ProDebNet: projector deblurring using a convolutional neural network.

Area of Science:

• Computer Vision
• Image Processing
• Display Technology

Background:

• Non-planar and multi-projection display surfaces introduce projection blur due to defocus and subsurface scattering.
• Existing methods struggle with real-time processing and require specific screen characteristics.

Purpose of the Study:

• To propose ProDebNet, an end-to-end deep learning network for real-time projection deblurring.
• To develop a method that synthesizes projection images minimizing blur without explicit screen geometry or scattering estimation.

Main Methods:

• ProDebNet is an end-to-end convolutional neural network architecture.
• A novel "pseudo-projected" synthetic dataset was created for training, generalizing to real-world blur.
• The network synthesizes a deblurred projection image directly.

Main Results:

• ProDebNet effectively compensates for defocus and subsurface scattering, dominant types of projection blur.
• The method achieves real-time processing speeds, significantly faster than baseline methods.
• Experimental results show successful deblurring even in real-world projection scenarios.

Conclusions:

• ProDebNet offers an efficient and effective solution for real-time projection deblurring.
• The synthetic dataset approach enables robust generalization without real captured data.
• The proposed network advances the quality of projected images on complex surfaces.