Learning Spatially Variant Linear Representation Models for Joint Filtering.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Joint filtering methods commonly use guidance images as priors to transfer structural information.
• Existing techniques often rely on local linear models or hand-designed functions, limiting their adaptability.
• Learning spatially variant linear representation models (SVLRMs) for vision tasks is inherently ill-posed.

Purpose of the Study:

• To propose a new joint filtering method utilizing a spatially variant linear representation model (SVLRM).
• To develop an effective deep convolutional neural network (CNN) approach for estimating SVLRM coefficients.
• To demonstrate the versatility and effectiveness of the proposed method across various image processing applications.

Main Methods:

• A novel joint filtering method based on a spatially variant linear representation model (SVLRM) is proposed.
• A deep convolutional neural network (CNN) is developed to estimate the spatially variant linear representation coefficients.
• The CNN is constrained by the SVLRM to model structural information from both guidance and input images.

Main Results:

• The proposed deep CNN-based SVLRM method effectively models structural information for joint filtering.
• The approach demonstrates successful application in depth/RGB image upsampling and restoration, flash deblurring, natural image denoising, and scale-aware filtering.
• The linear representation model is shown to be extendable to higher-order representations (e.g., quadratic, cubic).

Conclusions:

• The proposed deep CNN-constrained SVLRM method offers a powerful and flexible framework for joint image filtering.
• Experimental results confirm the method's superior performance compared to state-of-the-art task-specific approaches.
• This work advances joint filtering techniques by leveraging deep learning for robust structural information modeling.