A Lightweight Feature Distillation and Enhancement Network for Super-Resolution Remote Sensing Images

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning-based super-resolution (SR) models offer significant advancements.
• Large parameter counts in existing SR models limit practical applications on resource-constrained devices.

Purpose of the Study:

• To introduce a lightweight feature distillation and enhancement network (FDENet) for efficient super-resolution.
• To address the limitations of high parameter counts in current deep learning-based SR methods.

Main Methods:

• Proposed a novel feature distillation and enhancement block (FDEB) with two parts: feature distillation and feature enhancement.
• Feature distillation employs stepwise distillation and a stepwise fusion mechanism (SFM) with a shallow pixel attention block (SRAB).
• Feature enhancement utilizes bilateral bands to refine features and extract complex background information, particularly for remote sensing images.

Main Results:

• FDENet demonstrates a significant reduction in model parameters compared to existing advanced SR models.
• Experimental results show superior performance of FDENet in super-resolution tasks.
• The proposed network effectively enhances feature expression capabilities through fused features.

Conclusions:

• FDENet offers an efficient and effective solution for super-resolution imaging.
• The lightweight design makes FDENet suitable for real-world applications with limited hardware capabilities.
• The feature distillation and enhancement approach provides a promising direction for future SR research.