Multi-University Team Proposes Tree-Guided CNN for Image Super-Resolution

Abstract

Deep convolutional neural networks can capture richer structural information through deeper architectures, improving image super‑resolution. However, identifying the most influential layers within a single deep network is difficult. This work introduces a tree‑guided CNN that uses a binary‑tree topology to highlight critical layers, thereby strengthening the correlation of key hierarchical information and enhancing the model’s adaptability for high‑quality image reconstruction.

Contributions

The binary‑tree‑based network architecture amplifies the relevance of crucial layers, leading to better super‑resolution results.

A cosine‑based technique extracts cross‑domain features and local salient information, making the structural cues more robust for high‑quality image recovery.

An adaptive Nesterov momentum optimizer mitigates local gradient explosion and refines parameters, further boosting super‑resolution performance.

Method Overview

The proposed method integrates three key components:

Tree‑Guided Structure: A binary‑tree layout directs the flow of information, allowing the network to dynamically prioritize important layers during training.

Cosine Feature Extraction: Cosine similarity is employed to capture cross‑domain relationships, extracting directional features that enhance local detail preservation.

Adaptive Nesterov Momentum Optimizer: The optimizer adapts its momentum based on gradient statistics, preventing gradient explosion and accelerating convergence.

The complete network diagram is shown below:

Performance Validation

Extensive experiments compare the proposed tree‑guided CNN against baseline super‑resolution models. Quantitative metrics (e.g., PSNR, SSIM) and qualitative visual results demonstrate the superiority of the method. The following figures illustrate the experimental outcomes:

Conclusion

The tree‑guided CNN effectively leverages hierarchical relationships to locate and exploit critical structural information, resulting in notable improvements in image super‑resolution quality. Cosine‑based cross‑domain extraction preserves local salient features, while the adaptive Nesterov optimizer prevents training stagnation and gradient issues. Both quantitative and qualitative analyses confirm the method’s advantage. Future work will explore richer inter‑tree associations to further advance super‑resolution capabilities.

References

Paper: A Tree-Guided CNN for Image Super-Resolution (IEEE Xplore: https://ieeexplore.ieee.org/abstract/document/11010139)

Code repository: https://github.com/hellloxiaotian/TSRNet

Authors: Tian Chunwei, Song Mingjian, Fan Xiaopeng, Zheng Xiangtao, Zhang Bob, Zhang David