Selected work

Representative Papers

Beyond Convolution: A Taxonomy of Structured Operators for Learning-Based Image Processing

Mar 12, 2026

Standard convolutions, due to their fixed structure, linearity, and reliance on local averaging, struggle to capture complex image characteristics such as low-rank structures, adaptive basis representations, and non-uniform spatial dependencies. This work proposes a unified taxonomy encompassing five classes of structured operators—decomposition-based, adaptive weighting, basis-adaptive, integral/kernel-based, and attention-based—and systematically analyzes their differences along key dimensions including locality, linearity, and equivariance. By leveraging techniques such as singular value/tensor decomposition, content-adaptive weighting, learnable analysis bases, position-dependent nonlinear kernels, and attention mechanisms, the study comprehensively evaluates the performance of these operators across image-to-image and image-to-label tasks. The findings clarify the respective strengths and limitations of each operator class, offering both theoretical insights and practical guidance for future research.

0 citationsRead paper

Learning to Predict Aboveground Biomass from RGB Images with 3D Synthetic Scenes

Nov 28, 2025

Traditional methods struggle to accurately and cost-effectively estimate aboveground biomass (AGB) in dense forests. To address this, we propose the first end-to-end deep learning framework for dense AGB prediction from a single ground-level RGB image. Our key innovation is formulating pixel-wise AGB density maps as the regression target. We synthesize a large-scale 3D forest dataset—SPREAD—by leveraging instance segmentation masks and tree attributes (e.g., species, diameter at breast height) to generate pixel-level AGB ground truth via allometric equations. Critically, our method requires neither LiDAR nor multi-view imagery, greatly enhancing scalability and enabling broad public participation. Experiments demonstrate strong generalization: median absolute error is 1.22 kg/m² on the held-out SPREAD test set and 1.94 kg/m² on real-world forest images, confirming both accuracy and practical applicability.

0 citationsRead paper

Automated Monitoring of Cultural Heritage Artifacts Using Semantic Segmentation

Nov 25, 2025

Automated detection and segmentation of cracks in cultural heritage artifacts (e.g., statues, monuments) remain challenging due to poor generalizability across diverse, unseen artifact types and conditions. To address this, we propose a semantic segmentation framework based on the U-Net architecture, systematically evaluating multiple CNN encoders—including ResNet and EfficientNet—for fine-grained, pixel-level crack delineation. Evaluated on the OmniCrack30k dataset, our models achieve state-of-the-art performance in mIoU, Dice, and Jaccard scores. Crucially, we demonstrate that pre-trained encoders exhibit strong cross-domain generalization to previously unseen heritage objects without artifact-specific fine-tuning. Furthermore, out-of-distribution qualitative validation on unlabeled real-world damaged images confirms robustness under varying lighting conditions, material textures, and scale variations. This work establishes a plug-and-play, high-accuracy, and highly adaptable paradigm for automated crack diagnosis in cultural heritage conservation.

0 citationsRead paper

Optimal Weighted Convolution for Classification and Denosing

May 30, 2025

Standard convolutional neural networks neglect the relative spatial positions between neighboring and central pixels, limiting their capacity for spatial modeling. To address this, we propose a novel weighted convolution operator that embeds a learnable spatial density function directly into the convolution kernel—enabling adaptive, position-aware weight assignment based on relative coordinates—without increasing model parameters or compromising architectural compatibility. The operator is dimension-agnostic, supporting arbitrary-dimensional regular grid signals, and is designed for plug-and-play integration. Implemented in PyTorch with precomputation for efficient inference, it achieves significant improvements: on CIFAR-100, VGG accuracy increases by 10.05 percentage points to 66.94%; on the DIV2K image denoising benchmark, PSNR improves by 2.46 dB to 22.63—substantially outperforming standard convolution baselines.

0 citationsRead paper

Optimal Density Functions for Weighted Convolution in Learning Models

May 30, 2025

Traditional uniform convolutions neglect spatial distances among neighboring pixels, limiting modeling accuracy. To address this, we propose weighted convolution—a geometrically aware adaptive local aggregation mechanism that dynamically modulates the contribution of each neighbor to the central pixel via a learnable optimal density function. Our key innovation is the first integration of density function optimization directly into the convolution operation, decoupling kernel parameter learning from spatial weight optimization: convolutional kernels are updated via stochastic gradient descent, while the density function is optimized globally using the DIRECT-L algorithm. The method applies to regular-grid signals (2D/3D images) and demonstrates significant improvements in image denoising—achieving substantial PSNR gains, an average 53% reduction in loss, and markedly improved test accuracy, with only an 11% increase in computational overhead. It thus offers both high efficiency and strong robustness.

