多尺度知识引导局部增强的小样本细粒度图像分类方法

Multiscale knowledge-guided local enhancement method for few-shot fine-grained image classification

小样本细粒度图像分类任务中,由于支持样本与查询样本之间缺少局部关联性,导致图像关键可区分区域不易精确定位。针对这一问题,提出了多尺度知识引导局部增强的小样本细粒度图像分类方法,采用图像金字塔向下降低采样率获得多个不同分辨率的子图作为输入图像,融合多尺度特征,丰富了样本信息,利用知识引导模块捕获支持样本和查询样本之间的语义相关性,增强了支持样本重要区域的特征表示,对支持样本和查询样本的嵌入特征图进行克罗内克积操作,生成的空间相关图,更精确地定位样本特征之间的位置对应关系,突出了辨别性区域。实验结果表明该方法在小样本细粒度图像分类任务中表现出较好的分类性能。

In few-shot fine-grained image classification tasks,the absence of local correlation between support and query samples causes difficulty in precisely identifying discriminative regions in images.To solve this problem,this study proposes a multiscale knowledge-guided local enhancement method for few-shot fine-grained image classification.Multiple sub-images with different resolutions are obtained as input images by reducing the sampling rate downward using the image pyramid and by integrating multi-scale features to enrich the sample information.A knowledge-guided module is employed to capture semantic correlations between support and query samples,which enhances the feature representation of important regions in support samples.Kronecker product operations are performed on the embedded feature maps of support and query samples to generate spatial correlation maps.This way enables more accurate localization of the positional correspondence between sample features and highlights the discriminative area.Experimental results confirm the robust classification performance of the method in the classification task of few-shot fine-grained images.