💻 computer science

CASPA: Content-Aware Global Aggregation and Spatial prior Channel Attention for Thangka Image Super-Resolution

This paper proposes CASPA, a novel super-resolution network that combines Content-Aware Global Aggregation and Spatial prior Channel Attention modules to overcome the limitations of existing Vision Transformers, thereby achieving high-fidelity reconstruction of Thangka images with enhanced long-range semantic capture and fine geometric detail preservation.

Original authors: Mengyuan Zhang, Nianyi Wang, Yutong Wang, Yakun Xin, Chenyi Xia, Yanwen Gao

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mengyuan Zhang, Nianyi Wang, Yutong Wang, Yakun Xin, Chenyi Xia, Yanwen Gao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a priceless, ancient painting called a Thangka. These are intricate Tibetan religious scrolls filled with tiny, delicate lines, repeating patterns, and vibrant colors. Over time, the physical paintings have faded, and the digital photos we have of them are often blurry, low-resolution, and missing those fine details.

The goal of this paper is to build a "smart digital restorer" that can take a blurry, low-quality photo of a Thangka and magically reconstruct it into a sharp, high-definition masterpiece. The authors call their new tool CASPA.

Here is how they explain the problem and their solution, using simple analogies:

The Problem: Why Old AI Tools Fail

Current AI tools used for image restoration (called Vision Transformers) try to fix blurry images by looking at small, square "windows" of the picture at a time. The authors say this approach has two major flaws when dealing with Thangkas:

The "Fence" Problem (Local Windows):
Imagine you are trying to fix a torn map of a city, but you are only allowed to look at a tiny 3x3 inch square of the map at a time. If you see a missing street in your square, you can't fix it because you can't see the rest of the city to know where the street should go.
- In the paper: Thangkas have repeating patterns (like lotus flowers or clouds) that appear far apart from each other. Old AI tools get stuck in their "local windows" and can't see that a blurry flower here is identical to a sharp flower 10 inches away. They miss the big picture, leading to blurry, confused details.
The "Blind Colorist" Problem (Lost Direction):
Imagine a painter who knows what colors to use but has forgotten where to put them. They mix the colors well but lose the shape of the lines.
- In the paper: Traditional tools look at the whole image to decide which colors are important, but in doing so, they forget the "geometry" or the direction of the lines. Thangkas have thousands of thin, continuous lines (like wire strokes). When the AI forgets the direction, these lines get broken, disconnected, or turned into a muddy blur.

The Solution: Introducing CASPA

The authors built a new system with two special "superpowers" to fix these issues.

1. The "Content Detective" (Content-Aware Global Aggregation - CGA)

Instead of looking at the image in rigid, square boxes, this module acts like a smart librarian.

How it works: It scans the entire image and groups together pixels that "look" similar, no matter how far apart they are. If there is a red petal in the top left corner and a blurry red petal in the bottom right, the librarian says, "Ah! These belong together!"
The Result: It breaks the "fence" rule. It gathers all the similar patterns from across the whole painting to help reconstruct the blurry parts. It uses the clear parts of the image to fix the blurry parts, even if they are miles apart in the photo.

2. The "Directional Compass" (Spatial Prior Channel Attention - SCA)

This module acts like a compass and a ruler for the lines.

How it works: Instead of just looking at colors, it uses special "strip" tools (like long, thin brushes) that scan the image horizontally and vertically. It pays close attention to the direction of the lines.
The Result: It remembers that a line is supposed to go straight or curve in a specific way. This ensures that the thin, wire-like strokes in the Thangka stay continuous and sharp, rather than getting broken or turning into noise.

The Results: Does it Work?

The authors tested their new "CASPA" tool on a custom dataset of Thangka images they collected.

The Score: It scored higher on standard quality tests (PSNR) than any other existing tool, including the current "champions" of the field.
The Look: When they compared the results visually, the old tools produced images with "grid artifacts" (blocky errors) and broken lines. CASPA produced images where the lines were smooth, continuous, and the repeating patterns were perfectly aligned.
Speed: Despite doing this complex work, the tool is actually very lightweight and fast, using less computer memory than many of its competitors.

