| 尹冠千,欧阳春娟,高曦.基于图卷积神经网络的小尺寸图像分类算法[J].井冈山大学自然版,2025,46(3):84-90 |
| 基于图卷积神经网络的小尺寸图像分类算法 |
| A SMALL-SIZE IMAGE CLASSIFICATION ALGORITHM BASED ON GRAPH CONVOLUTIONAL NEURAL NETWORKS |
| 投稿时间：2025-02-28 修订日期：2025-03-28 |
| DOI：10.3969/j.issn.1674-8085.2025.03.011 |
| 中文关键词: 图像识别 图卷积神经网络 深度学习 训练方法 网络结构 |
| 英文关键词: image recognition graph convolutional networks deep learning training method network structure |
| 基金项目:国家自然科学基金项目(42061055)；江西省自然科学基金项目(20192BAB207021，20202BABL202047）；江西省教育厅科技计划项目(GJJ201008) |
|
| 摘要点击次数: 25 |
| 全文下载次数: 36 |
| 中文摘要: |
| 深度学习技术在图像识别领域取得了显著进展，从经典卷积神经网络到轻量化模型和迁移学习，可以大幅提升了图像分类与目标检测的精度与效率，其中小尺寸图像分类任务由于图像信息量有限，还是存在很多挑战。本研究提出了一类基于图卷积神经网络的小尺寸图像识别算法。该算法在模型架构设计中，通过引入残差连接将第三层的图卷积层重构为残差拼接的图卷积层，有效缓解了深层网络梯度消失问题，并增强了特征传递能力；同时采用注意力池化替代传统平均池化，通过动态分配节点权重强化关键特征的表达能力。基于Fashion MNIST数据集的实验表明，改进后的模型较基准模型分类准确率提升了7%，融合残差连接与注意力机制的图卷积神经网络架构，在超小尺寸图像分类和视角变形任务中展现出更强的鲁棒性，为相关领域的研究提供了新的技术路径。 |
| 英文摘要: |
| Recent years have significant progress in image recognition driven by deep learning, from classical convolutional neural networks (CNNs) to lightweight models and transfer learning, substantially improving classification and detection accuracy and efficiency. However, small-size image classification remains challenging due to the limited visual information. This study proposes a novel recognition algorithm for small images based on graph convolutional networks (GCNs). The architecture introduces a residual concatenation GCN layer by reconstructing the third graph convolutional layer with residual connections, effectively mitigating gradient vanishing and enhancing feature propagation. Additionally, attention-based pooling replaces average pooling to dynamically weight nodes and better represent key features. Experiments on the Fashion MNIST dataset show a 7% gain over the baseline. The results demonstrate that combining residual connections and attention mechanisms in GCNs boosts robustness in ultra-low-resolution image classification and viewpoint deformation, offering a new direction for related research. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
