A Fast Multi-Scale Generative Adversarial Network for Image Compressed Sensing.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Deep neural networks excel in image compressed sensing reconstruction.
• Existing methods face challenges with sampling patterns and high computational demands.

Purpose of the Study:

• To develop a computationally efficient deep learning model for image compressed sensing.
• To improve image reconstruction quality by addressing limitations of current methods.

Main Methods:

• Proposed a fast multi-scale generative adversarial network (FMSGAN).
• Introduced a multi-scale sampling structure with varying kernel sizes for effective image decomposition and multi-scale feature capture.
• Developed a lightweight multi-scale residual structure with smaller convolution kernels and channel attention for efficient deep image reconstruction.
• Utilized a combined optimization function including perceptual, MSE, and adversarial losses.

Main Results:

• FMSGAN demonstrates state-of-the-art image reconstruction quality.
• Achieved significantly lower computational complexity compared to existing methods.
• Effectively captures spatial features at multiple scales.

Conclusions:

• FMSGAN offers a superior balance between reconstruction quality and computational efficiency in image compressed sensing.
• The proposed multi-scale sampling and lightweight residual structures are key to its performance.
• This method advances the field of efficient and high-quality image reconstruction.