The Limitations (Where it Stumbles)

The authors are honest about where their tool isn't perfect yet:

Unique Details: If a part of the painting is totally unique (like a specific eye that doesn't appear anywhere else in the image), the "Content Detective" can't find a reference to help fix it. In these cases, the AI might smooth it out too much, making it look a bit soft.
Messy Tangles: If lines are crossing over each other at weird, irregular angles (like a tangled knot of hair), the "Directional Compass" (which prefers straight horizontal and vertical lines) can get confused, causing the lines to blur together.

Summary

In short, the paper presents CASPA, a new AI tool designed specifically to restore ancient Thangka paintings. It solves the problem of "local blindness" by grouping similar patterns from across the whole image, and it solves the problem of "lost direction" by using special tools to keep thin lines sharp and straight. The result is a clearer, more faithful digital reconstruction of these cultural treasures.

Technical Summary: CASPA for Thangka Image Super-Resolution

Problem Statement

The digital preservation of Thangka, a form of Tibetan Buddhist painting, faces critical challenges due to the irreversible degradation of physical artifacts and the low resolution of existing digital archives. While Image Super-Resolution (ISR) offers a solution, current state-of-the-art Vision Transformer (ViT) models struggle to reconstruct Thangka images with high fidelity. The paper identifies two specific mechanistic limitations in existing Transformer-based approaches when applied to Thangka:

Spatial Context Fragmentation: Thangka images feature complex, symmetrical structures and mineral pigments that exhibit strong non-local self-similarity (e.g., repetitive patterns like lotus petals spanning large distances). Traditional ViT models, such as SwinIR, rely on local window mechanisms. These rigid physical grids restrict feature interactions, preventing the model from leveraging long-range similarities to reconstruct dense, repetitive textures, leading to blurred details and structural dislocations.
Loss of Spatial Geometric Priors: Thangka art is characterized by abundant, slender, and coherent high-frequency geometric lines (e.g., wire-like strokes and intricate drapery). Existing channel attention mechanisms (e.g., in DAT) often rely on global pooling operations to establish channel dependencies. This process discards crucial spatial topological information, causing the model to focus solely on global color statistics while failing to maintain the structural coherence of directional lines, resulting in edge blurring and trajectory disconnection.

Methodology: CASPA Architecture

To address these challenges, the authors propose CASPA (Content-Aware Global Aggregation and Spatial prior Channel Attention), a novel super-resolution network built upon the Dual Aggregation Transformer (DAT) architecture. The network consists of three stages: shallow feature extraction, deep feature extraction (comprising 12 cascaded residual groups), and high-resolution reconstruction. The core innovation lies in two specialized modules that alternate within the residual groups:

1. Content-Aware Global Aggregation (CGA) Module

The CGA module replaces traditional spatial window self-attention to break physical distance constraints.

Mechanism: Instead of partitioning the image into fixed grids, CGA performs global token clustering based on feature similarity across the entire image.
Process: It utilizes a set of learnable global cluster centers to calculate cosine similarity between input tokens and these centers. Tokens are adaptively assigned to semantic groups regardless of their spatial location.
Interaction: Self-attention is executed within these semantic groups. This allows the model to bridge physical distances, aggregating scattered but semantically similar pixels (e.g., identical pigments) to restore long-range repetitive patterns with linear complexity $O(N)$ .
Efficiency: To ensure parallel efficiency, large groups are partitioned into sub-groups, and overlapping interactions are allowed at semantic boundaries.

2. Spatial Prior Channel Attention (SCA) Module

The SCA module is designed to inject directional and spatial priors into the channel branch, addressing the loss of topological information.

Mechanism: It employs decomposed cross-shaped strip convolutions (e.g., cascaded $1 \times 21$ and $21 \times 1$ kernels) to create an equivalent large receptive field (e.g., $21 \times 21$ ).
Process:
1. Channel Attention: Generates a channel attention vector using parallel global average and max pooling followed by an MLP, acting as a "soft-gating" classifier to highlight channels with critical geometric structures.
2. Spatial Priors: The channel-refined features are processed by parallel depth-wise convolution branches with different kernel sizes to capture anisotropic spatial geometric priors.
3. Fusion: These spatial features are summed and mixed via a $1 \times 1$ convolution to generate a dynamic spatial attention map, which is then multiplied with the channel-refined features.
Outcome: This design preserves the structural integrity of continuous high-frequency lines without introducing excessive computational overhead.

