Abstract: Aiming at the problems of small sample size and low detection accuracy of diseased takifugu rubripes, a detection method of diseased takifugu rubripes based on ResNet50 and migration learning was proposed. First, ResNet50 was used to pre-train the model on the ImageNet dataset; then based on the pre-training results, the takifugu rubripes detection ResNet50 network was constructed, and the model weights, which were pre-trained and contained in 16 residual blocks, were transferred to the ResNet50 network for model weight initialization to reduce the training cost; in order to further improve the accuracy of detection, a deconvolution layer was added after the last convolution layer of the constructed ResNet50 network model to learn the details of the target; finally, the data set was constructed from the images of healthy and diseased takifugu rubripes, and the data was augmented using methods such as flipping, rotation, random cropping, chromaticity changes, and adding noise to increase the diversity of data samples and improve the robustness of the detection method. Experiments were conducted on the constructed dataset, the accuracy of the detection method for disinfected puffer fish based on ResNet50 and transfer learning can reach 99%. Compared with ResNet18, ResNet34, ResNet101 and ResNet152, the detection accuracy of the proposed method was improved by 10.7%, 6.6%, 6.2% and 5.6% respectively. Compared with the ResNet50 residual network without deconvolution, the detection accuracy of the proposed method was improved by 0.4%. The results showed that the method based on ResNet50 and transfer learning could effectively solve the problems of fewer samples and low accuracy of diseased puffer fish, and provide a reference for the study of diseased puffer fish detection

Key words: ResNet50, transfer learning, data augmentation, takifugu rubripes, diseased fish detection

摘要： 针对红鳍东方鲀病鱼样本数量少、检测准确率不高等问题，提出一种基于ResNet50和迁移学习的红鳍东方鲀病鱼检测方法。首先用ResNet50在ImageNet数据集上进行模型预训练；然后基于预训练结果构建了红鳍东方鲀病鱼检测ResNet50网络，将经过预训练的、包含16个残差块的模型权重迁移到构建的ResNet50网络中进行模型权重初始化以降低训练成本；为进一步提高检测的准确性，在构建的ResNet50网络模型的最后一个卷积层后面加入反卷积层以学习目标中的细节信息；最后，用红鳍东方鲀健康鱼和病鱼图像构建了数据集，并采用翻转、旋转、随机裁剪、色度变化和添加噪声等方法进行了数据增广，以增加数据样本的多样性，进而提高检测方法的鲁棒性。在所构建的数据集上进行了试验，试验结果表明，基于ResNet50和迁移学习的红鳍东方鲀病鱼检测方法准确率可以达到99%，与ResNet18、ResNet34、ResNet101和ResNet152不同深度的残差网络相比，分别约提升了10.7%、6.6%、6.2%和5.6%，在与不加入反卷积的ResNet50残差网络相比，约提升0.4%的精度。研究表明，采用基于ResNet50和迁移学习的方法，有效地解决了红鳍东方鲀病鱼样本少和准确率不高的问题，为红鳍东方鲀病鱼检测提供了新方法。

关键词: ResNet50, 迁移学习, 数据增广, 红鳍东方鲀, 病鱼检测