基于YOLOv5算法对斑马鱼幼鱼的检测研究

A study on the detection of zebrafish larvae based on YOLOv5

斑马鱼幼鱼行为学分析常常是药物学研究与基因操作的直观表现,在封闭、复杂环境下对幼鱼进行检测是研究其功能特性的基本步骤。由于幼鱼属于小目标,本文通过去掉YOLOv5网络中的大尺度预测层和大、中尺度预测层,得到了YOLOv5m-sm模型以及YOLOv5m-s模型;由于没有公开的幼鱼数据集,本文使用DarkLabel标注软件将幼鱼头部作为特征标记,得到的Zebradata数据集,并按4:1的比例分为训练集与验证集,分别用来训练及验证模型;为了测试算法对幼鱼的检测能力,使用160张含有23条幼鱼的测试集对YOLOv3m、YOLOv5s、YOLOv5m-s、YOLOv5m-sm和YOLOv5m模型进行识别实验。实验结果表明,YOLOv5m-s算法具有较高的识别准确度,满足幼鱼目标检测要求。

Behavioral analysis of zebrafish larvae is often a visual representation of pharmacological studies and genetic manipulations.Therefore,the detection of zebrafish larvae in a closed and complex environment is a fundamental step in the study of their functional properties.Since the larva belong to small targets,YOLOv5m-sm model is obtained by removing the large prediction layer and YOLOv5m-s model is obtained by removing the large and medium prediction layer in YOLOv5m network.Meanwhile,since there is no publicly available larvae dataset,this paper uses DarkLabel software to label heads of the larva to obtain the Zebradata dataset that is divided into training and validation sets in a ratio of 4:1 to train and validate the model,respectively.In addition,160 images that containing 23 larvae are used in this paper to test the detection accuracy of YOLOv3m,YOLOv5s,YOLOv5m-s,YOLOv5m-sm and YOLOv5m.The experimental results show that the YOLOv5m-s has high recognition accuracy and meets the requirements for larvae detection experiments.