectiveness of the proposed method. The source code will be available at https://github.com/unscfr/Tri-FRNet .