高级检索

融合空间信息和分布关系的东巴画小样本分类

Dongba Painting Few-Shot Classification Combining Spatial Information and Distribution Relationship

  • 摘要: 纳西族东巴画具有纹理复杂、色彩鲜明的艺术特点,是民族文化的重要组成部分.针对东巴画数量稀少且主题多样、细节丰富,现有的深度学习方法直接应用于东巴画分类任务时易导致模型分类准确率低、泛化性差等问题,提出一种融合空间信息和分布关系的东巴画小样本分类方法.首先采用一对独立的编码操作将空间信息聚合到通道注意力,获取具有方向感知和位置敏感的注意力特征图,准确定位东巴画目标区域;然后将特征图输入自适应元权重生成器,加强支持集与查询集样本之间的信息关联,增强特征的表达能力;最后构建双向图神经网络,显式地利用样本实例信息和分布关系,通过优化重构样本图特征进行分类预测.在3个自建的东巴画数据集上进行模型训练的实验结果表明,与其他小样本算法相比,所提方法的分类准确率分别提升9.79,5.76和8.21个百分点,可以更有效地提取东巴画分类所需的图像全局信息和细节特征,并能扩展至其他少数民族文化的保护与传承等领域.

     

    Abstract: As an important part of national culture, Naxi Dongba painting has the artistic characteristics of rich texture and bright colors. However, due to the scarcity and complexity of Dongba painting, the existing deep learning methods perform poorly on the classification task. To improve this situation, we propose a few-shot classification method that combines spatial information and distribution relationship. Firstly, a pair of independent encoding operations are used to aggregate the spatial information into the channel attention, and the attention feature maps with orientation-aware and position-sensitive are obtained to accurately locate key regions. Secondly, feed the feature maps into the adaptive meta-weight generator to enhance the expressive ability of the features. Finally, a dual complete graph neural network is constructed to explicitly use the sample instance information and distribution relationship to perform classification prediction. Model training on 3 self-built Dongba painting datasets. Compared with other few-shot methods, the classification accuracy of the proposed method is improved by 9.79, 5.76 and 8.21 percentage points, respectively. The experimental results show that the method can more effectively exact the global information and detailed features of samples required for the classification of Dongba painting, and it can be effectively applied to the protection and inheritance of minority cultures.

     

/

返回文章
返回