Stereoscopic video deblurring transformer.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Stereoscopic cameras are increasingly prevalent in mobile devices and intelligent systems.
• Stereo video quality is susceptible to blur distortion from motion, with limited research on stereoscopic deblurring compared to monocular methods.

Purpose of the Study:

• To develop an advanced Transformer-based framework for effective stereo video deblurring.
• To enhance stereo video quality by mitigating motion blur and improving visual fidelity.

Main Methods:

• A novel Transformer architecture incorporating a self-attention layer and feed-forward layer to capture inter-frame correlations.
• Integration of 3D convolutional neural networks (3D CNNs) for processing successive frames and correcting motion blur.
• An improved parallax attention module (PAM) to leverage stereo-viewpoint and cross-view information for enhanced deblurring.

Main Results:

• The proposed framework demonstrates efficient and effective deblurring of stereo videos.
• Experimental validation on public datasets confirms the method's superior performance.
• The approach significantly outperforms existing image and video deblurring techniques.

Conclusions:

• The developed Transformer-based stereo video deblurring framework achieves state-of-the-art results.
• The novel attention mechanisms and 3D CNN integration effectively address stereo video blur.
• This work advances the field of stereoscopic video enhancement.