Robust deep convolutional neural network against image distortions

Feng-Tsun Chien,Liang-Yao Wang,Sau-Gee Chen

doi:10.1017/atsip.2021.14

Feng-Tsun Chien, Liang-Yao Wang + Show 1 more

Open Access

https://doi.org/10.1017/atsip.2021.14

Copy DOI

Abstract

Many approaches have been proposed in the literature to enhance the robustness of Convolutional Neural Network (CNN)-based architectures against image distortions. Attempts to combat various types of distortions can be made by combining multiple expert networks, each trained by a certain type of distorted images, which however lead to a large model with high complexity. In this paper, we propose a CNN-based architecture with a pre-processing unit in which only undistorted data are used for training. The pre-processing unit employs discrete cosine transform (DCT) and discrete wavelets transform (DWT) to remove high-frequency components while capturing prominent high-frequency features in the undistorted data by means of random selection. We further utilize the singular value decomposition (SVD) to extract features before feeding the preprocessed data into the CNN for training. During testing, distorted images directly enter the CNN for classification without having to go through the hybrid module. Five different types of distortions are produced in the SVHN dataset and the CIFAR-10/100 datasets. Experimental results show that the proposed DCT-DWT-SVD module built upon the CNN architecture provides a classifier robust to input image distortions, outperforming the state-of-the-art approaches in terms of accuracy under different types of distortions.

Highlights

Deep Convolutional Neural Network (DCNN) has been widely used in image classification due to its impressive capability of capturing relevant features of different classes in the data [1,2,3]
It is known that the classification accuracy of generic DCNN can be affected by image distortions in the input data: adding small amount of distortion to the test set usually results in a significant reduction in the classification accuracy of the network [4]
In this paper, we leverage the advantages of discrete cosine transform (DCT), discrete wavelets transform (DWT), and singular value decomposition (SVD) and propose a new hybrid preprocessing module in the “training” stage to further elevate the robustness and accuracy of the classifier against various types of image distortions

Summary

INTRODUCTION

Deep Convolutional Neural Network (DCNN) has been widely used in image classification due to its impressive capability of capturing relevant features of different classes in the data [1,2,3]. The existing approaches perform well for some types of distortions, but may not promisingly work as well for others To address this problem, in this paper, we leverage the advantages of DCT, discrete wavelets transform (DWT), and singular value decomposition (SVD) and propose a new hybrid preprocessing module in the “training” stage to further elevate the robustness and accuracy of the classifier against various types of image distortions. The experimental results show that the trained DCNN with the proposed DCT-DWT-SVD preprocessing module is robust to all five types of distortions with improved accuracy. We propose a hybrid DCT-DWT-SVD preprocessing module in the DCNN architecture that does not require fine-tuning, re-training, or data augmentation. Experimental results demonstrate the robustness of the proposed DCT-DWT-SVD module to various types of image distortions.

RELATED WORK

HYBRID MODULE WITH DCT-DWT-SVD

EXPERIMENTS AND DATASETS

CONCLUSION