User assisted separation of reflections from a single image using a sparsity prior

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Images captured through transparent materials like glass often contain superimposed reflections and the scene behind.
• Decomposing these single-input images into distinct layers is an ill-posed problem due to infinite possible solutions without prior information.

Purpose of the Study:

• To develop a user-assisted method for separating reflected and transmitted image layers.
• To investigate the effectiveness of natural image statistics as a prior for improving image decomposition.

Main Methods:

• Implemented a user-assisted approach where users label image gradients.
• Utilized a sparsity prior derived from natural image statistics.
• Optimized the sparsity prior using the iterative reweighted least squares (IRLS) algorithm.

Main Results:

• The proposed method significantly outperforms traditional Gaussian priors for image separation.
• Effective separation was achieved with a modest number of user-provided gradient labels.
• The sparsity prior derived from natural image statistics proved crucial for superior performance.

Conclusions:

• User-assisted gradient labeling combined with a natural image sparsity prior offers a robust solution for image decomposition through glass.
• This approach enhances the quality of image separation compared to methods relying on less informative priors.
• The findings demonstrate the power of incorporating statistical properties of natural images into computer vision tasks.