| 摘 要: 近年来,在卷积神经网络框架体系下,小样本细粒度图像分类研究取得颇为显著的进展,但其基础特征提取能力仍受到传统卷积算子表达能力的严重制约:一方面,卷积核尺寸预先设定且保持固定,难以针对不同尺度的目标实施自适应调整；另一方面,采样位置局限于规则的矩形感受野,难以处理轮廓不规则和姿态多变的细粒度目标,进而导致关键局部判别信息的缺失。针对这一现状,提出多方向自适应卷积(Multi-directional Adaptive Convolution,MDAConv)模块,替换骨干网络的底层标准卷积结构。通过动态学习卷积核的高度、宽度与方向参数,利用局部纹理复杂度确定采样点数量及其空间布局,MDAConv模块实现尺度、方向与采样密度等维度的联合自适应调节,极大地增强整体模型对于细微纹理差异和形状变化等底层特征的敏感性与鲁棒性。基于公开小样本细粒度数据集的实验结果表明,在相同的计算开销条件下,与多个经典模型相比,基于MDAConv模块的神经网络模型具有更高的分类准确率和平均精度,从而为小样本细粒度图像分类提供一种具有良好扩展潜力的卷积结构。 |
| 关键词: 卷积神经网络 小样本细粒度图像分类 自适应卷积 多方向自适应卷积 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然基金青年项目（62201335） |
|
| Few-shot Fine-grained Image Classification Based on Multi-directional Adaptive Convolution |
|
hedan, zhangweichuan, yangjunpo
|
陕西科技大学
|
| Abstract: In recent years,within the framework of convolutional neural network (CNN),research on few-shot fine-grained image classification has achieved remarkable progress.However,its basic feature extraction capability is still severely limited by the representational capacity of traditional convolutional operators:on the one hand,the size of convolutional kernels is pre-defined and fixed,making it difficult to perform adaptive adjustments for targets of different scales; on the other hand,sampling positions are confined to regular rectangular receptive fields,which hinders the handling of fine-grained targets with irregular contours and variable poses,thereby leading to the lack of key local discriminative information.To address this issue,we propose a Multi-directional Adaptive Convolution (MDAConv) module to replace the bottom standard convolutional structures of backbone networks.By dynamically learning the height,width,and direction parameters of the convolutional kernel,and determining the number of sampling points and their spatial layout based on local texture complexity,the MDAConv module achieves joint adaptive adjustment across dimensions such as scale,direction,and sampling density.This significantly enhances the sensitivity and robustness of the overall model to low-level features,including subtle texture differences and shape variations.Experimental results on public few-shot fine-grained datasets demonstrate that,under the same computational cost,the neural network model based on the MDAConv module achieves higher classification accuracy and mean average precision (mAP) compared with several classical models.Thus,this work provides a convolutional structure with excellent scalability for few-shot fine-grained image classification. |
| Keywords: convolutional neural network few-shot fine-grained image classification adaptive convolution multi-directional adaptive convolution |
