计算机科学
增采样
人工智能
目标检测
计算机视觉
稳健性(进化)
正规化(语言学)
探测器
模式识别(心理学)
特征(语言学)
特征提取
最小边界框
航空影像
跳跃式监视
特征跟踪
精确性和召回率
分割
斑点检测
特征向量
航空影像
预处理器
前景检测
特征匹配
标识
DOI:10.1088/1361-6501/ae7d4b
摘要
Abstract Accurate and real-time small object detection in unmanned aerial vehicle (UAV) aerial images requires robust feature discrimination across diverse target scales and visually similar categories under complex backgrounds. This study proposes a lightweight detection network, SMC-YOLO, which detects small objects in UAV aerial images by exploiting multi-class semantic relationships in the feature space based on an improved YOLOv8n architecture. To enhance the model’s perception of small targets, this study optimizes the detection hierarchy by replacing the P5 feature level with the higher-resolution P2 feature map, preserving fine-grained spatial details critical for small target localization. A Spatial-Channel Passthrough Downsampling module is introduced to losslessly reorganize spatial information into the channel dimension, effectively enlarging the receptive field while maintaining lightweight characteristics. Furthermore, through the introduction of the Skew Intersection over Union loss function and the design of a Scale-Aware Multi-Class Contrastive Regularization (SMCR) method incorporating class-specific prototypes, scale-aware weighting, and hard negative mining, the precision and robustness of both bounding box regression and multi-class discrimination are further enhanced. Experimental results demonstrated that SMC-YOLO achieved a mAP50 of 38.2% on the VisDrone dataset, with a precision of 50.3% and recall of 35.6%. Compared to the YOLOv8n baseline, SMC-YOLO achieved a 5.0 percentage point improvement in mAP50, and it reduced the parameter count by 73.3% compared to the baseline. The results indicated that the proposed network accurately detected small objects under complex UAV aerial scenarios, contributing a lightweight and effective solution for real-time object detection on resource-constrained UAV platforms.
科研通智能强力驱动
Strongly Powered by AbleSci AI