Infrared Physics & Technology最新文献_第7页

KANFuse: Enhancing infrared and visible image fusion through nonlinear representation modeling KANFuse：通过非线性表示建模增强红外和可见光图像的融合

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-15 DOI: 10.1016/j.infrared.2025.106319

Yongzi Zhang , Shengshi Li , Aolin Fang , Xinglong He , Daoheng Zhu , Keer Wu , Xiuchun Xiao

Infrared and Visible Image Fusion (IVIF) is a technique used to integrate thermal information from infrared images with the fine details and textures of visible images to achieve comprehensive scene perception. It has broad applications in night vision, surveillance, and autonomous driving, where highlighting thermal targets while retaining scene details is essential. However, existing methods often struggle to effectively preserve modality-specific features and suffer from limited nonlinear modeling capacity, which hinders their ability to fully exploit the complementary information across modalities. In this research, we proposed KANFuse, a novel fusion network in which Kolmogorov–Arnold Networks (KAN) are incorporated to model complex nonlinear cross-modal interactions. To further enhance representation, Wavelet Convolution Blocks (WCBs) are employed for edge-aware and noise-suppressing feature extraction, while Dynamic Fusion Modules (DFMs) are integrated into skip connections to balance multi-source contributions. Additionally, a spectral guided fidelity loss (SFL) is designed for second-phase training to better retain realistic visual information. Extensive evaluations on TNO,

M^{3} FD

, and LLVIP demonstrate that KANFuse consistently outperforms fourteen state-of-the-art methods in both qualitative and quantitative evaluations.

红外与可见光图像融合（IVIF）是一种将红外图像的热信息与可见光图像的精细细节和纹理进行融合以实现全面场景感知的技术。它在夜视、监视和自动驾驶中有着广泛的应用，在这些领域，突出热目标同时保留场景细节是必不可少的。然而，现有的方法往往难以有效地保留模态特定的特征，并且受到非线性建模能力的限制，这阻碍了它们充分利用模态之间的互补信息的能力。在这项研究中，我们提出了KANFuse，一种新的融合网络，其中包含Kolmogorov-Arnold网络（KAN）来模拟复杂的非线性跨模态相互作用。为了进一步增强表征，采用小波卷积块（WCBs）进行边缘感知和噪声抑制特征提取，并将动态融合模块（dfm）集成到跳过连接中以平衡多源贡献。此外，为第二阶段训练设计了光谱导引保真度损失（SFL），以更好地保留真实的视觉信息。对TNO、M3FD和LLVIP的广泛评估表明，KANFuse在定性和定量评估方面始终优于14种最先进的方法。

{"title":"KANFuse: Enhancing infrared and visible image fusion through nonlinear representation modeling","authors":"Yongzi Zhang , Shengshi Li , Aolin Fang , Xinglong He , Daoheng Zhu , Keer Wu , Xiuchun Xiao","doi":"10.1016/j.infrared.2025.106319","DOIUrl":"10.1016/j.infrared.2025.106319","url":null,"abstract":"<div><div>Infrared and Visible Image Fusion (IVIF) is a technique used to integrate thermal information from infrared images with the fine details and textures of visible images to achieve comprehensive scene perception. It has broad applications in night vision, surveillance, and autonomous driving, where highlighting thermal targets while retaining scene details is essential. However, existing methods often struggle to effectively preserve modality-specific features and suffer from limited nonlinear modeling capacity, which hinders their ability to fully exploit the complementary information across modalities. In this research, we proposed KANFuse, a novel fusion network in which Kolmogorov–Arnold Networks (KAN) are incorporated to model complex nonlinear cross-modal interactions. To further enhance representation, Wavelet Convolution Blocks (WCBs) are employed for edge-aware and noise-suppressing feature extraction, while Dynamic Fusion Modules (DFMs) are integrated into skip connections to balance multi-source contributions. Additionally, a spectral guided fidelity loss (SFL) is designed for second-phase training to better retain realistic visual information. Extensive evaluations on TNO, <span><math><mrow><msup><mrow><mi>M</mi></mrow><mrow><mn>3</mn></mrow></msup><mi>FD</mi></mrow></math></span>, and LLVIP demonstrate that KANFuse consistently outperforms fourteen state-of-the-art methods in both qualitative and quantitative evaluations.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106319"},"PeriodicalIF":3.4,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787449","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

