YOLO-FR: An enhanced object detector for small object detection in complex scenes

Abstract

As a key technology, small object detection has shown broad application prospects in the fields of UAV aerial images, remote sensing images, medical diagnosis, intelligent transportation and industrial quality inspection. However, due to the small size of the small object itself, it is susceptible to factors such as low resolution, limited semantic information, and complex background interference, resulting in its detection performance still facing severe challenges. To this end, we propose an improved object detection model YOLO-FR, which aims to improve the detection accuracy and robustness of small objects in complex scenes. Firstly, the C3k2_WT module is designed to expand the receptive field layer by layer and capture the context information by using the cascaded frequency decomposition characteristics of wavelet transform. At the same time, a C3k2_ECA module integrated with Efficient Channel Attention is designed to effectively reduce the number of parameters and enhance the selective response of channel features. Secondly, a small object detection head is introduced to improve the model's support for small object detection. Additionally, the Quad Fusion Downsample downsampling module is proposed to reduce information loss and improve downsampling efficiency. Finally, we compare a variety of emerging Mamba modules to determine the most suitable module for the task of this article, and optimize it on this basis. Our model improves mAP@50:95 by 6.6% and 2.2% on the public benchmark datasets VisDrone2019 and AI-TOD, respectively. At the same time, the parameters of the model are reduced by 36.5%. We also evaluate the segmentation performance of the model on the UAV instance segmentation dataset UAVid, which demonstrates the advancement and applicability of the proposed model in the face of different tasks. The source code will be available at https://github.com/unscfr/YOLO-FR.