基于迁移学习与模型融合的犬种识别方法

Dog breed identification method based on transfer learning and model fusion

犬种识别研究属于细粒度图像分类的典型代表,使用传统图像分类方法与普通卷积神经网络进行犬种识别,会出现准确率普遍很低等问题。本文提出了一种将迁移学习与模型融合相结合的方法。通过运用四种常用的卷积神经网络模型分别进行部分图像的特征提取,选取表现最佳的两种模型Inception_v3以及Resnet152_v1进行双模型融合,将得到的融合网络用于犬种图像进行迁移学习训练。针对120类犬种图片,训练得到了验证集精度可达93.02%的网络模型。同时考虑将测试集图片经过YOLO目标检测算法识别,定位目标区域后再送入网络,实验结果表明该方法在融合模型中能进一步提高犬种识别检测精度。

Dog breed identification research is a typical representative of fine-grained image classification. The accuracy of recognition is generally low due to the use of traditional image classification and common convolutional neural networks for dog breed identification studies. This paper proposed a method to combine transfer learning with model fusion. Firstly,the method extracted features of partial images by using four convolutional network models separately,the two best performing models--Inception_v3 and Resnet152_v1,are selected for dual model fusion. The fusion network was applied to the dog breed image for transfer learning training. For the 120 kinds of dog breed pictures,the network model obtain the accuracy of 93. 02% on the verification dataset. Meanwhile,the testset was targeted by the YOLO object detection algorithm and cropped into the network. The result shows that the method can further improve the dog species classification accuracy through the model.