IR signature surrogate model of ground-based rocket exhaust plumes based on convolutional neural network 基于卷积神经网络的陆基火箭排气羽流红外特征替代模型

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-13 DOI: 10.1016/j.infrared.2025.106320

Qinglin Niu , Panpan Yan , Xiaying Meng , Rong Lv , Hongli Wang

The traditional and widely used method for calculating the infrared (IR) radiation of the rocket exhaust plume is time consuming and not conducive to the generation of massive data. In this study, an “end-to-end” prediction surrogate model of IR radiation of exhaust plumes of ground motor based on a convolutional neural network (CNN) is proposed. It is used to calculate the IR radiation intensity of plume quickly only under the condition of using engine design and operating parameters as input. The dataset required for training is provided by a high-fidelity, full-link, plume IR radiation numerical method. Based on the obtained IR radiation intensity, the bilinear interpolation method was used to normalize the thermal images of radiation intensities of different sizes. Parallel decoders, asymmetric convolution blocks, and bottleneck layers were used to train the model. The root mean square error calculation loss function was used as the criterion for model accuracy. The results of the fine physics model were used to evaluate the calculation accuracy of the plume IR radiation surrogate model, CNN-IR. The results show that the regression coefficient of the established plume IR radiation CNN-IR model was close to 1, and the average error of the test set was 6.04%. Under different expansion states, the prediction error of the plume IR radiation was less than 10%, and the model showed high prediction accuracy and generalization ability. Compared with the traditional detailed numerical calculation model, the computational efficiency of the CNN-IR model was improved by four orders of magnitude, demonstrating extremely high computational efficiency. The CNN-IR model can provide a method for the rapid real-time prediction of plume IR radiation signals and the generation of massive numbers of sample data.

传统的计算火箭排气羽流红外辐射的方法既耗时又不利于大量数据的生成。本文提出了一种基于卷积神经网络（CNN）的接地电机排气羽流红外辐射“端到端”预测代理模型。该方法仅在以发动机设计参数和运行参数为输入条件下，才能快速计算出烟羽的红外辐射强度。训练所需的数据集由高保真度、全链接、羽流红外辐射数值方法提供。基于获得的红外辐射强度，采用双线性插值方法对不同尺寸辐射强度的热图像进行归一化处理。使用并行解码器、非对称卷积块和瓶颈层来训练模型。采用均方根误差计算损失函数作为模型精度的判据。利用精细物理模型的结果对羽流红外辐射替代模型CNN-IR的计算精度进行了评价。结果表明，所建立的羽流红外辐射CNN-IR模型回归系数接近于1，测试集的平均误差为6.04%。在不同膨胀状态下，羽流红外辐射预测误差小于10%，模型具有较高的预测精度和泛化能力。与传统的精细数值计算模型相比，CNN-IR模型的计算效率提高了4个数量级，显示出极高的计算效率。CNN-IR模型为羽流红外辐射信号的快速实时预测和大量样本数据的生成提供了一种方法。

