计算机科学
目标检测
计算机视觉
人工智能
算法
模式识别(心理学)
作者
Shicong Cheng,Keke Geng,Zhichao Liu,Guodong Yin
出处
期刊:
日期:2024-07-28
卷期号:: 8595-8600
被引量:1
标识
DOI:10.23919/ccc63176.2024.10662387
摘要
UAV aerial images possess several characteristics, including large size, high resolution, tiny and dense objects. Existing algorithms may not meet the requirements for detection accuracy. To address these issues, this paper proposes a small object detection (SOD) algorithm model, which is an improvement over YOLOv5s. Since the model adopts SAHI, improved CBAM and EIOU techniques, we call it as SCE-YOLO. The proposed algorithm first performs a slicing operation on the input image, which enlarges the detection area without increasing the model’s complexity. Secondly, the network introduces an improved attention mechanism called Res_CBAM to enhance the feature extraction ability. To solve the problem that the general loss function is sensitive to small objects, the loss function of NWD_EIOU is modified. And finally, a small object detection layer is added to the traditional YOLOv5s network structure to deepen the combination of shallow semantics and deep features. The experimental results show that SCE-YOLO improves the mAP to 41.0% compared to the YOLO v5s algorithm, which is 6.4% higher than the baseline.
科研通智能强力驱动
Strongly Powered by AbleSci AI