Tree-Guided CNN Boosts Image Super-Resolution in Joint University Study

Overview

The paper proposes a tree‑guided convolutional neural network (TSRNet) for single‑image super‑resolution. By embedding a binary‑tree structure into a deep CNN, the model explicitly strengthens the correlation among key layers, enabling more effective extraction of structural information from low‑resolution inputs.

Network Architecture

The architecture consists of three tightly coupled components:

Binary‑tree‑guided backbone – a hierarchical arrangement of convolutional blocks where each node aggregates features from its two child branches, thereby amplifying salient representations at critical depths.

Cosine‑similarity cross‑domain extractor – a module that computes cosine similarity between feature vectors from different branches to capture directional similarity and local salient cues, which are then fused back into the tree.

Adaptive Nesterov momentum optimizer (Adan) – an optimizer that adapts Nesterov momentum based on gradient statistics, suppressing local gradient explosion and providing smoother parameter updates.

The overall network diagram is shown below:

Key Modules

Tree‑guided feature aggregation

Each internal node receives feature maps from its left and right child nodes, concatenates them, and applies a 1×1 convolution to re‑weight the combined representation. This operation increases the mutual information between layers that are otherwise distant in a plain sequential CNN, allowing the network to discover and exploit critical structural cues.

Cosine‑similarity cross‑domain extractor

For a pair of feature vectors f_i and f_j, the module computes cosine_similarity = (f_i·f_j) / (||f_i||·||f_j||). High similarity scores indicate consistent directional features across branches, which are then emphasized through a gating mechanism. This enhances robustness of local salient details such as edges and textures.

Adaptive Nesterov momentum optimizer (Adan)

The optimizer follows the Adan formulation (Zhang et al., 2023), adjusting the momentum term based on the moving average of past gradients and their variance. By dynamically scaling the momentum, the method mitigates sudden gradient spikes that can destabilize training, especially in deep tree‑structured networks.

Training Details

All experiments use the standard DIV2K training set with L1 loss. The learning rate is initialized at 1e‑4 and decayed by a factor of 0.5 every 200 k iterations. The Adan optimizer is employed with default hyper‑parameters (β1=0.9, β2=0.999, ε=1e‑8). Data augmentation includes random horizontal flips and 90° rotations.

Experimental Results

Quantitative evaluation on benchmark datasets (Set5, Set14, BSD100, Urban100) shows that TSRNet consistently outperforms representative SR models such as EDSR, RCAN, and SwinIR. For example, on Set5 (×4 up‑sampling) TSRNet achieves a PSNR gain of +0.12 dB over RCAN while maintaining comparable model size.

Qualitative comparisons illustrate sharper edge reconstruction and more faithful texture recovery, especially in regions with fine patterns. The following figures present PSNR/SSIM tables and visual examples:

Conclusion

The tree‑guided CNN effectively discovers and leverages critical structural information across hierarchical branches, leading to superior super‑resolution performance. Cosine‑similarity based cross‑domain extraction preserves local salient details, while the adaptive Nesterov (Adan) optimizer stabilizes training and prevents gradient explosion. Both quantitative metrics and visual inspection confirm the advantage of the proposed method.

Paper: https://ieeexplore.ieee.org/abstract/document/11010139

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