基于多分辨率自蒸馏网络的小样本图像分类  

Few-Shot Image Classification Based on Multi-Resolution Self-Distillation Network

作　　者：仇真 奚雪峰 崔志明 盛胜利 胡伏原[1,2,3] QIU Zhen;XI Xuefeng;CUI Zhiming;SHENG Shengli;HU Fuyuan(School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou,Jiangsu 215000,China;Suzhou Key Laboratory of Virtual Reality Intelligent Interaction and Application Technology,Suzhou,Jiangsu 215000,China;Suzhou Smart City Research Institute,Suzhou,Jiangsu 215000,China;Computer Science Department,Texas Tech University,Lubbock 79401,USA)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215000 [2]苏州市虚拟现实智能交互及应用重点实验室,江苏苏州215000 [3]苏州智慧城市研究院,江苏苏州215000 [4]德州理工大学计算机科学系,美国拉伯克79401

出　　处：《计算机工程》2022年第12期232-240,共9页Computer Engineering

基　　金：国家自然科学基金(61876217,61876121,62176175);江苏省“六大人才高峰”高层次人才项目(XYDXX-086);苏州市科技计划项目(SGC2021078)。

摘　　要：因图像数据具有大量的空间冗余信息,传统的多分辨率网络在处理图像数据时会产生较高的计算成本。而自蒸馏学习方法能够在精度与计算成本之间实现动态平衡,使模型在不增加网络深度和宽度的基础上,有效地提高模型精度。提出一种多分辨率自蒸馏网络(MRSDN),用于解决小样本学习中输入样本的空间冗余问题。从原始网络中分出一个浅层子网络以识别图像的低分辨率表示,并且保持该原始网络识别高分辨率图像特征的能力。同时,在多分辨率网络中引入改进的全局注意力机制,以减少信息损失且放大全局交互表示。利用自蒸馏学习方法将网络中更深层的知识压缩到浅层子网络中,以提升浅层子网络的泛化能力。在此基础上,将低分辨率网络中的粗粒度特征融合到高分辨率网络中,从而提高模型提取图像特征的能力。实验结果表明,在Mini-ImageNet数据集上MRSDN网络对5-way 1-shot与5-way 5-shot任务的准确率分别为56.34%和74.35%,在Tiered-ImageNet数据集上对5-way 1-shot与5-way 5-shot任务的准确率分别为59.56%和78.96%,能有效缓解高分辨率图像输入时的空间冗余问题,提高小样本图像分类的准确率。The traditional multi-resolution network incurs high computing costs when processing image data owing to a large amount of spatial redundancy information in the image data.The Self-Distillation(SD)learning method can achieve a dynamic balance between accuracy and calculation cost,effectively improving the accuracy of the model without increasing the depth and width of the network.A Multi-Resolution Self-Distillation Network(MRSDN)is proposed to solve the spatial redundancy of input samples in Few-Shot Learning(FSL).A shallow sub-network is separated from the original network to recognize the low-resolution representation of the image,and the ability of the original network to recognize the high-resolution image features is maintained.In addition,an improved Global Attention Mechanism(GAM)is introduced into the multi-resolution network to reduce information loss and enlarge the global interactive representation.The SD learning method is used to compress the in-depth knowledge of the network into a shallow sub-network to improve the generalization ability of the shallow sub-network.The coarse granularity features in the low-resolution network are fused into the high-resolution network to improve the ability of the model to extract image features.The experimental results show that accuracy of the MRSDN network for five-way one-shot and five-way five-shot tasks on the Mini-ImageNet dataset are 56.34%and 74.35%,respectively.The accuracy of the network for fiveway one-shot and five-way five-shot tasks on the Tiered-ImageNet dataset are 59.56%and 78.96%,respectively.The proposed network can effectively alleviate the spatial redundancy when inputting high-resolution images,improving the accuracy of few-shot image classification.

关 键 词：自蒸馏学习 小样本学习 多分辨率网络 空间冗余 全局注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]