Joint Image Filtering with Deep Convolutional Networks.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Joint image filters utilize a guidance image as a prior to transfer structural details for noise suppression or resolution enhancement.
• Current methods often involve complex filter constructions or manual objective functions, hindering understanding and optimization.

Purpose of the Study:

• To propose a learning-based framework for constructing joint image filters using Convolutional Neural Networks (CNNs).
• To develop a method that selectively transfers salient structures consistent with both guidance and target images, unlike existing approaches.

Main Methods:

• A novel joint filter construction is proposed, leveraging Convolutional Neural Networks (CNNs).
• The approach enables selective transfer of structural information, considering consistency between guidance and target images.

Main Results:

• The proposed CNN-based joint filter demonstrates effective selective structure transfer.
• The model exhibits strong generalization capabilities, performing well across diverse image modalities (e.g., RGB/depth, flash/non-flash, RGB/NIR).

Conclusions:

• The learning-based approach offers a coherent framework for understanding, improving, and accelerating joint filters.
• The proposed joint filter achieves state-of-the-art performance, validated through extensive experiments.