基于视觉特征低维嵌入的细粒度图像分类
Fine-grained Image Classification with Low-dimensional Visual Feature Embedding
-
摘要: 为了减少原图像特征空间中高维数据的冗余,解决细粒度数据分布在特征空间中无法线性可分的问题,提出一种结合视觉特征低维嵌入和非线性映射的细粒度图像分类算法.首先将视觉特征嵌入到低维空间来减少冗余数据对分类造成的干扰,提高分类模型对测试数据的泛化能力;然后通过基于排序的目标函数来训练多个线性分类器,建立类别和低维视觉嵌入之间的非线性关系,有效地区分不同类别的细粒度样本之间的细微差异.实验结果表明,该算法有效地改进了现有的细粒度图像分类方法,显著提高对未知测试样本的分类精度.Abstract: In order to reduce the redundancy of high-dimensional data in the original image space and solve the problem that the distribution of fine-grained data in the feature space cannot be linearly separable, this paper proposes a novel fine-grained image classification method combining low-dimensional visual embedding and non-linear mapping. Firstly, the visual features are embedded into low-dimensional space to reduce the inference caused by the redundant data on the classification and improve the generalization ability of the classification model on the test data. Then, a number of linear classifiers are trained with a ranking based objective function, and the non-linear relationship between the class label and low-dimensional visual embedding is established to effectively distinguish the nuances between the samples of different fine-grained classes. The experimental results demonstrate that, the proposed method improves the classification accuracy on unknown samples and performs better than the existing fine-grained image classification algorithms under the zero-shot setting.