| 汪鲁才,陈春江,邹伊雯,谢婷.长尾分布下的微藻显微图像轻量级目标检测[J].电子测量与仪器学报,2024,38(4):46-54 |
| 长尾分布下的微藻显微图像轻量级目标检测 |
| Lightweight target detection for micro-algae microscopicimages under long-tailed distribution |
| |
| DOI： |
| 中文关键词: 目标检测 长尾分布 延迟重采样 知识蒸馏 |
| 英文关键词:object detection long-tail distribution delayed re-sampling knowledge distillation |
| 基金项目:国家自然科学基金(62101188)、国家自然科学基金(62261001)、湖南省自然科学基金(2023JJ40450)项目资助 |
|
|
| Author | Institution |
| Wang Lucai | College of Engineering and Design, Hunan Normal University, Changsha 410081, China |
| Chen Chunjiang | College of Engineering and Design, Hunan Normal University, Changsha 410081, China |
| Zou Yiwen | College of Engineering and Design, Hunan Normal University, Changsha 410081, China |
| Xie Ting | College of Engineering and Design, Hunan Normal University, Changsha 410081, China |
|
| 摘要点击次数: 357 |
| 全文下载次数: 13012 |
| 中文摘要: |
| 微藻显微图像目标检测技术是生物学研究和环境监测等领域重要研究方向之一。电子显微镜采集到的微藻图像数据集存在长尾数据问题。传统的微藻检测方法需要大量人工操作，耗时长且结果容易受到操作人员技术经验的影响。结合解决长尾分布的方法，本文提出了一种基于延迟重采样和知识蒸馏相结合的目标检测算法（DDM-YOLO）。先对微藻显微图像进行数据增强，然后针对长尾分布数据，采用延迟重采样，并在二阶段采用反向采样，关注难以分类的少数类别样本，改善目标检测性能。设计了一种轻量级目标检测网络架构，通过知识蒸馏来减少模型复杂度和计算量。实验结果表明，DDM-YOLO算法的mAP@0.5/%为77.1%，与YOLOv5s相比提高了6.1%，模型参数量为3.88 MiB，减少了45.4%。所提出的方法在微藻显微图像数据上取得了显著的性能提升，同时在资源受限条件下实现了高效的目标检测，大大降低了检测人员的工作量。 |
| 英文摘要: |
| Microalgae microscopic image target detection technology is one of the important research directions in fields such as biology and environmental monitoring. The dataset of microalgae images captured by electron microscope exhibits a long-tail data issue. Traditional methods for microalgae detection are notoriously labor-intensive, time-consuming, and heavily influenced by operator expertise. In this context, combining methods to address the long-tail distribution, this paper proposes a target detection algorithm called DDM-YOLO, which combines delayed resampling and knowledge distillation. The approach involves data augmentation for microalgae microscopic images and utilizes delayed resampling for long-tail data. In the second stage, reverse resampling is applied to focus on the challenging minority class samples, thereby enhancing the performance of target detection. Additionally, a lightweight target detection network architecture is designed, and knowledge distillation is employed to reduce model complexity and computational requirements. Experimental outcomes reveal that the DDM-YOLO algorithm achieves an mAP@0.5/% of 77.1%, surpassing the YOLOv5s algorithm by a notable 6.1%. The model parameter size is 3.88 megabytes, a significant 45.4% decrease. This proposed method significantly enhances performance on microalgae microscopic image data and efficiently performs target detection under resource-constrained conditions, substantially reducing the workload of detection personnel. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
