Learning Shadow Removal From Unpaired Samples via Reciprocal Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Shadows in images degrade visual quality and complicate image analysis.
• Existing shadow removal methods often require pixel-level annotations, which are labor-intensive to obtain.
• Weakly supervised learning offers a promising alternative by leveraging less detailed labels.

Purpose of the Study:

• To develop a weakly supervised shadow removal model that does not rely on pixelwise-paired training data.
• To propose a deep reciprocal learning framework for joint optimization of shadow removal and detection.
• To enhance shadow removal accuracy using image-level shadow presence labels.

Main Methods:

• A deep reciprocal learning model interactively optimizes a shadow remover and a shadow detector.
• Shadow removal is formulated as an optimization problem with a latent shadow mask.
• A shadow detector is trained using priors from the shadow remover, employing a self-paced learning strategy.
• Color-maintenance loss and a shadow-attention discriminator are incorporated for improved optimization.

Main Results:

• The proposed model achieves superior performance in shadow removal.
• Demonstrated effectiveness on diverse datasets, including ISTD, SRD, and unpaired USR datasets.
• The weakly supervised approach successfully handles the absence of pixelwise-paired training samples.

Conclusions:

• The deep reciprocal learning model offers an effective solution for weakly supervised shadow removal.
• Interactive optimization of shadow removal and detection significantly enhances model capabilities.
• The method provides a practical approach for shadow removal in scenarios with limited annotated data.