Deep Learning for Speckle Noise Quantification: A Novel CNN-Based Approach

Speckle noise is a major challenge in digital image processing, particularly in computer vision and medical imaging, where it degrades visual quality and hampers critical tasks such as object recognition and edge detection. In modalities like ultrasound imaging, the presence of speckle significantly reduces image clarity and diagnostic accuracy, underscoring the need for effective noise level quantification. This paper presents a Convolutional Neural Network (CNN)-based approach to quantify speckle noise levels in random images. A custom dataset was created by applying varying speckle noise levels (0%–100%) to 1,000 base images, generating 11 noise classes. A custom CNN model was developed and trained to classify these noise levels, and its performance was compared with pre-trained models including VGG16, InceptionV3, ResNet50, and Xception. Experimental results show that the proposed CNN model outperforms the pre-trained models in terms of accuracy and training efficiency, achieving 89.75% accuracy, with minimal misclassification and stable convergence. This work provides a robust solution for noise level quantification, which is crucial for improving preprocessing in image denoising applications. Future work will focus on improving generalizability and integrating the model into real-time ultrasound systems for enhanced medical imaging.