SF-ConvNeXt在低分辨率图像分类的研究  

Research on Low-Resolution Image Classification Using SF-ConvNeXt

作　　者：陈阵 郑长勇 CHEN Zhen;ZHENG Chang-yong(School of Electronics and Information Engineering,Anhui Jianzhu University,Hefei,230601,Anhui)

机构地区：[1]安徽建筑大学电子与信息工程学院,安徽合肥230601

出　　处：《蚌埠学院学报》2025年第2期52-60,共9页Journal of Bengbu University

基　　金：安徽省自然科学基金资助项目(1808085MF206)。

摘　　要：针对资源受限设备难以部署识别精度高的低分辨率图像分类模型的问题,提出轻量分融卷积网络模型(Split Fusion Convolutional Network, SF-ConvNeXt)。以ConvNeXt_V2_atto模型为基线,首先,为提高模型在低分辨率图像分类的精度,对模型框架和激活函数进行优化,设计SF-ConvNeXt网络架构。其次,针对ConvNeXt_V2残差块中深度卷积核尺寸过大带来的特征提取不丰富问题,在深度卷积后,采用通道分裂技术减少多路特征提取的通道数,通过融合3×3分组卷积增强特征提取能力,并结合通道混洗技术增强通道信息交流,进而提出分融卷积块(Split Fusion Convolutional Block, SF-ConvBlock)。实验结果表明,仅有344.20 k参数量的SF-ConvNeXt模型在公开低分辨率数据集CIFAR-10和Fashion-MNIST上分别取得了93.57%和94.88%的准确率,表现出良好的性能。Addressing the challenge of deploying high-accuracy low-resolution image classification models on resource-constrained devices,it proposed a lightweight Split Fusion Convolutional Network(SF-ConvNeXt)in the paper.Based on the ConvNeXt_V2_atto model as a baseline,the SF-ConvNeXt architecture was designed by optimizing the model framework and activation functions to enhance classification accuracy on low-resolution images.Furthermore,to mitigate the issue of insufficient feature extraction caused by the large kernel size in the ConvNeXt_V2 residual blocks,a channel splitting technique was introduced after depthwise convolution to reduce the number of channels for multi-path feature extraction.It was combined with 3×3 group convolution to enhance feature extraction capability and channel shuffle technology to improve inter-channel information exchange,resulting in the proposed Split Fusion Convolutional Block(SF-ConvBlock).Experimental results demonstrated that the SF-ConvNeXt model,with only 344.20 k parameters,achieves accuracies of 93.57%and 94.88%on the CIFAR-10 and Fashion-MNIST low-resolution datasets,respectively,showcasing its superior performance.

关 键 词：低分辨率 图像分类 轻量化 卷积神经网络 残差网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]