基于显著性增强和迁移学习的鱼类识别研究

针对水下鱼类无法快速准确识别的难点,提出一种具有图像主体自动增强功能的鱼类迁移学习方法。该方法将鱼类RGB图像转换至Lab颜色空间后,利用中央周边算子计算得到整个输入图像的显著性值,进而提供鱼类目标的潜在区域,并结合GrabCut算法获取鱼类分割图像,最终将融合分割图的原始图像送入优化后的残差网络中进行训练。通过对23种鱼类进行识别试验,结果显示,固定ImageNet数据集上ResNet-50预训练模型的conv1层和conv2层参数,微调高层参数的方法能够取得最好的识别效果,且在公开的Fish4Knowledge数据集上,该模型取得了最高的识别准确率,平均识别精度达到99.63%。与其他卷积神经网络方法的对比结果显示,本方法在Fish4Knowledge和Fish30Image数据集上的识别精度和时间性能均具有较大优势,其中识别准确率至少提升4.98%。多个数据集上的试验验证了模型的有效性。     

基于深度学习的鱼类识别相关技术研究现状及展望

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。     

基于DCGAN数据增强的水产动物分类方法

针对公开大规模水产动物数据集少、人为采集数据工作量大以及传统数据增强方法对数据的特征提升有限的问题,提出一种基于深度卷积生成对抗网络的数据增强方法用于水产动物图像识别。首先,使用深度卷积生成对抗网络(DCGAN)对样本数据进行增强,然后分别使用VGG16、InceptionV3、ResNet50这三个训练模型,以微调的方式,对样本进行训练、识别。结果显示,所提出的方法在水产动物数据集上,与非生成式的数据增强方法相比,在3种模型上分类的准确率可分别提高9.8%、2.7%、1.2%。试验证实,DCGAN可有效增强水产动物图像数据,提高深度神经网络模型对水产动物图像分类的准确率。     

基于图像特征的模板匹配成鱼多目标识别算法研究与程序设计

为解决渔业养殖及转运自动计数要求，提出一种基于机器视觉的鱼类识别算法设计与程序编制，以证明使用数字图像识别分类技术的尺度不变特征变换SIFT及加速稳健特征SURF算法等可有效地检测与标注鱼类图像特征点。采用快速近邻匹配FLANN匹配算法测试了基于图像特征的鱼类旋转和泛化目标检测，得出SURF特征对个体检测效果好、SIFT特征对泛化目标检测效果优的结果。考察FLANN图像匹配碎片化特性、结合图像信息区域聚集实际、借鉴模板检测方法，设计了图像分割扫描及特征匹配的模板检测算法，并使用最大稳定极值区域MSER方法对识别结果进行冗余排除，达到能正确识别多目标鱼类的算法预设目标。研究发现，该算法及软件能成功识别图片中的多个鱼类目标，测试效果好，有较强的实用意义。     

基于成像声呐的鱼类三维空间分布

针对海洋牧场中鱼群的三维空间分布问题,本研究提出一种利用成像声呐进行位置计算的方法。将成像声呐固定在船舷外侧的水下,并保证波束发射方向和声呐移动方向一致,通过走航的方式采集水下鱼群信息。首先对采集的原始数据进行图像处理,包括图像构建、背景去除、目标提取等,然后利用基于交互式多模型联合概率数据关联算法对水下目标进行关联处理,得到同一个目标在声呐水平视场中不同帧图像中的对应关系,在此基础上根据连续两帧图像中目标位置关系计算目标的空间坐标,最后结合关联算法获得多目标在三维空间中的运动轨迹以及深度分布情况。研究表明,本方法可以有效获取鱼群在水下的三维运动轨迹及其分布情况,这将为鱼类行为分析以及海洋牧场的资源评估提供技术支持。     

