基于卷积语义增强的细粒度图像分类方法

Fine-Grained Image Classification Method Based on Convolutional Semantic Enhancement

细粒度图像分类是指基于已划分的基本类别而进行的更细粒度的子类别划分。由于细粒度图像分类具有类间差异小、类内差异大的数据特征,使其成为了一项非常具有挑战性的研究任务。通过对现有细粒度图像分类算法和模型的分析研究,提出一种基于卷积增强多尺度特征语义的弱监督细粒度分类方法。该方法通过卷积关联高低层特征,运用高层特征语义突出底层有意义的特征,抑制语义无效的底层特征,进而获得更具表达能力的多尺度特征。在以ResNeXt-101网络作为骨干网络和特征提取网络的基础上,在3个常用的细粒度图像数据集上对该方法进行实验验证,取得的分类正确率分别为88.3%、93.7%和94.3%。实验结果表明,与增强全局特征子特征的语义方法SEF、采用并行卷积块的多层特征融合方法MFF等其他多个主流细粒度分类算法相比,所提方法取得了更好的分类效果。

Fine-grained image classification refers to the finer-grained subcategory division based on the divided basic categories.Due to the data features of small inter class differences and large intra class differences in fine-grained image classification,it has become a very chal⁃lenging research task.Through the analysis and research of existing fine-grained image classification algorithms and models,a weakly super⁃vised fine-grained classification method based on convolution-enhanced multi-scale feature semantics is proposed.This method correlates high-level and low-level features through convolution,uses high-level feature semantics,highlights underlying meaningful features,sup⁃presses low-level features with invalid semantics,and obtains multi-scale features with more expressive capabilities.Based on the ResNeXt101 network as the backbone network and the feature extraction network,the method is verified experimentally on three commonly used finegrained image datasets,and the classification accuracy rates are 88.3%,93.7%and 94.3%respectively.Experimental results show that this method achieves better classification results than other mainstream fine-grained classification algorithms such as the semantics method(SEF)which enhances the sub-features of global features,and multi-layer feature fusion method(MFF)which uses parallel convolution blocks.