基于深度神经网络的弱监督信息细粒度图像识别

Fine-grained image recognition of weak supervisory informationbased on deep neural network

强监督识别算法需要大量的人工标注信息,消耗较多的人力物力资源。为了解决上述问题,满足实际需求,提出了两种基于弱监督信息图像识别方法用于细粒度图像分类(FGVC)。一种是联合残差网络和Inception网络,通过优化卷积神经网络的网络结构提高捕捉细粒度特征的能力。另一种是对双线性CNN模型进行改进,特征提取器选取Google提出的Inception-v3模组和Inception-v4模组,最后把不同的局部特征汇集起来进行分类。通过在CUB200-2011鸟类公开数据集和Stanford Cars汽车类型数据集上进行测试,实验结果表明,提出的方法在两种数据集上的分类精度分别到达了88.3%和94.2%的分类精度,实现了较好的分类性能。

Strong supervisory recognition algorithm requires a large amount of annotation information and consumes a lot of manpower and material resources. In order to solve the above problems and meet the practical requirements, two image recognition methods based on weak supervisory information are proposed for fine-grained vision classification(FGVC). One is the combination of ResNet and Inception network, which improves the ability of capturing fine-grained features by optimizing the network structure of convolutional neural network. The other is to improve the Bilinear CNN model, feature extractor selects Inception-v3 module and Inception-v4 module proposed by Google, and finally gathers different local features for classification. The experimental results on CUB200-2011 and Stanford Cars fine-grained image datasets show that the proposed method achieves classification accuracy of 88.3% and 94.2% on the two data sets, and achieves better classification performance.