基于集成迁移学习的细粒度图像分类算法

针对现有的大部分细粒度图像分类算法都忽略了局部定位和局部特征学习是相互关联的问题，提出了一种基于集成迁移学习的细粒度图像分类算法。该算法的分类网络由区域检测分类和多尺度特征组合组成。区域检测分类网络通过类别激活映射(class activation mapping,CAM)方法获得局部区域，以相互强化学习的方式，从定位的局部区域中学习图像的细微特征，组合各局部区域特征作为最终的特征表示进行分类。该细粒度图像分类网络在训练过程中结合提出的集成迁移学习方法，基于迁移学习，通过随机加权平均方法集成局部训练模型，从而获得更好的最终分类模型。使用该算法在数据集CUB-200-2011和Stanford Cars上进行实验，结果表明，与原有大部分算法对比，该算法具有更优的细粒度分类结果。

Fine-grained image classification algorithm based on ensemble methods of transfer learning

Aiming at the problem that most existing fine-grained image classification algorithms ignore the correlation between local localization and local feature learning, this paper proposes a fine-grained image classification algorithm based on ensemble methods of transfer learning. The classification network of the algorithm consists of region detection classification and multi-scale feature combination. The regional detection classification network obtains local regions by class activation mapping (CAM), and learns fine-grained features from the localized regions in a mutually reinforcing way. Finally, the local features are combined as the final feature representation to classify. The classification network combines the proposed ensemble methods of transfer learning in the training process,and ensemble the local training model by stochastic weight averaging method based on transfer learning to obtain a better classification model. Experiments on datasets CUB-200-2011 and Stanford Cars show that the algorithm has better fine-grained classification results than most of the previous algorithms.