{"title":"IR signature surrogate model of ground-based rocket exhaust plumes based on convolutional neural network","authors":"Qinglin Niu , Panpan Yan , Xiaying Meng , Rong Lv , Hongli Wang","doi":"10.1016/j.infrared.2025.106320","DOIUrl":"10.1016/j.infrared.2025.106320","url":null,"abstract":"<div><div>The traditional and widely used method for calculating the infrared (IR) radiation of the rocket exhaust plume is time consuming and not conducive to the generation of massive data. In this study, an “end-to-end” prediction surrogate model of IR radiation of exhaust plumes of ground motor based on a convolutional neural network (CNN) is proposed. It is used to calculate the IR radiation intensity of plume quickly only under the condition of using engine design and operating parameters as input. The dataset required for training is provided by a high-fidelity, full-link, plume IR radiation numerical method. Based on the obtained IR radiation intensity, the bilinear interpolation method was used to normalize the thermal images of radiation intensities of different sizes. Parallel decoders, asymmetric convolution blocks, and bottleneck layers were used to train the model. The root mean square error calculation loss function was used as the criterion for model accuracy. The results of the fine physics model were used to evaluate the calculation accuracy of the plume IR radiation surrogate model, CNN-IR. The results show that the regression coefficient of the established plume IR radiation CNN-IR model was close to 1, and the average error of the test set was 6.04%. Under different expansion states, the prediction error of the plume IR radiation was less than 10%, and the model showed high prediction accuracy and generalization ability. Compared with the traditional detailed numerical calculation model, the computational efficiency of the CNN-IR model was improved by four orders of magnitude, demonstrating extremely high computational efficiency. The CNN-IR model can provide a method for the rapid real-time prediction of plume IR radiation signals and the generation of massive numbers of sample data.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106320"},"PeriodicalIF":3.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787446","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Thermal evolution of skin temperature during rehabilitation from severe knee injury: A 12-month longitudinal case study of a female junior elite alpine ski racer 重度膝关节损伤康复过程中皮肤温度的热演化：一名年轻优秀高山滑雪女运动员12个月的纵向个案研究

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-13 DOI: 10.1016/j.infrared.2025.106315

Sebastian Färber, Ronja Mittermeier, Christoph Ebenbichler, Christian Raschner

This case study used infrared thermography (IRT) to monitor skin temperature (Tsk) during rehabilitation from a severe knee injury in a 21-year-old female professional alpine ski racer. The study aimed to track knee Tsk changes and detect thermal asymmetries between the injured leg (IL) and the non-injured leg (NIL) over 12 months (72–409 days post-surgery).

Initial measurements showed marked IL hyperthermia, with a mean contralateral Tsk difference (ΔTsk_mean) of 1.13 °C that decreased as healing progressed. Although ΔTsk_mean declined, the IL consistently exhibited higher Tsk in the anterior (p < 0.001, r = 0.97), medial (p = 0.028, r = 0.73), and lateral (p = 0.002, r = 0.92) regions. Posterior IL regions showed lower Tsk (p < 0.001, r = −0.97), suggesting compensatory mechanisms. Reductions in ΔTsk_mean coincided with lower training volumes and an intra-articular anti-inflammatory injection, whereas increases in ΔTsk_mean coincided with higher training volumes. By 13.5 months post-surgery, anterior asymmetries returned to healthy ranges. Menstrual cycle status did not affect Tsk at the knees.

These results demonstrate region-specific thermographic patterns and support IRT as a non-invasive tool for monitoring knee rehabilitation, assessing load tolerance, tracking healing progress, and supporting return-to-sport decisions. Further research is needed to define objective reference values.

本案例研究采用红外热像仪（IRT）监测一名21岁职业高山滑雪女运动员严重膝关节损伤康复期间的皮肤温度（Tsk）。该研究旨在追踪膝关节Tsk变化，并检测受伤腿（IL）和未受伤腿（NIL）在12个月内（术后72-409天）的热不对称性。最初的测量显示明显的IL热疗，平均对侧Tsk差（ΔTskmean）为1.13°C，随着愈合的进展而降低。虽然ΔTskmean下降，但IL在前区（p < 0.001, r = 0.97）、内侧区（p = 0.028, r = 0.73）和外侧区（p = 0.002, r = 0.92）始终表现出较高的Tsk。后IL区显示较低的Tsk (p < 0.001, r = - 0.97)，提示代偿机制。ΔTskmean的减少与训练量的减少和关节内抗炎注射相一致，而ΔTskmean的增加与训练量的增加相一致。术后13.5个月，前路不对称恢复到健康范围。月经周期状况不影响膝盖处的Tsk。这些结果证明了区域特异性热成像模式，并支持IRT作为监测膝关节康复、评估负荷耐受性、跟踪愈合进展和支持重返运动决策的非侵入性工具。需要进一步的研究来确定客观的参考值。

