基于语义分割的视频鱼类特征提取方法研究 |
A Semantic Segmented Framework for Extracting Fish Features from Videos |
投稿时间:2024-08-08 修订日期:2024-08-30 |
DOI:10.15928/j.1674-3075. 202408080298 |
中文关键词:弱监督学习 语义分割 视觉模型 鱼类特征提取 |
英文关键词:Weekly supervised learning,Semantic segmentation, Vision transformer model, Feature extraction of fish |
基金项目:国家重点研发计划(2022YFC3204200)。 |
|
摘要点击次数: 182 |
全文下载次数: 36 |
中文摘要: |
从视频图像中快速、准确提取水生生物(如鱼类)的特征信息,是信息科学与水生态研究结合的热点。基于Transformer的视觉模型,提出一种基于弱监督语义分割的视频鱼类特征提取方法,在无需预训练或微调的条件下,可以实现对鱼的身体、头部和鳍3类形态区域标签的分割提取。采用DeepFish分割数据集构建计算机视觉模型(vision transformer,ViT),通过对水下拍摄的鱼类视频进行实验,结果实现了鱼体形态主体特征的有效提取,对拟定的3类形态标签区域进行了良好的分割标记。研究方法具有较高的效率、分割准确度和标记区域的连续平滑性,可提供良好的语义特征,为人工智能技术在鱼类等水生生物监测实践应用提供一种低成本、高效率的新方法。 |
英文摘要: |
Fast and accurate extraction of information on features of aquatic organisms from video images is a research hotspot that draws from information science and ecological research. In this study, we developed a fish feature extraction method based on weakly supervised semantic segmentation and the vision transformer. Our aim was to realize the segmentation and extraction of three types of fish morphological regions (body, head, and fins) without the need for pre-training or fine-tuning. First, a self-attention model was created using a DeepFish segmentation dataset, and then applied to extract information from underwater videos of Schizothorax oconnori. Results show that the method we proposed effectively extracted the three morphological features of the test fish, appropriately segmenting, marking and labeling the three features. In general, the process is highly efficient, accurate, and smoothly labeled the semantic features. It is a low-cost, highly efficient method for the practical application of artificial intelligence technology in the monitoring of fish and other aquatic organisms. |
李潇洋,陈 健,常剑波.2024.基于语义分割的视频鱼类特征提取方法研究[J].水生态学杂志,45(5):204-212. |
查看全文 查看/发表评论 下载PDF阅读器 HTML |