Pixelwise Estimation of Signal-Dependent Image Noise Using Deep Residual Learning

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional image denoising relies on a single noise level parameter.
• Existing methods often assume uniform noise distribution across images.
• Real-world noise is frequently signal-dependent and non-uniform.

Purpose of the Study:

• To develop a method for precise, pixelwise noise-level estimation.
• To address the limitations of global scalar noise estimation.
• To introduce the first deep learning approach for pixelwise noise estimation.

Main Methods:

• Proposed a deep convolutional neural network named deep residual noise estimator (DRNE).
• Designed DRNE with customized residual blocks, avoiding pooling or interpolation.
• Formulated noise estimation as a pixel-to-pixel prediction task.

Main Results:

• DRNE achieved superior performance in non-homogeneous noise estimation compared to state-of-the-art methods.
• Demonstrated improved denoising performance with signal-dependent Gaussian noise.
• Validated the effectiveness of pixelwise noise estimation for advanced denoising.

Conclusions:

• The proposed DRNE effectively estimates pixelwise noise levels.
• This approach offers significant advantages over traditional global noise estimation.
• DRNE enhances the capabilities of deep learning-based image denoising techniques.