{"title":"Thermal evolution of skin temperature during rehabilitation from severe knee injury: A 12-month longitudinal case study of a female junior elite alpine ski racer","authors":"Sebastian Färber, Ronja Mittermeier, Christoph Ebenbichler, Christian Raschner","doi":"10.1016/j.infrared.2025.106315","DOIUrl":"10.1016/j.infrared.2025.106315","url":null,"abstract":"<div><div>This case study used infrared thermography (IRT) to monitor skin temperature (Tsk) during rehabilitation from a severe knee injury in a 21-year-old female professional alpine ski racer. The study aimed to track knee Tsk changes and detect thermal asymmetries between the injured leg (IL) and the non-injured leg (NIL) over 12 months (72–409 days post-surgery).</div><div>Initial measurements showed marked IL hyperthermia, with a mean contralateral Tsk difference (ΔTsk<sub>mean</sub>) of 1.13 °C that decreased as healing progressed. Although ΔTsk<sub>mean</sub> declined, the IL consistently exhibited higher Tsk in the anterior (p < 0.001, r = 0.97), medial (p = 0.028, r = 0.73), and lateral (p = 0.002, r = 0.92) regions. Posterior IL regions showed lower Tsk (p < 0.001, r = −0.97), suggesting compensatory mechanisms. Reductions in ΔTsk<sub>mean</sub> coincided with lower training volumes and an intra-articular anti-inflammatory injection, whereas increases in ΔTsk<sub>mean</sub> coincided with higher training volumes. By 13.5 months post-surgery, anterior asymmetries returned to healthy ranges. Menstrual cycle status did not affect Tsk at the knees.</div><div>These results demonstrate region-specific thermographic patterns and support IRT as a non-invasive tool for monitoring knee rehabilitation, assessing load tolerance, tracking healing progress, and supporting return-to-sport decisions. Further research is needed to define objective reference values.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106315"},"PeriodicalIF":3.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787017","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Modification of Al/porous SiO2 composite films via Ag particles for infrared-visible compatible stealth Ag粒子改性铝/多孔SiO2复合膜的红外-可见兼容隐身研究

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-13 DOI: 10.1016/j.infrared.2025.106303

Tingzhen Chen , Quanjiang Lv , Siwei Liu , Haigang Hou , Guiwu Liu , Junlin Liu , Guanjun Qiao

The advancement of infrared-visible compatible stealth materials is pivotal for strengthening national defense security and driving the next frontier in stealth material technologies. Here, we propose a modification strategy for Al/porous SiO₂ (Al/p-SiO₂) composite films by incorporating Ag particles to modulate their spectral response, thereby achieving infrared-visible compatible stealth. The most outstanding Al/p-SiO₂/Ag-20 exhibits excellent thermal stability, with emissivity of 0.27 in the mid-wave infrared (3-5 μm) and 0.19 in the long-wave infrared (8-14 μm) bands. In addition, compared with Al/p-SiO₂, its visible reflectivity is reduced by 0.35, reaching as low as 0.38 (380-780 nm). To meet diverse application requirements, flexible and patterned Al/p-SiO₂/Ag-20 composite film was fabricated, exhibiting excellent stealth performance, particularly suitable for wearable devices and fabrics. It is worth mentioning that the fabrication of Al/p-SiO₂/Ag composite films only requires a simple, low-cost and scalable process, which provides a solid foundation for large-scale application of infrared-visible compatible stealth materials.

