An enhanced framework for small object detection with middle-order interaction and adaptive cross-scale aggregation
Zihan Guo, Xingyu Mu, Chao Chang, Weijie Cheng, Xincheng Tian · Engineering Applications of Artificial Intelligence · 2025
Small object detection holds significant research value in fields such as autonomous driving, intelligent traffic monitoring, and military reconnaissance. However, in practical applications, the performance of commonly used detection algorithms is often suboptimal due to limited pixel information, dense distribution, and susceptibility to background interference for small objects. To enhance the accuracy of small object detection, we propose an improved algorithm based on You Only Look Once version 8 (YOLOv8), termed You Only Look Once version 8-For Small-object Detection (YOLOv8-FSD). First, we introduce Spatial-Channel Aggregation Module (SCA Module) into feature extraction network, leveraging multi-order gated aggregation mechanism to adaptively capture and integrate interactions among different orders of features, thereby enhancing representation of multi-order features. Second, we design Attention-Guided Path Aggregation Feature Pyramid Network (AG-PAFPN) to add pathways for extracting and aggregating information from large feature maps, generating more discriminative features. Additionally, Multi-Scale Feature Extraction Module (MSFE Module) and Three-Branch Feature Fusion Module (TBFF Module) are proposed to extract and fuse multi-scale spatial information across multiple scales and finer granularity. Finally, we optimize the bounding box regression loss function to accelerate convergence and improve detection accuracy. Experimental results on Visual Detection of Drones (VisDrone) dataset show that the proposed YOLOv8s-FSD achieves mean Average Precision at Intersection over Union (IoU) = 0.50 (mAP50) and mean Average Precision at IoU = 0.50-0.95 (mAP50-95) improvements of 8.3% and 5.5% over YOLOv8s. Migration experiments on TinyPerson and Dataset for Object deTection in Aerial images (DOTA) further validate the effectiveness and robustness of the proposed enhancements.