计算机工程

• •    

基于轻量级特征融合卷积网络的图像分类算法

  

  • 发布日期:2020-12-09

Image classification algorithm based on lightweight feature fusion convolutional network

  • Published:2020-12-09

摘要: 针对传统卷积神经网络的卷积核过于单一、不具备多样性以及网络结构复杂、参数冗余的问题,设计了一种轻量级 特征融合卷积神经网络 MS-FNet。该网络架构的 Fusion 模块采用多路结构增加了卷积神经网络的宽度,对输入特征图通过不 同尺寸的卷积核处理,提高了网络在同一层中提取不同特征的能力,并且在每一次卷积后采用 BN、ReLU 等方法去除冗余特 征。在网络的最后使用卷积层代替传统的全连接层,不仅加快了模型的训练,还缓解了由于参数过多造成过拟合的问题。实 验结果表明,MS-FNet 在达到较低错误率的同时,大大减少了网络的参数量,有更强的学习能力。

Abstract: Aiming at the problem of traditional convolutional neural networks where the convolutional kernels are too single, not diverse, and the network structure is complex with redundant parameters. A lightweight feature fusion convolutional neural network MS-FNet is designed for this purpose. The Fusion module of the network architecture used a multi-branch structure to increase the width of the convolutional neural network, and the input feature map was processed by convolutional kernels of different sizes, which improved the ability of network to extract different features in the same layer. And the redundant features are removed after each convolution using BN, ReLU, etc. At the end of the network, convolutional layers were used to replace the traditional fully connected layer, which not only accelerated the training of model but also alleviated overfitting problems due to too many parameters. The experimental results show that MS-FNet, while achieving a lower error rate, it greatly reduces the amount of network parameters and has better learning capabilities.