基于深度学习的鱼类智能识别系统的设计与实现

中国拥有种类繁多的鱼类,外形是其分类的重要依据。但目前主要采用人工识别方法进行分类,为解决鱼类人工识别存在的问题,提出一种基于深度学习的鱼类智能识别系统的设计,以实现对中国1 400种鱼类的智能识别。系统首先采用卷积神经网络的Efficient模型,将含有1 400种鱼类,50万张鱼类图片的数据集进行训练,最终得到的模型识别精度达到了95%,单张图片识别时间仅为0.2 s,模型大小为74.5 MB。系统前端使用微信小程序,后端采用Spring+SpringMVC+Mybatis的SSM架构,调用训练好的模型文件进行识别,实现了鱼类识别、页面呈现、统计分析和相邻种类推荐等功能。该系统所提出的设计和实现方法对鱼类智能识别技术在移动端的应用提供了一种可行的思路,对海洋科研人员和开发人员有一定的借鉴作用。     

基于迁移学习和金字塔卷积网络的河蟹个体图像识别方法研究

针对目前河蟹追溯成本高、消费者无法细粒度地追溯单体河蟹信息等问题,提出一种基于迁移学习和金字塔卷积的河蟹背甲图像个体识别算法。该算法使用金字塔卷积层替换普通残差卷积块构建网络模型,可以从蟹背图像中提取多尺度、深层次的特征信息。结果显示:采用金字塔卷积结构的Resnet34和Resnet50的准确率分别为98.38%、98.51%,与使用普通卷积层的模型相比,准确率提升5.49%、1.3%,而当模型深度达到101层时,模型性能不再明显提升。与使用金字塔卷积结构的全新学习模型相比,迁移学习方法的训练收敛迭代轮次从20轮降低至5轮,此时模型准确率为98.88%,较全新学习的准确率提升0.37%,同时弥补了样本量较少的问题。该研究为河蟹个体识别追溯提供了理论依据和技术支持。     

基于鱼体背部弯曲潜能算法的四种主养鱼类识别

以四种主养淡水鱼鲫(Carassius auratus)、草鱼(Ctenopharyngodon idellus)、团头鲂(Megalobrama amblycephala)、鲤(Cyprinus carpio)为例,提出一种基于图像处理技术的鱼种类自动识别方法。首先通过鱼体信息采集系统获取待识别鱼体图像,并对其进行灰度化与二值化,得到鱼体轮廓信息;然后采用邻域边界算法对鱼体的轮廓进行提取,根据轮廓曲线建立鱼体背部轮廓数学模型;最后根据轮廓模型,采用鱼体背部弯曲潜能算法对不同种类鱼体样本的背部弯曲潜能值进行计算和聚类统计,得到不同鱼类样本的背部弯曲潜能值分布区间,从而通过比对待识别鱼体背部弯曲潜能值的区间实现对不同鱼类的自动识别。对四种主养鱼类的实验测试结果表明,对团头鲂的识别精度可以达到100%,对鲫、鲤和草鱼的识别精度达到96%,基本上能准确实现四种鱼体的分类识别,具有较好的实际应用价值。     

基于TensorFlow的水族馆鱼类目标检测APP开发

近年来深度学习在图像识别研究中取得突破进展,带动了目标检测技术的快速发展。利用目标检测技术开发水族馆鱼类目标检测APP,可以增强游客参观体验,提升科普效果。针对水族馆拍摄的80种鱼类,首先,使用LabelImg软件进行目标标记,再利用标记的目标导出成tfrecord数据;其次,选择ssd_mobilenet_v1模型进行数据训练,通过20万次的迭代训练获取到鱼类目标检测模型;最后,利用TensorFlow多目标检测API调用模型,定义2个接口和12个类,开发出Android系统手机APP。经过80种鱼类1620张图片测试,正确率为92.59%,华为MHA-AL00手机目标检测平均时间40 ms。使用鱼类目标检测APP,能实现水族馆鱼类快速识别、多鱼类目标实时检测,可提升游客的参观体验,辅助科普量化评价。     

