Jul 2026· International Conference on Pattern Recognition· Vol abs/2607.06424· 0 citations· 30 references
Computer Science
TL;DR
XRFormer is introduced, a transformer architecture tailored to XRF spectra through a multiscale convolutional tokenizer that injects locality and multi-resolution inductive biases before global self-attention before transformer-based XRF modeling under data-limited conditions.
Abstract
X-ray fluorescence (XRF) spectroscopy is a key modality for material analysis in cultural heritage. However, automated learning from XRF spectra remains challenging: XRF spectra are complex one-dimensional signals composed of sharp elemental peaks, broader structures, and background variations that are not taken into account by existing learning-based models. This paper introduces XRFormer, a transformer architecture tailored to XRF spectra through a multiscale convolutional tokenizer that injects locality and multi-resolution inductive biases before global self-attention. The tokenizer progressively reduces spectral resolution while increasing embedding dimensionality, and the resulting token sequence is processed by a standard transformer encoder. We further investigate self-supervised pretraining for XRF representation learning using Masked Spectral modeling (MSM) and a physics-informed Peak Presence Prediction (PPP) objective. Experiments on the Pigments Checker STANDARD v.5 dataset for pigment identification and unmixing show that XRFormer consistently outperforms ViT, SpectralFormer (with and without CAF), and a 1D-CNN baseline for pigment identification. For pigment unmixing, XRFormer achieves robust abundance estimation while maintaining significantly higher parameter efficiency than SpectralFormer, operating at a lower token resolution (128 vs. 512 tokens) and with less than half the number of parameters (1.5M vs. 3.37M). MSM yields consistent gains across both tasks, while PPP further enhances performance for both identification and unmixing when tuned with an appropriate peak prominence. These results highlight multiscale, modality-aware tokenization as an effective and parameter efficient foundation for transformer-based XRF modeling under data-limited conditions. A GitHub repository is provided at https://github.com/sofiane1010/XRFormer.
Speckle reconstructive spectrometers (RSs) leverage the wavelength sensitivity of disordered media for spectral encoding, where the spectral information must be decoded via computational reconstruction algorithms. In recent years, machine learning has emerged as a powerful computational tool and has been applied to spectral reconstruction. However, existing learning-based methods predominantly rely on multi-layer artificial neural network architectures, which suffer from limited physical interpretability and insufficient generalization across diverse spectral types. Grounded in the insights of spectral-to-spatial mapping, especially its linear scattering nature, we introduce a physics-embedded model whose structure implicitly represents the inverse of this mapping. Compared with the standard architecture of convolutional neural networks and multilayer perceptrons, our learning-based method features a task-oriented design that incorporates domain knowledge of RSs, demonstrating superior generalization capability and a 30-fold improvement in inference speed. The proposed approach offers significant potential in applications such as real-time optical monitoring and hyperspectral imaging, among others.
Uni-XAS is presented, a unified benchmark and learning framework that reframes bidirectional XAS modeling as a cross-modal alignment and conditional generation problem, and introduces Permutation-Rectified Flow Matching, which integrates type-wise optimal transport into a continuous generative flow to provide a principled solution to ligand permutation ambiguity without relying on heavy high-order equivariant architectures.
Suyang Zhong, Yuhao Zhao, Boying Huang et al.· 0 citations
Recovering hyperspectral images (HSIs) from RGB observations is a highly ill-posed problem due to severe spectral information loss. However, current methods either rely on costly paired RGB–HSI datasets that are difficult to obtain or on unpaired RGB–HSI data for spectral guidance, which increases training costs and often leads to physically distorted spectral reconstructions. To address these challenges, we propose a self-supervised score-distilled spectral prior (SDSP) framework, which consists of a wavelet-based cross-attention reconstruction network (WCAR-Net) as the generator and a pretrained spectral diffusion model as the data-driven spectral expert. Score distillation sampling (SDS) is applied to the diffusion model to compute spectral gradients, which are incorporated into the generator’s loss to iteratively refine its predictions and produce high-quality hyperspectral reconstructions. In WCAR-Net, a cross-attention refinement module is designed to fuse spatial details with high-level semantic features, while wavelet-based feature decomposition preserves fine frequency-domain structures. Guided by the spectral expert provided by the diffusion model, the network is trained in a self-supervised and iterative manner, thereby eliminating the need for paired RGB–HSI data. Extensive experiments on multiple datasets demonstrate that our method significantly outperforms state-of-the-art approaches in both spectral fidelity and spatial reconstruction.