红外-可见光兼容隐身材料的发展对于加强国防安全、推动隐身材料技术发展具有重要意义。本文提出了一种Al/多孔SiO2 （Al/p-SiO2）复合薄膜的改性策略，通过加入Ag颗粒来调节其光谱响应，从而实现红外-可见兼容隐身。Al/p-SiO2/Ag-20表现出优异的热稳定性，在中波红外（3-5 μm）和长波红外（8-14 μm）波段的发射率分别为0.27和0.19。此外，与Al/p-SiO2相比，其可见光反射率降低了0.35，低至0.38 （380 ~ 780 nm）。为了满足多样化的应用需求，制备了具有柔性和图案化的Al/p-SiO2/Ag-20复合薄膜，具有优异的隐身性能，特别适用于可穿戴设备和织物。值得一提的是，Al/p-SiO2/Ag复合薄膜的制备只需要一个简单、低成本和可扩展的工艺，这为红外可见兼容隐身材料的大规模应用提供了坚实的基础。

{"title":"Modification of Al/porous SiO2 composite films via Ag particles for infrared-visible compatible stealth","authors":"Tingzhen Chen , Quanjiang Lv , Siwei Liu , Haigang Hou , Guiwu Liu , Junlin Liu , Guanjun Qiao","doi":"10.1016/j.infrared.2025.106303","DOIUrl":"10.1016/j.infrared.2025.106303","url":null,"abstract":"<div><div>The advancement of infrared-visible compatible stealth materials is pivotal for strengthening national defense security and driving the next frontier in stealth material technologies. Here, we propose a modification strategy for Al/porous SiO<sub>2</sub> (Al/p-SiO<sub>2</sub>) composite films by incorporating Ag particles to modulate their spectral response, thereby achieving infrared-visible compatible stealth. The most outstanding Al/p-SiO<sub>2</sub>/Ag-20 exhibits excellent thermal stability, with emissivity of 0.27 in the mid-wave infrared (3-5 μm) and 0.19 in the long-wave infrared (8-14 μm) bands. In addition, compared with Al/p-SiO<sub>2</sub>, its visible reflectivity is reduced by 0.35, reaching as low as 0.38 (380-780 nm). To meet diverse application requirements, flexible and patterned Al/p-SiO<sub>2</sub>/Ag-20 composite film was fabricated, exhibiting excellent stealth performance, particularly suitable for wearable devices and fabrics. It is worth mentioning that the fabrication of Al/p-SiO<sub>2</sub>/Ag composite films only requires a simple, low-cost and scalable process, which provides a solid foundation for large-scale application of infrared-visible compatible stealth materials.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106303"},"PeriodicalIF":3.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787019","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Single-spectral passive infrared imaging method for gas concentration inversion based on reference gas 基于参比气体的气体浓度反演单光谱被动红外成像方法

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-13 DOI: 10.1016/j.infrared.2025.106307

Xianghui Wang, Zhihao Yuan, Gui Xiao, Chong Kang, Haoyan Liu

In recent years, hazardous gas leakage incidents have occurred frequently, attracting heightened global attention to industrial safety. Infrared optical gas imaging technology has received significant attention due to its advantages, including distinct gas absorption spectral lines, suitability for low-concentration leak detection, relatively simple structure, and low cost. However, its performance limitations in quantitative measurement processes still hinder its widespread adoption. To address this issue, this study proposes a single-spectral passive infrared imaging method for gas concentration inversion based on reference gas. By preconfiguring the internal parameters of the reference gas and relevant backgrounds, this method enables concentration inversion in practical applications by merely acquiring the radiation values of the target gas. To verify the feasibility, inversion accuracy, and anti-interference capability of the proposed method, gas concentration inversion simulations were conducted, and experimental comparisons were performed with two conventional methods: the absolute difference-based concentration inversion method and the dual background temperature-based concentration inversion method. Experimental results demonstrate that: in multi-concentration methane tests, the reference gas method exhibits an average error of ≤2.6 %, representing a reduction of over 60 % compared to the two comparative methods (≤7.1 % and ≤11.6 %, respectively); under mixed noise scenarios, its average error is ≤6.7 %, with a reduction of over 40 % compared to the aforementioned methods (≤11.3 % and ≤11.8 %, respectively). Furthermore, the proposed method can still achieve global gas density reconstruction even under low leakage rates.

