计算机视觉
目标检测
人工智能
计算机科学
对象(语法)
图像分割
航空影像
图像处理
可视化
边缘检测
特征提取
对象类检测
噪音(视频)
机器视觉
像素
变更检测
图像分辨率
数学形态学
特征(语言学)
特征检测(计算机视觉)
信号处理
人脸检测
角点检测
视觉对象识别的认知神经科学
迭代重建
稳健性(进化)
图像(数学)
模式识别(心理学)
作者
Fan Long,Jing Zhang,Rui Xia,Jianwen Huo,Aimin Wang,Heng Liu
标识
DOI:10.1117/1.jei.34.5.053029
摘要
To address key challenges in small object detection for unmanned aerial vehicle (UAV) imagery, such as illumination variations, insufficient representation of shallow-layer features, and the complexity of multiscale object detection, we propose an enhanced Mamba-YOLO model, termed IM-YOLO. Built upon the conventional Mamba-YOLO architecture, IM-YOLO incorporates an image preprocessing enhancement module to adaptively optimize image quality under varying illumination conditions. In addition, a spatial-channel synergistic attention is integrated to enhance the feature representation of shallow-layer small objects while suppressing background interference. A feature complementary fusion module is designed to dynamically integrate high-resolution details with global semantic features, significantly improving detection robustness in dense multiscale scenarios. On the VisDrone2019 dataset, IM-YOLO achieves mean average precision (mAP) values of 53.7% and 33.6%, representing improvements of 3.5% and 2.7% over the baseline model, respectively, while maintaining an inference speed of 187.3 frames per second to meet real-time detection requirements. On the DIOR and UAVDT datasets, IM-YOLO achieves mAP values of 86.4% and 65.3%, and 51.1% and 31.5%, respectively. Experiments across multiple datasets demonstrate that the proposed model exhibits superior detection performance compared with existing mainstream methods. Furthermore, detection experiments using real-world UAV imagery confirm IM-YOLO’s superior small object detection performance across diverse complex scenes, validating its robustness and adaptability in practical settings. These findings establish IM-YOLO as an efficient and viable solution for UAV vision tasks.
科研通智能强力驱动
Strongly Powered by AbleSci AI