基于显著性融合的细粒度图像分类方法研究

Research of Improved Fine-Grained Image Classification Based on Saliency

摘要: 针对细粒度图像存在的类内差异大、类间差异小和依赖数据标注的问题，提出了一种基于显著度融合改进细粒度图像分类的算法。该算法基于一种双输入的深度神经网络，包括显著性特征融合结构和特征提取网络两个部分。首先，根据Fusion层网络结构将原RGB图与显著图进行特征融合，显著图是由SALICON显著性检测算法计算产生；其次，为充分利用更高分辨显著特征的调制潜力，利用最大池化操作对数据空间进行降维操作；最后，借助迁移学习思想，把在ImageNet数据集上预训练好的深度神经网络模型Inception_V3.0作为基础特征提取模型，进一步提取高层语义特征。在公开数据集CUB200-2011和Stanford Dogs中进行对比实验，结果表明，该算法的分类准确率分别达到84.36%、84.94%，相较于Part R-CNN、LRBP等多个主流细粒度分类算法，本文方法能取得更好的分类效果。

Abstract: In view of large intraclass differences, small differences between classes and the problems of dependency on data annotation in fine-grained images, an algorithm based on saliency fusion to improve fine-grained image classification is proposed. This paper introduced a two-input deep neural network, which integrated two components in a single framework: the salient feature fusion structure and the feature extractor. Firstly, the SALICON saliency detection algorithm is used to generate the saliency map. The original RGB image is fused with the saliency map according to the fusion network structure. Secondly, in order to make full use of higher resolution, the modulation potential of the salient features, maximum pooling operation is used to reduce the dimensionality of the data space so that the modulation potential of higher resolution salient features can be fully utilized. Finally, with the help of migration learning, the deep neural network model Inception_V3.0 pretrained on the ImageNet dataset is used as the basic feature extraction model to extract high-level semantic features. The comparison experiments in the public datasets CUB200-2011 and Stanford Dogs show that the classification accuracy of the algorithm is 84.36%, 84.94%, compared with Part R-CNN, LRBP and other mainstream fine-grained classification algorithms, this method can achieve better classification results.