Automatic Search Dense Connection Module for Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Increasing demand for high-resolution displays challenges current camera imaging systems.
• Deep convolutional neural networks (CNNs) show promise for image super-resolution (SR), but suffer from high computational costs.
• Existing SR methods often have restrictive image resolution due to physical limitations of imaging systems.

Purpose of the Study:

• To develop a lightweight network for image super-resolution (SR) that minimizes memory and computation.
• To improve feature utilization and reduce redundancy in dense network connections for SR tasks.
• To leverage neural architecture search (NAS) for optimizing dense connections in SR models.

Main Methods:

• Proposed a novel lightweight network architecture for image super-resolution (SR).
• Employed neural architecture search (NAS) to automatically discover optimal dense connections.
• Developed an automated search for dense connections (ASDCN) to reduce network redundancy and focus on salient features.

Main Results:

• The derived ASDCN model demonstrated superior performance compared to state-of-the-art SR models.
• Experiments on five public datasets confirmed the effectiveness of the proposed lightweight network.
• The network successfully reduced redundancy in dense connections, enhancing focus on valuable features for SR.

Conclusions:

• The proposed lightweight network with ASDCN offers an effective solution for practical image super-resolution.
• This approach addresses the limitations of existing CNN-based SR methods regarding computational overhead.
• The findings suggest a promising direction for developing efficient and high-performance image super-resolution systems.