Year: 2025
Type: article
Abstract: In the field of computer vision, small target detection is extremely challenging due to the small size, low pixel density, and lack of contextual information of targets, making it difficult for traditional algorithms to effectively identify small targets because of frequent missed and false detections. Therefore, designing efficient and accurate detection algorithms is crucial. This paper proposes a novel small target detection method named YOLO_GF, which is based on the YOLOv5s framework. It introduces an improved DAMO-YOLO dense association memory mechanism to address the insufficient exchange of information between high- and low-level features during feature fusion. Furthermore, by integrating the improved efficient multi-scale attention EMA, the method utilizes the wide receptive field of parallel subnetworks to collect multi-scale spatial information and establish interdependencies among different spatial locations, thereby realizing a cross-spatial learning mechanism to enhance model accuracy. Additionally, the method employs a novel convolutional CSPStage to construct the feature fusion module, reducing model inference latency while improving detection accuracy. YOLO_GF achieves an end-to-end small target detection process and has been validated on the VisDrone2019 dataset, with experimental results demonstrating its high detection accuracy and good detection performance.
Author: <p>Yi Shi, Lei Ding, Shan Li, Xin Wang</p>
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4409032402