Performance comparison of machine learning driven approaches for classification of complex noises in quick response code images

Abstract

Quick response codes (QRCs) are found on many consumer products and often encode security information. However, information retrieval at receiving end may become challenging due to the degraded clarity of QRC images. This degradation may occur because of the transmission of digital images over noise channels or limited printing technology. Although the ability to reduce noises is critical, it is just as important to define the type and quantity of noises present in QRC images. Therefore, this study proposed a simple deep learning-based architecture to segregate the image as either an original (normal) QRC or a noisy QRC and identifies the noise type present in the image. For this, the study is divided into two stages. Firstly, it generated a QRC image dataset of 80,000 images by introducing seven different noises (speckle, salt & pepper, Poisson, pepper, localvar, salt, and Gaussian) to the original QRC images. Secondly, the generated dataset is fed to train the proposed convolutional neural network (CNN)-based model, seventeen pre-trained deep learning models, and two classical machine learning algorithms (Naïve Bayes (NB) and Decision Tree (DT)). XceptionNet attained the highest accuracy (87.48%) and kappa (85.7%). However, it is worth noting that the proposed CNN network with few layers competes with the state-of-the-art models and attained near to best accuracy (86.75%). Furthermore, detailed analysis shows that all models failed to classify images having Gaussian and Localvar noises correctly.