基于深度学习的鱼类识别相关技术研究现状及展望

Review and prospect of fish recognition and related techniques based on deep learning

为促进渔业生产智能化、现代化发展,综述了基于深度学习的鱼类识别相关技术。首先,从数据集构建、数据预处理、神经网络模型设计以及模型训练等4个方面阐述了基于深度学习的鱼类识别工作流程。然后,从图像分类、目标检测、图像分割3个角度总结了近几年鱼类识别相关技术的研究进展及应用成果。其中,图像分类主要用于识别个体鱼的色泽与种类,目标检测侧重于估计鱼群的数量和体型,而图像分割则在推断鱼类的状态和行为方面发挥着重要作用。同时,分析了不同方法所具备的优势,比较了各方法在数据集中的性能指标。最后,对深度学习在鱼类识别领域的下一步发展方向和研究重点进行了展望。综上,深度学习方法效率普遍较高、泛化能力普遍较强,深度学习技术在鱼类识别中的广泛应用能够为渔业科研人员提供有效的技术支撑。

Deep learning technology is an important research field of artificial intelligence,which can simulate the operation of neurons in human brain,build network models,and complete various learning tasks such as image recognition.In marine fisheries,deep learning technology can be used to classify fish species,detect fish population,calculate fish size and infer fish behavior,it can provide strong support for the intellectualization and modernization of fishery production.Therefore,it is of great significance to summarize the research on fish recognition techniques based on deep learning.The workflow of deep learning in fish recognition is divided into four parts:data set construction,data preprocessing,neural network model design and model training.The data set construction is the most important part of fish identification,and the size and quality of the samples determine the performance of the model directly.Data preprocessing is to use data enhancement techniques to expand data sets,increase sample diversity and help improve model generalization capabilities.The neural network model determines how the fish image is extracted.Common neural network model is composed of convolution,pool,and full connection network.The model training is based on deep learning framework and the artificial optimization method and loss function training model to fit the function mapping behind the application of fish recognition.The application of deep learning in fish recognition includes image classification,object detection and image segmentation.Fish image classification is applied to the study of fish diversity,and fish image with a single sample is classified by depth network.In general situation,using general networks such as VGG and ResNet can achieve good classification results.In complex scenes,removing background interference,adding attention mechanism,using transfer learning and combining with traditional pattern recognition can effectively improve classification accuracy and enhance model generalization ability.Fish object detection is the generation of categories and detection frames for multiple fish contained in an image,which is often used in the study of fish size and quantity,the accuracy of model detection is mainly improved from image quality and feature information enhancement.Underwater image enhancement technology is often used to reduce the interference of background and underwater impurities.Feature information enhancement is used to improve the ability of the network model to distinguish the classifiable features of fish,including the introduction of feature fusion module,attention mechanism,integrated network model and so on.Fish image segmentation is the process of labeling fish pixels with the same properties.Fish body parts can be separated by this method,and then the body length and weight can be calculated.Image segmentation often uses encoder and decoder networks,in which we can introduce additional monitoring loss function to improve the segmentation accuracy,but also can introduce the attention mechanism to improve the segmentation performance.^The experimental results in three application directions of fish recognition show that the deep learning method has high efficiency and strong generalization ability.The deep learning technology can provide reference for fishery researchers and better meet the needs of fishery modernization development.In the future,we should pay more attention to the construction of fish data sets for unified standards,and to the speed of deep learning model reasoning and the reduction of hardware power consumption,so as to achieve the best application effect.