期刊文献+

双路特征提取与度量的少样本细粒度图像分类方法

Dual-Path Feature Extraction and Metrics for Few-Shot Fine-Grained Image Classification
下载PDF
导出
摘要 少样本学习旨在利用少量数据训练深度学习模型,并将其快速泛化到新任务中.在这一领域,少样本细粒度图像分类是最具有挑战性的任务之一,原因在于细粒度图像具有类内方差大、类间方差小的特点.为了解决这一问题,本文提出了一种基于距离与方向双重度量的神经网络,分别利用欧氏距离衡量特征间的绝对距离差异和余弦相似度衡量特征间的相对方向差异,以提升度量信息多样性和样本特征的判别性.同时,为了与当前先进的少样本细粒度图像分类方法对比,将特征提取器在不增加深度的前提下设置为双路形式,以适应不同度量方法对嵌入特征信息的需要.此外,设计了彼此分离的通道和空间注意力机制,分别通过自适应通道注意力和空间信息交叉注意力对不同阶段的提取特征进行增强,从而挖掘重要分类信息.最后,通过双相似度模块分别计算两种差异信息的度量结果,并选取一定权重融合得到最终的相似度分数,实现绝对差异与相对差异在度量空间中的协调补充.在4个主流细粒度图像分类数据集上进行实验对比与分析,最终结果表明了所提方法在相同设置下最多实现了7.0%左右的分类准确率提升. Few-shot learning aims at training deep-learning models with limited data and then quickly generalizing them to new tasks.Few-shot fine-grained image classification is a highly challenging task in this field,primarily due to the large intraclass and small interclass variance of fine-grained images.To address this issue,the twins of distance-direction metric network is proposed,which uses Euclidean distance and cosine similarity to measure the absolute distance difference and the relative direction difference among features,respectively,thereby improving the diversity of metric information and discrimination of sample features.Furthermore,the feature extractor is equipped with a dual-path output without the added depth to meet the embedding feature information demand for different metric methods.This ensures that the feature extractor remains competitive with advanced few-shot fine-grained image classification methods currently available.Moreover,the separate channel and spatial attention mechanisms are designed to enhance the extracted features at different stages,wherein the important classification information is mined via adaptive channel attention and spatial-aggregation cross attention,respectively.Finally,the twin-similarity modules calculate the metric results of the two different pieces of information and fuse them into the final similarity scores by selecting specific weights,thereby realizing the coordination and complement between absolute and relative differences in the metric space.The experimental contrast and analyses were conducted on four benchmark fine-grained datasets,and the results demonstrate that the proposed method enhances the classification accuracy by up to 7.0%under the same settings.
作者 冀中 吴伊兵 王轩 Ji Zhong;Wu Yibing;Wang Xuan(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处 《天津大学学报(自然科学与工程技术版)》 EI CAS CSCD 北大核心 2024年第2期137-146,共10页 Journal of Tianjin University:Science and Technology
基金 国家自然科学基金资助项目(62176178).
关键词 细粒度图像 少样本 欧氏距离 余弦相似度 fine-grained image few-shot Euclidean distance cosine similarity
  • 相关文献

参考文献2

二级参考文献4

共引文献7

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
使用帮助 返回顶部