基于轻量级卷积神经网络的红鳍东方鲀个体身份无损识别方法

A LIGHTWEITGH CONVOLUTIONAL NEURAL NETWORK-BASED METHOD FOR LOSSLESS IDENTIFICATION OF INDIVIDUAL TAKIFUGU RUBRIPES

文章充分利用红鳍东方鲀(Takifugu rubripes)体侧纹理特征提出了一种基于轻量级卷积神经网络的鱼类个体身份识别方法,可在无损前提下实现红鳍东方鲀个体身份的高精度识别。首先,采用SOLOv2模型进行前景分割,并结合红鳍东方鲀体型特点,通过质心和哈希值计算方法完成数据集生成和筛选;随后,从多维度分别测试主流深度学习图像分类骨干网络和不同损失函数在红鳍东方鲀身份识别中的效果;继而,在MobileNet v2骨干网络基础上,耦合Softmax Loss函数,建立了一种适用于红鳍东方鲀的个体身份无损识别的最优组合方法。研究结果表明,文章方法准确率可达90.2%,优于其他相关主流方法(准确率73.6%—89.3%),相关研究成果将为循环水养殖鱼类个体身份无损识别和精准生物量估算提供技术支撑。

In aquaculture,avoiding duplicate individual biomass estimation is an important prerequisite for achieving accurate fish biomass estimation,the key is to perform individual fish identification,while few relevant studies have been reported.In this paper,a lightweight convolutional neural network-based identification method for individual fish identity was proposed,which can achieve high accuracy identification of individual Takifugu rubripes without loss.Firstly,SOLOv2 model was used for foreground segmentation,and combined with the characteristics of the body size of Takifugu rubripes,the dataset generation and filtering were completed by the method of calculating the center of mass and Different Hash Algorithms;subsequently,the effectiveness of mainstream deep learning image classification backbone networks and different loss functions in Takifugu rubripes identity recognition were tested separately from multiple dimensions;following that,an optimal combination method for the lossless identification of individual identity of Takifugu rubripes was established based on MobileNet v2 backbone network coupled with Softmax Loss function.The results showed that the accuracy of the proposed method can reach 90.2%,which is better than other related mainstream methods(accuracy 73.6%—89.3%),and related research results will provide technical support for non-destructive identification of individual fish identity and accurate biomass estimation in recirculating water culture.