Image Reconstruction Based on Progressive Multistage Distillation Convolution Neural Network

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Current algorithms face feature loss during distillation and information loss in compressed channel attention.
• This necessitates improved methods for preserving key information in image reconstruction networks.

Purpose of the Study:

• To propose a progressive multistage distillation network (PMDN) to enhance feature distillation and channel attention.
• To maximize network performance by minimizing information loss during image reconstruction.

Main Methods:

• Developed a progressive multistage distillation network for gradual feature refinement.
• Introduced a weight-sharing information lossless attention block to enhance channel characteristics without compression.
• Utilized convolution layers to model interchannel dependencies effectively.

Main Results:

• The proposed PMDN achieves a balance between network performance, parameter count, and computational complexity.
• Demonstrated highly competitive performance in both objective metrics and subjective visual quality on benchmark datasets.
• Validated the effectiveness of gradual feature distillation from coarse to fine for improved network performance.

Conclusions:

• The progressive multistage distillation network offers significant advantages for image reconstruction tasks.
• The proposed attention mechanism effectively preserves channel information, outperforming existing methods.
• The approach provides a robust solution for high-quality image reconstruction with optimized resource utilization.