Abstract: In order to solve the problems such as lack of large-scale aquatic animal data set, heavy workload of artificial data collection and limited feature enhancement of data by traditional data enhancement methods, a data enhancement method based on deep convolutional generative adversarial network was proposed for aquatic animal image recognition. First of all, deep convolutional generative adversarial network (DCGAN) was used to enhance the sample data, and then three training models: VGG16, InceptionV3 and ResNet50 were used to train and identify the samples by fine tuning. The results show that the classification accuracy of the proposed method on the aquatic animal data set can be improved by 9.8%, 2.7% and 1.2% respectively compared with the non-generated data enhancement method. The experiment proves that DCGAN can effectively enhance the image data of aquatic animals and improve the accuracy of deep neural network model in classifying aquatic animal images.

Key words:  deep convolutional generative adversarial network, data enhancement, classification of aquatic animals
 

摘要： 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，在三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。

关键词: 深度卷积生成对抗网络, 数据增强, 水产动物分类