摘要
为提高卷积神经网络图像分类精度的同时实现网络轻量化,本文提出一种基于坐标重要性池化和解耦类别对齐蒸馏的图像分类算法.首先,设计一种坐标重要性池化模块并将其嵌入ResNet34,充分利用图像像素的位置信息,以增强其判别重要性特征的能力;其次,采用BlurPool缓解在下采样过程中移位等变性丢失对网络性能的影响,以此构建教师网络;最后,构造一种解耦类别对齐蒸馏算法,分别考虑目标类和非目标类的知识并引入类别之间的关联信息,以高效地将分类知识从教师网络迁移到轻量级MobileNetV3学生网络.在不同数据集上的实验结果表明,本文提出的教师网络有效提高了分类性能,且蒸馏训练后的学生网络明显优于其他同量级网络,实现了更优越的综合性能,能够更好地应用于计算和内存资源受限的实际场景.
An image classification algorithm based on coordinate importance pooling and decoupled class alignment distillation is proposed to improve the image classification accuracy of convolutional neural networks while achieving network lightweighting.Firstly,a coordinate importance pooling module is designed and embedded it into ResNet34,in order to fully utilize the positional information of image pixels to enhance the ability to discriminate important features.Secondly,BlurPool is used to mitigate the impact on network performance due to shift equivariance during down-sampling,and to construct the teacher network.Finally,the decoupled class alignment distillation algorithm was constructed to efficiently migrate image classification knowledge from the teacher network to the lightweight MobileNetV3 network,which considers the knowledge of target and non-target class separately and introduces correlation information between the class.The experimental results on different datasets showed that the proposed teacher network effectively improves the classification performance,and the distillation-trained student network achieves superior overall performance than other networks of the same magnitude,making it better applicable to practical scenarios with limited computational and storage power.
作者
刘颖
薛家昊
张伟东
许志杰
LIU Ying;XUE Jia-hao;ZHANG Wei-dong;XU Zhi-jie(Center for Image and Information Processing,Xi’an University of Posts and Telecommunications,Xi’an,Shaanxi 710121,China;International Joint-Research Center for Wireless Communication and Information Processing,Xi’an,Shaanxi 710121,China;University of Huddersfield,West Yorkshire HD13DH,United Kingdom of Great Britain and Northern Ireland)
出处
《电子学报》
北大核心
2025年第3期962-973,共12页
Acta Electronica Sinica
基金
国家自然科学基金(No.62106195)。
