Video Super-Resolution Method Using Deformable Convolution-Based Alignment Network.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Advancements in sensors have driven progress in image and video processing for visual sensing.
• Video Super-Resolution (VSR) reconstructs high-resolution video sequences from low-resolution inputs.
• Effective VSR leverages spatial and temporal information from consecutive low-resolution frames.

Purpose of the Study:

• To propose a novel convolutional neural network-based VSR method.
• To introduce the Deformable Convolution-based Alignment Network (DCAN) for generating quadrupled high-resolution video sequences.
• To improve VSR performance and reduce computational complexity.

Main Methods:

• The proposed DCAN utilizes a feature extraction block, two deformable convolution-based alignment blocks, and an up-sampling block.
• Deformable convolutions are employed to capture spatial and temporal characteristics more effectively.
• The network is designed to reconstruct high-resolution sequences from low-resolution inputs.

Main Results:

• The DCAN achieved superior performance in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) compared to existing methods.
• Experimental results demonstrate significant reductions in network complexity, including fewer parameters, lower memory usage, and faster inference speed.
• The method successfully generates high-resolution sequences with quadruple the size of the original low-resolution sequences.

Conclusions:

• The Deformable Convolution-based Alignment Network (DCAN) presents an effective approach for video super-resolution.
• DCAN offers a favorable trade-off between performance enhancement and computational efficiency.
• This method contributes to the advancement of VSR techniques, particularly for real-time applications.