Shadow Removal by a Lightness-Guided Network With Training on Unpaired Data.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Shadow removal is crucial for enhancing image quality and has diverse applications.
• Deep learning, particularly Convolutional Neural Networks (CNNs), is highly effective for shadow removal.
• Training CNNs on unpaired data is preferred due to easier data collection.

Purpose of the Study:

• To present a novel Lightness-Guided Shadow Removal Network (LG-ShadowNet) for shadow removal.
• To enable effective shadow removal using unpaired training data.

Main Methods:

• A two-stage CNN approach is employed: first, a module compensates for lightness.
• A second CNN module refines shadow removal, guided by the lightness information from the first module.
• A novel loss function is introduced to leverage color priors.

Main Results:

• The proposed LG-ShadowNet demonstrates superior performance compared to state-of-the-art methods.
• Experiments were conducted on widely used datasets: ISTD, adjusted ISTD, and USR.
• The method excels in shadow removal when trained on unpaired data.

Conclusions:

• LG-ShadowNet offers an effective solution for shadow removal using unpaired data.
• The lightness-guided approach and color prior utilization contribute to improved performance.
• This method advances the field of shadow removal in computer vision.