近年来，危险气体泄漏事件频发，引起了全球对工业安全的高度关注。红外光学气体成像技术因其具有气体吸收谱线清晰、适合低浓度泄漏检测、结构相对简单、成本低等优点而备受关注。然而，它在定量测量过程中的性能限制仍然阻碍了它的广泛采用。针对这一问题，本研究提出了一种基于参比气体的单光谱被动红外成像气体浓度反演方法。该方法通过预先配置参比气体的内部参数和相关背景，在实际应用中仅通过获取目标气体的辐射值即可实现浓度反演。为验证所提方法的可行性、反演精度和抗干扰能力，进行了气体浓度反演仿真，并与基于绝对差分的浓度反演方法和基于双背景温度的浓度反演方法进行了实验比较。实验结果表明：在多浓度甲烷测试中，参考气体法的平均误差≤2.6%，与两种比较方法（分别≤7.1%和≤11.6%）相比，误差减小60%以上；在混合噪声情况下，其平均误差≤6.7%，比上述方法（分别≤11.3%和≤11.8%）降低了40%以上。此外，即使在低泄漏率下，该方法仍然可以实现全局气体密度重建。

{"title":"Single-spectral passive infrared imaging method for gas concentration inversion based on reference gas","authors":"Xianghui Wang, Zhihao Yuan, Gui Xiao, Chong Kang, Haoyan Liu","doi":"10.1016/j.infrared.2025.106307","DOIUrl":"10.1016/j.infrared.2025.106307","url":null,"abstract":"<div><div>In recent years, hazardous gas leakage incidents have occurred frequently, attracting heightened global attention to industrial safety. Infrared optical gas imaging technology has received significant attention due to its advantages, including distinct gas absorption spectral lines, suitability for low-concentration leak detection, relatively simple structure, and low cost. However, its performance limitations in quantitative measurement processes still hinder its widespread adoption. To address this issue, this study proposes a single-spectral passive infrared imaging method for gas concentration inversion based on reference gas. By preconfiguring the internal parameters of the reference gas and relevant backgrounds, this method enables concentration inversion in practical applications by merely acquiring the radiation values of the target gas. To verify the feasibility, inversion accuracy, and anti-interference capability of the proposed method, gas concentration inversion simulations were conducted, and experimental comparisons were performed with two conventional methods: the absolute difference-based concentration inversion method and the dual background temperature-based concentration inversion method. Experimental results demonstrate that: in multi-concentration methane tests, the reference gas method exhibits an average error of ≤2.6 %, representing a reduction of over 60 % compared to the two comparative methods (≤7.1 % and ≤11.6 %, respectively); under mixed noise scenarios, its average error is ≤6.7 %, with a reduction of over 40 % compared to the aforementioned methods (≤11.3 % and ≤11.8 %, respectively). Furthermore, the proposed method can still achieve global gas density reconstruction even under low leakage rates.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106307"},"PeriodicalIF":3.4,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787451","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Tomato seed variety identification using hyperspectral imaging with deep learning classification 基于深度学习分类的高光谱成像番茄种子品种识别

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-12 DOI: 10.1016/j.infrared.2025.106317

Hengnian Qi , Shuqiang Hu , Junyi Chen , Xiaoping Wu , Xuhua Zhu , Jianfang Yan , Chu Zhang

