摘要
人工识别鱼类的方法需要直接接触鱼体,传统机器学习方法需要人工提取特征并设计图像向量化方法,而深度学习方法能够从输入数据中获得高级特征,进而挖掘数据的分布规律.使用深度学习方法识别鱼类能够解放人力、规避主观识别的倾向性、减少鱼类应激反应,对发展智慧渔业和精准养殖有积极推动作用.本研究综述了深度学习在鱼类识别任务上的应用研究,针对研究中遇到的识别问题,提出未来应推广统一的成果检测标准以明确研究方向,不断提高应用研究水平以扩展实用性更强和智能化更高的应用任务,解决设备和模型间的接口兼容问题以增强科研人员在设备和模型选择上的灵活性,以期为使用深度学习方法研究鱼类识别任务提参考依据.
Abstract
Fish recognition is essential for fisheries management and ecosystem protection measures due to provide a variety of information on the aquatic ecosystem,especially the abundance of fish resources and aquatic health.Time consuming,laborious and subjective manual fish recognition will lead to fish stress reaction after artificial contact,and is extremely unfavorable to the growth of fish.The traditional machine learning method avoids the adverse effects caused by contact with the fish directly,and still needs to manually extract the features and design the image vectorization method according to the recognition problem.However,deep learning known as end to end learning,different from the above recognition methods,is capable of obtaining high level features from the input data and mining the distribution patterns of the data,can free the labor,circumvent the tendency of subjective recognition,and minimize the stress reaction of fish,which is a positive contribution to the development of smart fisheries and precision aquaculture.The application of deep learning in fish recognition is described,and uniform recognition standard is proposed to clarify future research directions.Also,the level of applied research should be continuously improved to enhance practicability and intelligence.Addressing the interface compatibility between equipment and models is also recommended to increase flexibility for researchers in choosing tools,aiming to provide reference for researchers working on fish recognition using deep learning methods.
维普
万方数据