Dr. Md. Jarez Miah

Thesis title: An Efficient CNN-Based Regression Model for Noise Prediction in Color Images

Abstract—Noise estimation is a critical task in image processing, especially when dealing with various noise types like Gaussian and Poisson noise. Accurately estimating the noise level before denoising is essential for optimizing noise removal, balancing noise reduction with detail preservation, and improving the efficiency of denoising processes. This study proposes two novel deep learning models, NoiseNet and its enhanced version NoiseNetV2, for predicting noise levels in images under mixed noise conditions, including both Gaussian and Poisson noise.

The NoiseNet model leverages deep learning techniques, particularly convolutional neural networks (CNNs), to predict the noise level in color images. Trained using datasets like Flickr30K, CelebA, and COCO, NoiseNet accurately estimates varying noise levels and outperforms several traditional and modern deep learning models, including VGG16, ResNet50, MobileNetV2, DenseNet121, and AutoencoderKL. Additionally, it surpasses non-deep learning methods such as the Scikit-learn approach and BM3D technique. The model achieves superior performance in predicting Gaussian noise and Poisson noise individually, as measured by metrics such as mean absolute error, root mean squared error, mean absolute percentage error and R² score.

Building upon the success of NoiseNet, NoiseNetV2 introduces advanced architectural improvements that enhance its performance in mixed noise conditions, making it more robust in estimating both Gaussian and Poisson noise levels simultaneously. Through extensive validation and testing, NoiseNetV2 demonstrates superior generalization capabilities across different data, providing highly accurate noise estimations for real-world applications in areas such as medical imaging, autonomous driving, and surveillance.

This research contributes to the advancement of noise estimation techniques, offering an effective solution for image restoration by accurately predicting noise levels, ensuring better preservation of image quality while removing noise. The improved models provide a foundation for future developments in denoising and image enhancement techniques, making them suitable for a wide range of applications where high-quality, noise-free images are essential.