LCSCNet: Linear Compressing Based Skip-Connecting Network for Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs.
• Existing SR networks like ResNet and DenseNet utilize skip connections but have limitations in feature handling and parameter efficiency.
• There is a need for efficient network architectures that can effectively leverage hierarchical features for improved SR performance.

Purpose of the Study:

• To propose a novel and efficient network architecture, LCSCNet, for image super-resolution.
• To enhance feature representation by distinguishing between former and newly-explored feature maps.
• To improve the exploitation of hierarchical information through an adaptive fusion strategy.

Main Methods:

• Developed a Linear Compressing based Skip-Connecting Network (LCSCNet) architecture.
• Incorporated a linear compressing layer within skip connections to differentiate feature maps.
• Implemented an adaptive element-wise fusion strategy inspired by LSTM gate units, coupled with multi-supervised training.

Main Results:

• LCSCNet demonstrates superior performance compared to state-of-the-art image super-resolution algorithms.
• The proposed architecture achieves a balance between DenseNet's feature treatment and ResNet's parameter efficiency.
• Experimental validation confirms the effectiveness of LCSCNet in enhancing image super-resolution.

Conclusions:

• LCSCNet offers a concise and efficient solution for image super-resolution tasks.
• The novel architecture effectively handles feature maps and exploits hierarchical information.
• The proposed methods contribute to advancing the field of deep learning-based image super-resolution.