0 citationsRead paper

Recent publications

Latest Papers

Beyond Convolution: A Taxonomy of Structured Operators for Learning-Based Image Processing

Mar 12, 2026

Standard convolutions, due to their fixed structure, linearity, and reliance on local averaging, struggle to capture complex image characteristics such as low-rank structures, adaptive basis representations, and non-uniform spatial dependencies. This work proposes a unified taxonomy encompassing five classes of structured operators—decomposition-based, adaptive weighting, basis-adaptive, integral/kernel-based, and attention-based—and systematically analyzes their differences along key dimensions including locality, linearity, and equivariance. By leveraging techniques such as singular value/tensor decomposition, content-adaptive weighting, learnable analysis bases, position-dependent nonlinear kernels, and attention mechanisms, the study comprehensively evaluates the performance of these operators across image-to-image and image-to-label tasks. The findings clarify the respective strengths and limitations of each operator class, offering both theoretical insights and practical guidance for future research.

0 citationsRead paper

Learning to Predict Aboveground Biomass from RGB Images with 3D Synthetic Scenes

Nov 28, 2025

Traditional methods struggle to accurately and cost-effectively estimate aboveground biomass (AGB) in dense forests. To address this, we propose the first end-to-end deep learning framework for dense AGB prediction from a single ground-level RGB image. Our key innovation is formulating pixel-wise AGB density maps as the regression target. We synthesize a large-scale 3D forest dataset—SPREAD—by leveraging instance segmentation masks and tree attributes (e.g., species, diameter at breast height) to generate pixel-level AGB ground truth via allometric equations. Critically, our method requires neither LiDAR nor multi-view imagery, greatly enhancing scalability and enabling broad public participation. Experiments demonstrate strong generalization: median absolute error is 1.22 kg/m² on the held-out SPREAD test set and 1.94 kg/m² on real-world forest images, confirming both accuracy and practical applicability.

0 citationsRead paper

Automated Monitoring of Cultural Heritage Artifacts Using Semantic Segmentation

Nov 25, 2025

Automated detection and segmentation of cracks in cultural heritage artifacts (e.g., statues, monuments) remain challenging due to poor generalizability across diverse, unseen artifact types and conditions. To address this, we propose a semantic segmentation framework based on the U-Net architecture, systematically evaluating multiple CNN encoders—including ResNet and EfficientNet—for fine-grained, pixel-level crack delineation. Evaluated on the OmniCrack30k dataset, our models achieve state-of-the-art performance in mIoU, Dice, and Jaccard scores. Crucially, we demonstrate that pre-trained encoders exhibit strong cross-domain generalization to previously unseen heritage objects without artifact-specific fine-tuning. Furthermore, out-of-distribution qualitative validation on unlabeled real-world damaged images confirms robustness under varying lighting conditions, material textures, and scale variations. This work establishes a plug-and-play, high-accuracy, and highly adaptable paradigm for automated crack diagnosis in cultural heritage conservation.

0 citationsRead paper

Optimal Weighted Convolution for Classification and Denosing

May 30, 2025

Standard convolutional neural networks neglect the relative spatial positions between neighboring and central pixels, limiting their capacity for spatial modeling. To address this, we propose a novel weighted convolution operator that embeds a learnable spatial density function directly into the convolution kernel—enabling adaptive, position-aware weight assignment based on relative coordinates—without increasing model parameters or compromising architectural compatibility. The operator is dimension-agnostic, supporting arbitrary-dimensional regular grid signals, and is designed for plug-and-play integration. Implemented in PyTorch with precomputation for efficient inference, it achieves significant improvements: on CIFAR-100, VGG accuracy increases by 10.05 percentage points to 66.94%; on the DIV2K image denoising benchmark, PSNR improves by 2.46 dB to 22.63—substantially outperforming standard convolution baselines.

0 citationsRead paper

Optimal Density Functions for Weighted Convolution in Learning Models

May 30, 2025

Traditional uniform convolutions neglect spatial distances among neighboring pixels, limiting modeling accuracy. To address this, we propose weighted convolution—a geometrically aware adaptive local aggregation mechanism that dynamically modulates the contribution of each neighbor to the central pixel via a learnable optimal density function. Our key innovation is the first integration of density function optimization directly into the convolution operation, decoupling kernel parameter learning from spatial weight optimization: convolutional kernels are updated via stochastic gradient descent, while the density function is optimized globally using the DIRECT-L algorithm. The method applies to regular-grid signals (2D/3D images) and demonstrates significant improvements in image denoising—achieving substantial PSNR gains, an average 53% reduction in loss, and markedly improved test accuracy, with only an 11% increase in computational overhead. It thus offers both high efficiency and strong robustness.

0 citationsRead paper