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

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Quick Response Codes (QRCs) are prevalent on consumer products for security information.
• Image degradation in QRCs poses challenges for information retrieval.
• Identifying noise types in QRC images is crucial for accurate data processing.

Purpose of the Study:

• To propose a deep learning architecture for classifying QRC images as normal or noisy.
• To identify the specific types of noise present in degraded QRC images.
• To evaluate the performance of the proposed model against state-of-the-art deep learning and classical machine learning algorithms.

Main Methods:

• Generated a dataset of 80,000 QRC images with seven distinct noise types (speckle, salt & pepper, Poisson, pepper, localvar, salt, Gaussian).
• Trained a proposed convolutional neural network (CNN)-based model, seventeen pre-trained deep learning models, and two classical algorithms (Naïve Bayes, Decision Tree).
• Evaluated model performance based on accuracy and kappa statistics.

Main Results:

• The proposed CNN model achieved an accuracy of 86.75%, closely rivaling the best-performing XceptionNet (87.48%).
• All tested models demonstrated difficulty in correctly classifying images with Gaussian and Localvar noise.
• The study highlights the effectiveness of a simple CNN for QRC noise classification.

Conclusions:

• A deep learning approach can effectively segregate normal from noisy QRC images and identify noise types.
• The proposed CNN model offers a competitive alternative to more complex deep learning architectures for this task.
• Further research is needed to improve classification accuracy for Gaussian and Localvar noise in QRC images.