Rapid and non-destructive identification of tomato seed varieties is critical for ensuring seed quality and supporting agricultural production. In this study, near-infrared hyperspectral imaging was applied to acquire spectral data from 7894 seeds across four representative tomato varieties. To improve classification performance, a deep learning framework that integrates Convolutional Neural Networks (CNN) with Kolmogorov-Arnold Networks (KAN) was proposed. Spectral data were preprocessed using six methods—Multiplicative Scatter Correction (MSC), Standard Normal Variate (SNV), First Derivative (FD), Local Trend First Derivative (LTFD), Second Derivative (SD), and Local Trend Second Derivative (LTSD)—prior to model training. Results showed that the CNN–KAN fusion model consistently outperformed both the standalone CNN and KAN models. Under MSC preprocessing, a peak accuracy of 89.87 % was attained, and variance across repeated trials was reduced, indicating superior robustness and stability. Furthermore, feature visualization with the DisCo-FFS method revealed that the 1100–1500 nm region contributed most to varietal discrimination. These findings highlight the effectiveness of combining hyperspectral imaging with advanced deep learning models and feature selection techniques, offering a rapid, non-destructive, and interpretable approach for seed variety identification and quality assessment.

快速、无损的番茄种子品种鉴定是保证种子质量和支持农业生产的重要手段。利用近红外高光谱成像技术，对4个代表性番茄品种7894颗种子进行了光谱分析。为了提高分类性能，提出了一种融合卷积神经网络（CNN）和Kolmogorov-Arnold网络（KAN）的深度学习框架。在模型训练之前，使用六种方法对光谱数据进行预处理：乘法散点校正（MSC）、标准正态变量（SNV）、一阶导数（FD）、局部趋势一阶导数（LTFD）、二阶导数（SD）和局部趋势二阶导数（LTSD）。结果表明，CNN - KAN融合模型始终优于单独的CNN和KAN模型。在MSC预处理下，达到89.87%的峰值准确率，并且重复试验的方差减少，表明具有较好的鲁棒性和稳定性。此外，利用DisCo-FFS方法进行特征可视化，发现1100 ~ 1500 nm区域对品种识别贡献最大。这些发现强调了将高光谱成像与先进的深度学习模型和特征选择技术相结合的有效性，为种子品种鉴定和质量评估提供了一种快速、无损和可解释的方法。

{"title":"Tomato seed variety identification using hyperspectral imaging with deep learning classification","authors":"Hengnian Qi , Shuqiang Hu , Junyi Chen , Xiaoping Wu , Xuhua Zhu , Jianfang Yan , Chu Zhang","doi":"10.1016/j.infrared.2025.106317","DOIUrl":"10.1016/j.infrared.2025.106317","url":null,"abstract":"<div><div>Rapid and non-destructive identification of tomato seed varieties is critical for ensuring seed quality and supporting agricultural production. In this study, near-infrared hyperspectral imaging was applied to acquire spectral data from 7894 seeds across four representative tomato varieties. To improve classification performance, a deep learning framework that integrates Convolutional Neural Networks (CNN) with Kolmogorov-Arnold Networks (KAN) was proposed. Spectral data were preprocessed using six methods—Multiplicative Scatter Correction (MSC), Standard Normal Variate (SNV), First Derivative (FD), Local Trend First Derivative (LTFD), Second Derivative (SD), and Local Trend Second Derivative (LTSD)—prior to model training. Results showed that the CNN–KAN fusion model consistently outperformed both the standalone CNN and KAN models. Under MSC preprocessing, a peak accuracy of 89.87 % was attained, and variance across repeated trials was reduced, indicating superior robustness and stability. Furthermore, feature visualization with the DisCo-FFS method revealed that the 1100–1500 nm region contributed most to varietal discrimination. These findings highlight the effectiveness of combining hyperspectral imaging with advanced deep learning models and feature selection techniques, offering a rapid, non-destructive, and interpretable approach for seed variety identification and quality assessment.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106317"},"PeriodicalIF":3.4,"publicationDate":"2025-12-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787444","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Determination of fatty acids in peanuts by NIR combined with CARS-PLS 近红外光谱联合CARS-PLS测定花生中脂肪酸

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-12 DOI: 10.1016/j.infrared.2025.106316