改进的DeepLabCut鱼类游动轨迹提取

针对现有的鱼类游动轨迹提取方法在提取效率和准确率方面不能同时兼顾的问题,提出了一种改进的DeepLabCut方法用于鱼类背部关键点识别和定位。首先,选择了轻量级卷积神经网络模型EfficientNet-B0作为DeepLabCut的主干网络模型,用于提取鱼类背部关键点的特征,为了增强EfficientNet-B0的表征能力,在网络模型中引入了改进的CBAM(Convolutional Block Attention Module)注意力机制模块,将CBAM中的空间注意力模块和通道注意力模块从原来的串行连接方式改为并行连接,以解决两种注意力模块之间因串行连接而导致的互相干扰问题。其次,基于MSE(Mean Squared Error)损失函数提出了一种分段式损失函数H＿MSE用于模型的训练,分段式损失函数H＿MSE相对于传统的损失函数具有较强的鲁棒性,其在处理数据中的异常值时能表现出较低的敏感性。最后,采用了半监督学习方法对关键点进行自动标记来减少人工标记数据时产生的误差。结果显示：相比于DeepLabCut原始算法,识别误差RMSE(Root Mean Squared Error)平均...     

Deep learning-based appearance features extraction for automated carp species identification

Fish species identification is vital for aquaculture and fishery industries, stock management of water bodies and environmental monitoring of aquatics. Traditional fish species identification approaches are costly, time consuming, expert-based and unsuitable for large-scale applications. Hence, in this study, a deep learning neural network as a smart, real-time and non-destructive method was developed and applied to automate the identification of four economically important carp species namely common carp (Cyprinus carpio), grass carp (Ctenopharingodon idella), bighead carp (Hypophtalmichthys nobilis) and silver carp (Hypophthalmichthys molitrix). The obtained results proved that our approach, evaluated through 5-fold cross-validation, achieved the highest possible accuracy of 100 %. The achieved high level of classification accuracy was due to the ability of the suggested deep model to build a hierarchy of self-learned features, which was in accordance with the hierarchy of these fish’s identification keys. In conclusion, the proposed convolutional neural network (CNN)-based method has a single and generic trained architecture with promising performance for fish species identification.     

基于深度学习的青海湖裸鲤产卵场遥感识别方法

鱼类产卵场会随着外部环境条件的改变而发生变化, 因此, 快速、有效地定位鱼类产卵场对于开展水生生物资源调查、珍稀水生动物保护等工作具有重要的意义。本研究基于无人机航拍影像和青海湖裸鲤(Gymnocypris przewalskii)产卵场实地调查结果, 构建了深度学习模型, 以分析将深度学习模型应用于青海湖裸鲤产卵场识别中的可行性。模型训练交并比精度和像素精度分别为 0.870 和 0.996, 验证交并比精度和像素精度分别为 0.648 和 0.985, 虽然精度低于一般的遥感影像或图像分割精度, 但从测试的结果来看, 深度学习模型可以识别到约 79%的产卵场, 但尚不能精确地分割出产卵场, 可以作为一种辅助手段, 应用到青海湖裸鲤产卵场的识别中。     

Shrimp egg counting with fully convolutional regression network and generative adversarial network

Accurate egg counting is the basic demand in the hatcheries of the aquaculture. However, the time-consuming, fallible manual counting is now still adopted in most situations. Also, many traditional automatic egg counting methods cannot provide enough high efficiency and accuracy, especially in densely-distributed cases. In this paper, we propose a novel convolutional neural network (CNN) based method for shrimp egg counting. Compared to traditional methods mainly based on contour detection or image segmentation, the proposed method exploits the density map regression and is more efficient in densely-distributed case even with severe occlusion. Firstly, a new dataset of the redclaw crayfish Cherax quadricarinatus eggs is collected, which includes 450 images with about 272,000 eggs annotated accurately. Also, a synthetic dataset generation method based on generative adversarial network (GAN) is proposed to avoid the onerous manual labeling, and has potential applications in the supervised learning for the densely-distributed eggs or larvae counting. Then, considering the advantages of no manual preprocessing and fine-grained feature extraction in CNNs, a shrimp egg counting network (SECNet) based on fully convolutional regression network (FCRN) is proposed to realize the counting through regressing the input image into its density map. The test results show the average counting accuracy of the proposed SECNet can be up to 99.2 % when the SECNet is pre-trained on the synthetic dataset and finetuned on the collected dataset. Finally, a simple and cheap computer vision based counting setup is built by using three off-the-shelf devices and a convenient operation program integrated with the SECNet is developed for a person computer, which provides an accurate, real-time, and highly-efficient egg counting way.     