Ningjia Lv, Feiwang Yuan, Wei He et al.· IEEE Transactions on Geoscie...· 0 citations
PixRestore is presented, a VAE-free pixel-space Diffusion Transformer (DiT) for UIR, where the diffusion backbone is trained entirely from scratch, without relying on T2I pretraining.
Lingchen Sun, Rongyuan Wu, Xiangtao Kong et al.· 0 citations
Hyperspectral image classification (HSIC) relies on effective modeling of coupled spectral–spatial interactions. While recent CNN-, Transformer-, and Mamba-based methods have improved feature extraction, most of them organize token interaction through layerwise aggregation and repeated stacking, leaving cross-layer propagation only implicitly modeled. To address this issue, we propose SSH-Net, a symplectic-inspired spectral–spatial dynamic interaction network for HSIC. The proposed framework parameterizes token representations as latent position and momentum states and performs gated leapfrog-style dynamic updates to realize structured multistep feature propagation. On top of this dynamic backbone, a terminal-state descriptor branch summarizes the final position, momentum, and interaction-response statistics as terminal-state evidence, and a discriminative evidence aggregation (DEA) branch integrates such evidence with multiscale spatial and global contextual features for prediction. In this way, SSH-Net combines Hamiltonian-inspired dynamic propagation with complementary state-aware readout for hyperspectral classification. Experiments on four public benchmarks, namely Indian Pines, Houston2013, WHU-Hi-LongKou, and WHU-Hi-HanChuan, show that SSH-Net achieves consistently competitive or superior performance in terms of overall accuracy (OA), average accuracy (AA), and kappa coefficient compared with representative baseline methods. These results suggest that symplectic-inspired dynamic propagation provides a useful inductive bias for organizing spectral–spatial interaction in HSIC. Codes are available at https://github.com/yi1275174812/SSH-Net
Zihang Luo, Xiaofei Yang, Fei Yu et al.· IEEE Transactions on Geoscie...· 0 citations
Hyperspectral and multispectral image fusion (HMIF) aims to reconstruct a high-resolution hyperspectral image (HR-HSI) by combining the fine spatial details of a high-resolution multispectral image (HR-MSI) with the rich spectral information of a low-resolution hyperspectral image (LR-HSI). Recent advances in implicit neural representations (INRs) have enabled flexible coordinate-based modeling for HMIF; however, existing INR-based approaches may not fully capture fine-grained spatial structures and rich spectral dependencies. Moreover, the LR-HSI and HR-MSI are primarily incorporated through degradation-consistency constraints, leaving their complementary information underexploited. To address these limitations, we propose Two-Stage Reconstruction with Implicit Tensor Neural Representation (TSR-ITNR), a unified self-supervised framework integrating representation refinement and observation-guided calibration. In Stage 1, TSR-ITNR learns an implicit Tucker representation and refines its low-rank spatial coefficient tensor and spectral basis to better capture fine spatial structures and interband correlations. A fixed pretrained denoiser further provides a deep prior for the preliminary reconstruction. In Stage 2, parameter-free calibration derives complementary and noninterfering corrections from both observations to recover information insufficiently captured in Stage 1. Theoretical analysis establishes the geometry-preserving property of spectral refinement and the orthogonal complementarity of calibration. Extensive experiments on multiple benchmark datasets demonstrate strong quantitative, visual, and spectral reconstruction performance without ground-truth HR-HSI supervision. Beyond conventional reconstruction metrics, we further assess the effectiveness of TSR-ITNR using downstream semantic segmentation accuracy.
Unknown authors· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