Auxiliary Components

Adaptive Cross-Gating (ACG): A mechanism using parallel depth-wise convolutions to achieve bidirectional soft-mask recalibration, ensuring seamless fusion of macroscopic long-range semantics and microscopic local textures.
Dynamic Feed-Forward Network (DFFN): Replaces standard activation functions with a parameterized Dynamic Hyperbolic Tangent (DynamicTanh) to enhance non-linear fitting capabilities for complex details while mitigating gradient saturation.

Key Contributions

The paper outlines three primary contributions:

CGA Module: A global interaction mechanism that abandons physical grid partitioning in favor of feature similarity-driven global token clustering. This successfully activates non-local self-similarity priors, enabling efficient collaborative reconstruction of long-range repetitive patterns and scattered pigments.
SCA Module: A channel attention mechanism that utilizes decomposed cross-shaped strip convolutions to inject robust multi-scale spatial and directional priors. This overcomes the topological information loss caused by traditional global pooling, ensuring the high-fidelity reconstruction of complex continuous lines.
Comprehensive Validation: The proposal of a custom Thangka dataset and the demonstration of the model's effectiveness in restoring both long-range semantics and fine structural details, outperforming existing SOTA models on both domain-specific and general benchmarks.

Experimental Results

Dataset and Setup

Custom Dataset: 5,000 HR Thangka images (512 $\times$ 512) for training, 680 for validation, and 1,000 for testing.
Benchmarks: Urban100 and Manga109 for cross-domain generalization.
Baselines: EDSR, SwinIR, HAT, SRFormer, and DAT.
Metrics: PSNR, SSIM, LPIPS, and FID.

Quantitative Performance

Thangka Dataset ( $\times$ 4 Scale): CASPA achieved a PSNR of 32.9045 dB and SSIM of 0.8817, outperforming the baseline DAT (32.7774 dB) by 0.1271 dB. It also achieved the best scores in LPIPS and FID, indicating superior perceptual quality and realism.
General Benchmarks: On Urban100 and Manga109 ( $\times$ 4), CASPA established new state-of-the-art performance, demonstrating strong cross-domain generalization capabilities.
Efficiency: Despite high performance, CASPA is highly lightweight with only 5.02 M parameters and 21.97 G FLOPs (at 64 $\times$ 64 input), significantly lower than competitors like HAT (20.77 M params) and EDSR (43.09 M params). It also demonstrated faster inference times (19.38 ms) compared to most baselines.

Qualitative Analysis

Visual comparisons revealed that traditional window-based methods often produced grid artifacts and structural dislocations in dense, parallel lines. In contrast, CASPA successfully reconstructed clear, coherent geometric lines and high-frequency textures, closely approximating the ground truth.

Ablation Study

Removing either module resulted in performance degradation:

Without SCA: Loss of spatial priors led to distorted trajectories and blurred edges in fine lines.
Without CGA: Spatial context fragmentation caused incorrect detail fusion and blurring in repetitive patterns.
Full Model: The combination of both modules yielded the highest PSNR, confirming their complementary roles.

Significance and Limitations

The paper claims that CASPA represents a significant advancement in the digital preservation of cultural heritage, specifically for Thangka art. By resolving the trade-off between global context awareness and local geometric fidelity, the model offers a robust solution for restoring complex, non-local structures and directional lines that have historically been difficult for Transformer-based models to handle.

The authors also note the model's generalization potential, successfully applying it to Dunhuang frescoes and European medieval church frescoes with strong results, suggesting broad applicability in ancient mural restoration.

However, the paper modestly acknowledges limitations:

Unique Micro-structures: The model exhibits over-smoothing on unique, isolated details (e.g., specific eye features) that lack global references for the CGA module to leverage.
Complex Intersections: Dense, non-orthogonally intertwined lines (e.g., eyebrows) can suffer from blurring due to the fixed horizontal and vertical sampling directions of the SCA module.

Future work is proposed to explore more flexible sampling methods and external generative priors to address these specific edge cases.