Image super-resolution with an enhanced group convolutional neural network.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) are effective for image super-resolution but often require deep architectures, increasing computational costs.
• Existing methods struggle with extracting comprehensive low-frequency information and long-range contextual dependencies.

Purpose of the Study:

• To develop a shallow yet effective CNN architecture for single image super-resolution (SISR).
• To improve the extraction of low-frequency information and long-distance contextual information.
• To create a model adaptable to low-resolution images of varying sizes.

Main Methods:

• Proposed an Enhanced Super-Resolution Group CNN (ESRGCNN) with a shallow architecture.
• Implemented full fusion of deep and wide channel features to capture inter-channel correlations.
• Introduced a signal enhancement operation for long-distance contextual information inheritance.
• Integrated an adaptive up-sampling operation for handling diverse image sizes.

Main Results:

• ESRGCNN demonstrates superior performance compared to state-of-the-art methods in SISR.
• Achieved improvements in complexity, execution speed, and overall image quality.
• Visual effects in super-resolved images were significantly enhanced.

Conclusions:

• The proposed ESRGCNN offers an efficient and effective solution for single image super-resolution.
• Shallow architectures can achieve high performance by optimizing feature fusion and contextual information inheritance.
• The model provides a practical advancement in image super-resolution technology.