首页|多区域注意力的细粒度图像分类网络

多区域注意力的细粒度图像分类网络

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
  • 维普
目前细粒度图像分类的难点在于如何精准定位图像中高度可辨的局部区域以及其他辅助判别特征。提出一种多区域注意力的细粒度图像分类网络来解决这个问题。首先使用Inception-V3对图像特征进行提取，通过重复使用注意力擦除的方法使模型关注次要特征;然后通过背景去除以及上采样的方法获取图像更精准的局部图像，对提取到的局部特征进行位置统计，并以矩形框的方式获取图像整体，减少细节信息丢失;最后对局部与整体图像进行更加细致的学习。此外，设计联合损失函数，通过动态平衡难易样本和缩小类内差距的方法改善模型的识别效果。实验结果表明，该方法在公开的细粒度图像数据集CUB-200-2011、Stanford-Cars和FGVC-Aircraft上的准确率分别达到89。2%、94。8%、94。0%，相较于对比方法性能更优。
Multi-Region Attention Network for Fine-Grained Image Classification
The current challenge in fine-grained image classification lies in accurately identifying highly recognizable local areas and auxiliary distinguishing features within an image.To address this issue,a multi-region attention network for fine-grained image classification is proposed.The process begins with the use of Inception-V3 for feature extraction.The model is then directed to focus on secondary features via repeated application of attention erasure.Subsequently,more precise local images are generated by removing the background and employing up-sampling techniques.This is followed by the analysis of the position statistics of the extracted local features.The entire image is then represented as a rectangular box,minimizing the loss of detailed information.Further detailed learning is conducted on local and overall images.Additionally,a joint loss function is designed to enhance the model's recognition capabilities.This is realized by dynamically balancing between difficult and easy samples and reducing intra-class variance.Experimental results on the public fine-grained image datasets CUB-200-2011,Stanford-Cars,and FGVC-Aircraft demonstrate that this method can realize accuracies of 89.2%,94.8%,and 94.0%,respectively.These figures surpass those achieved by other methods.

multi-region attentionfine-grained image classificationerasure strategyjoint lossdeep learningConvolutional Neural Network(CNN)

白尚旺、王梦瑶、胡静、陈志泊

展开 >

太原科技大学计算机科学与技术学院,山西 太原 030024

北京林业大学信息学院,北京 100091

多区域注意力 细粒度图像分类 擦除策略 联合损失 深度学习 卷积神经网络

国家自然科学基金山西省自然科学基金博士科研启动基金

3207177520220302121118920202057

2024

10.19678/j.issn.1000-3428.0066426

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(1)
  • 7