MSViT:融合多尺度特征的轻量级图像分类模型

MSViT: A Lightweight Image Classification Model Integrating Multi-Scale Features

关键词:  图像分类  多尺度特征融合  多阶损失函数  特征金字塔网络  Transformer

Key words:  Image classification  Multi-scale feature fusion  multi-order loss function  Feature pyramid Network  Transformer

摘要:

针对现有Vison Transformer模型在局部特征捕捉和多尺度特征融合方面的局限性，本文提出了一种新型的融合多尺度特征的轻量化图像分类混合模型MSViT（Multi-scale Vision Transformer）。首先，在编码器中设计了捕获通道特征的多尺度前馈神经网络（MSFFN）模块，能够更好地提取空间和多尺度通道特征。其次，设计一个新的级联特征融合解码器，通过整合特征金字塔网络（FPN）和多阶段特征融合解码器，显著提升了模型对不同尺度特征的交互和融合能力。最后，模型引入多阶损失函数，以全面优化不同尺度特征在图像分类任务中的表现。为了验证MSViT的有效性，本文在四个实验数据集ImageNet的子集、Cifar100、APTOS2019和Mushroom66上进行大量的实验。其中在ImageNet子集上的实验结果显示，MSViT实现了87.58%的Top-1准确率，较Edgevits_xxs提升了2.27%。这些结果证明了MSViT在图像分类任务中的有效性。

Abstract:

Aiming at the limitations of existing Vison Transformer models in local feature capture and Multi-scale feature fusion, this paper proposes a new lightweight image classification hybrid model MSViT (Multi-scale Vision Transformer), which integrates multi-scale features. Firstly, a multi-scale feedforward neural network (MSFFN) module is designed to capture channel features in the encoder, which can better extract spatial and multi-scale channel features. Secondly, a new cascaded feature fusion decoder is designed. By integrating the feature pyramid network (FPN) and the multi-stage feature fusion decoder, the interaction and fusion ability of the model for different scale features are significantly improved. Finally, a multi-order loss function is introduced to optimize the performance of different scale features in image classification tasks. In order to verify the validity of MSViT, a large number of experiments were conducted on a subset of four experimental datasets ImageNet, Cifar100, APTOS2019 and Mushroom66. The experimental results on the ImageNet subset show that MSViT achieves 87.58% Top-1 accuracy, which is 2.27% higher than Edgevits_xxs. These results demonstrate the effectiveness of MSViT in image classification tasks.