Evaluating fish feeding intensity in aquaculture with convolutional neural networks

This paper presents a novel method to evaluate fish feeding intensity for aquaculture fish farming. Determining the level of fish appetite helps optimize fish production and design more efficient aquaculture smart feeding systems. Given an aquaculture surveillance video, our goal is to improve fish feeding intensity evaluation by proposing a two-stage approach: an optical flow neural network is first applied to generate optical flow frames, which are then inputted to a 3D convolution neural network (3D CNN) for fish feeding intensity evaluation. Using an aerial drone, we capture RGB water surface images with significant optical flows from an aquaculture site during the fish feeding activity. The captured images are inputs to our deep optical flow neural network, consisting of the leading neural network layers for video interpolation and the last layer for optical flow regression. Our optical flow detection model calculates the displacement vector of each pixel across two consecutive frames. To construct the training dataset of our CNNs and verify the effectiveness of our proposed approach, we manually annotated the level of fish feeding intensity for each training image frame. In this paper, the fish feeding intensity is categorized into four, i.e., ‘none,’ ‘weak,’ ‘medium’ and ‘strong.’ We compared our method with other state-of-the-art fish feeding intensity evaluations. Our proposed method reached up to 95 % accuracy, which outperforms the existing systems that use CNNs to evaluate the fish feeding intensity.     

Research of dissolved oxygen prediction in recirculating aquaculture systems based on deep belief network

Recirculating aquaculture has received more and more attention because of its high efficiency of treatment and recycling of aquaculture wastewater. The content of dissolved oxygen is an important indicator of control in recirculating aquaculture, its content and dynamic changes have great impact on the healthy growth of fish. However, changes of dissolved oxygen content are affected by many factors, and there is an obvious time lag between control regulation and effects of dissolved oxygen. To ensure the aquaculture production safety, it is necessary to predict the dissolved oxygen content in advance. The prediction model based on deep belief network has been proposed in this paper to realize the dissolved oxygen content prediction. A variational mode decomposition (VMD) data processing method has been adopted to evaluate the original data space, it takes the data which has been decomposed by the VMD as the input of deep belief network (DBN) to realize the prediction. The VMD method can effectively separate and denoise the raw data, highlight the relations among data features, and effectively improve the quality of the neural network input. The proposed model can quickly and accurately predict the dissolved oxygen content in time series, and the prediction performance meets the needs of actual production. When compared with bagging, AdaBoost, decision tree and convolutional neural network, the VMD-DBN model produces higher prediction accuracy and stability.     

A modified YOLOv3 model for fish detection based on MobileNetv1 as backbone

This paper proposes a new approach combining YOLOv3 with MobileNetv1 for fish detection in real breeding farm. The feature maps of MobileNet are reselected as per their receptive fields for better fish detection instead of fixed chosen strategy in the original YOLOv3 framework. A set of fish image data acquired in breeding farm is used to evaluate the proposed method. The high accuracy of detection results is achieved to confirm the effectiveness of the proposed method. Furthermore, taking the place of “ImageNet”, a slighter dataset including fish images with 16 species for backbone network pretraining is picked out from “ImageNet” to extract fish features. On this basis, the effect of detection of the model is further improved due to that the extracted features are more closed to fish objects. Therefore, the proposed method is proved to have the capability of providing necessary and accurate number of fish, which will then be used to determine the breeding actions accordingly.     

鱼目标三维信息采集平台中干扰现象的判定与消除

为了保证鱼类行为数据采集的准确性和鲁棒性,本实验提出了一种干扰消除算法,对鱼目标三维定位算法进行优化。主要针对鱼目标三维信息采集平台中出现的一种由全反射引起的干扰现象,利用光学原理分析其出现原因,总结干扰现象特点,之后通过基于轮廓的运动目标跟踪方法,设计干扰消除算法,优化鱼目标三维定位算法,实现对干扰现象影响的消除,提高定位精度。通过对红鲫和仿真鱼的定位实验,结果发现,本实验提出的去干扰优化算法可以有效消除干扰像对于目标定位的影响,准确获取鱼目标的三维坐标,保证鱼目标三维定位跟踪的准确性和鲁棒性,提高鱼目标行为信息采集精度。