Research on a CEF-YOLOv8n-Based Method for Small Object Detection in UAV Aerial Imagery

Abstract

Highlights What are the main findings? A CEF-YOLOv8n model is proposed for small-object detection in UAV aerial images by introducing the CPF module into the YOLOv8n backbone to enhance feature extraction for low-resolution small objects. The FGFPN, FSFM, and LWSD modules are designed to improve multi-scale feature fusion, strengthen the interaction between shallow detail features and deep semantic features, and reduce redundant computation in the detection head. What are the implications of the main findings? The CPF-based feature extraction scheme proposed in this paper can better retain the detailed features of small objects in UAV aerial images and alleviate feature loss arising from repeated convolution and downsampling operations. The combination of FGFPN, FSFM, and LWSD improves the representation of dense and multi-scale small objects while maintaining lightweight model design, enabling more accurate and efficient small-object detection in UAV aerial images. Abstract To address the recognition challenges caused by the high proportion, low resolution, and significant multi-scale variations of small objects in UAV small-object detection tasks, a UAV small-object detection and recognition algorithm based on CEF-YOLOv8n is proposed. The proposed algorithm uses YOLOv8n as the baseline network and introduces a Partial Convolution-based Cross Partial Feature (CPF) module into the backbone network to enhance the local feature extraction capability for low-resolution small objects. In the neck network, the concept of feature focusing and diffusion is adopted to construct a Focusing Generalized Feature Pyramid Network (FGFPN). A Feature Semantic Fusion Module (FSFM) based on a cross-attention mechanism is designed to complementarily fuse shallow detail features with deep semantic features, thereby enhancing information interaction among objects at different scales. In addition, a Lightweight Weight-Sharing Detection Head (LWSD) is proposed to improve the computational efficiency and real-time performance of the model while maintaining detection accuracy. Publicly available datasets are used for network training and detection evaluation, and comparative experiments are conducted with other algorithms. The results show that the proposed detection and recognition algorithm achieves 37.6% and 22.6% in terms of mAP50 and mAP50-95, respectively, representing improvements of 3.4 and 2.6 percentage points over the original YOLOv8n. Meanwhile, the number of parameters and FLOPs are reduced from 3.2 M and 8.7 G to 2.5 M and 6.9 G, respectively.