Image Classification Model Based on Deep Learning in Internet of Things

Songshang Zou,Hao Chen,Wenshu Chen,Hongju Cheng

doi:10.1155/2020/6677907

Abstract

In the environment of Internet of Things, the convolutional neural network (CNN) is an important tool and method of image classification. However, the features that are extracted by each layer of CNN are all high dimensional, and the features differ among the layers. In addition, these features contain substantial amounts of redundant information. To prevent the increase in the computational burden and the decline of the model generalization performance that are caused by high dimensionality, this paper proposes an improved image classification algorithm based on deep feature fusion, which designs and builds an 8-layer CNN. In addition, it reduces the dimensionality of the features via the principal component analysis (PCA) dimensionality reduction algorithm and fuses the features that have undergone dimensionality reduction to make the obtained features more typical and differential. The experimental results demonstrate that the proposed algorithm improves the performance of the model and achieves satisfactory accuracy.

Highlights

In the era of Internet of Things, image classification plays an important role in multimedia information processing
When using deep learning algorithms to represent complex data distributions, nonlinear network models with deep layers can be used to learn the deep features from the data in the case of few samples
To evaluate the classification performance of the convolutional neural network (CNN) model that is designed in this paper, which is based on deep feature fusion, experiments have been conducted on two image datasets, namely, Food-101 and Places2, and the results are compared with those of other image classification methods

Summary

Introduction

In the era of Internet of Things, image classification plays an important role in multimedia information processing. (i) This paper analyzes the structure of CNN, studies the principles of activation functions, specifies the role that nonlinear activation functions play in neural networks, and shows that via facilitation by nonlinear functions; CNN has stronger feature representation performance and can realize complex image classification (ii) To address the problems that conventional image classification algorithms that are based on deep learning cannot effectively fuse multilayered deep features and perform poorly in terms of classification accuracy, this paper proposes an improve image classification algorithm that is based on deep feature fusion and improves the diversity and expressiveness of the extracted features to improve the classification performance (iii) By comparing the classification performances of the CNN model on the Food-101 and Places datasets under various activation functions, it is demonstrated that the activation function that is used in this paper can improve the classification accuracy of the model on image datasets and ensure its convergence (iv) This paper conducts a performance analysis and evaluation of the proposed algorithm in comparison with other algorithms.

Related Work

CNN Image Recognition and Classification Based on Feature Fusion

Analysis of the Experimental Result

Findings

Conclusions