A Benchmark for Edge-Preserving Image Smoothing.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Edge-preserving image smoothing is crucial for low-level vision tasks.
• Existing methods struggle with diverse image content and lack objective evaluation metrics.
• Subjectivity in performance evaluation hinders algorithm comparison.

Purpose of the Study:

• To introduce a comprehensive benchmark for edge-preserving image smoothing.
• To provide an objective evaluation framework for smoothing algorithms.
• To advance the state-of-the-art in edge-preserving image smoothing.

Main Methods:

• Developed a benchmark including an image dataset with ground truth smoothing results.
• Created baseline algorithms using deep convolutional neural networks (CNNs).
• Designed novel loss functions tailored for edge-preserving smoothing within CNNs.

Main Results:

• The benchmark dataset comprises 500 training and testing images across various object categories.
• Baseline deep networks achieve competitive smoothing results, outperforming existing methods.
• Trained deep networks offer faster processing speeds compared to most state-of-the-art algorithms.
• Quantitative and qualitative evaluations demonstrate leading smoothing performance.

Conclusions:

• The proposed benchmark addresses limitations in existing edge-preserving smoothing methods.
• The benchmark facilitates objective comparison and advancement of smoothing algorithms.
• The developed deep learning models provide efficient and effective edge-preserving smoothing solutions.