Single Image Super-Resolution Based on Global Dense Feature Fusion Convolutional Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep neural networks (DNNs) are successful in single image super-resolution (SISR).
• Hierarchical features in deeper networks are often underutilized in existing SISR methods.
• Limited influence of previous hierarchical information impacts SISR performance.

Purpose of the Study:

• To propose a novel deep neural network, the global dense feature fusion convolutional network (DFFNet), for SISR.
• To fully leverage global intermediate hierarchical features in SISR.
• To improve the performance of SISR by enhancing feature utilization.

Main Methods:

• Introduced a feature fusion block (FFblock) as the core module of DFFNet.
• FFblock enables direct reading of raw global features from preceding layers.
• FFblock learns spatial and channel correlations holistically, creating a continuous global information memory.

Main Results:

• DFFNet effectively utilizes global intermediate features.
• The feature fusion block facilitates a continuous global information memory mechanism.
• Experimental results demonstrate favorable performance of DFFNet compared to state-of-the-art methods.

Conclusions:

• The proposed DFFNet effectively addresses the underutilization of hierarchical features in SISR.
• The novel FFblock design enhances feature correlation learning and information flow.
• DFFNet represents a significant advancement in single image super-resolution technology.