Wanqing Yao , Junhao Ye , Xiangning Sun , Pingjing Chang , Jilin Guo , Xianhu Liu , Guohua Zhong

The traditional method for determining peanut fatty acid content, which relies on labor-intensive and destructive chromatographic techniques, faces practical limitations when applied to large-scale industrial processes. This study introduces a novel Competitive Adaptive Reweighted Sampling-Partial Least Squares (CARS-PLS) algorithm designed to enable rapid, non-destructive quantification of six major fatty acids, palmitic, stearic, oleic, linoleic, arachidic, and cis-11-eicosenoic acids, using near-infrared (NIR) spectroscopy. Through iterative optimization of wavelength selection using Monte Carlo sampling and adaptive reweighting, CARS-PLS effectively tackles spectral redundancy and nonlinearity, outperforming traditional linear methods such as full-spectrum PLS and PCA-PLS. The model exhibited outstanding predictive accuracy, with calibration coefficients (R_c²) and prediction coefficients (R_p²) exceeding 0.97 for major fatty acids, root mean square errors of prediction (RMSEP) below 0.35 g/100 g, and residual prediction deviation (RPD) values greater than 2.5. External validation using independent samples confirmed the robustness of the method, with relative prediction errors remaining under 10 % for all analytes. The CARS-PLS framework, known to resist overfitting and its adaptability to complex spectral datasets, provides a transformative tool for real-time quality control in peanut processing chains, with potential applications extending to other edible oil analyses.

传统的花生脂肪酸含量测定方法依赖于劳动密集型和破坏性的色谱技术，在大规模工业生产过程中存在实际局限性。本研究介绍了一种新颖的竞争自适应重加权采样-偏最小二乘（car - pls）算法，该算法旨在使用近红外（NIR）光谱快速、无损地定量六种主要脂肪酸，棕榈酸、硬脂酸、油酸、亚油酸、花生酸和顺式11-二十烯酸。通过蒙特卡罗采样和自适应加权的波长选择迭代优化，CARS-PLS有效地解决了光谱冗余和非线性问题，优于传统的线性方法，如全光谱PLS和PCA-PLS。该模型对主要脂肪酸的校正系数（Rc2）和预测系数（Rp2）均大于0.97，预测均方根误差（RMSEP）小于0.35 g/100 g，剩余预测偏差（RPD）大于2.5。使用独立样本的外部验证证实了该方法的稳健性，所有分析物的相对预测误差保持在10%以下。CARS-PLS框架具有抗过拟合和对复杂光谱数据集的适应性，为花生加工链的实时质量控制提供了一种变革性工具，具有扩展到其他食用油分析的潜在应用前景。

{"title":"Determination of fatty acids in peanuts by NIR combined with CARS-PLS","authors":"Wanqing Yao , Junhao Ye , Xiangning Sun , Pingjing Chang , Jilin Guo , Xianhu Liu , Guohua Zhong","doi":"10.1016/j.infrared.2025.106316","DOIUrl":"10.1016/j.infrared.2025.106316","url":null,"abstract":"<div><div>The traditional method for determining peanut fatty acid content, which relies on labor-intensive and destructive chromatographic techniques, faces practical limitations when applied to large-scale industrial processes. This study introduces a novel Competitive Adaptive Reweighted Sampling-Partial Least Squares (CARS-PLS) algorithm designed to enable rapid, non-destructive quantification of six major fatty acids, palmitic, stearic, oleic, linoleic, arachidic, and <em>cis</em>-11-eicosenoic acids, using near-infrared (NIR) spectroscopy. Through iterative optimization of wavelength selection using Monte Carlo sampling and adaptive reweighting, CARS-PLS effectively tackles spectral redundancy and nonlinearity, outperforming traditional linear methods such as full-spectrum PLS and PCA-PLS. The model exhibited outstanding predictive accuracy, with calibration coefficients (<em>R<sub>c</sub><sup>2</sup></em>) and prediction coefficients (<em>R<sub>p</sub></em><sup><em>2</em></sup>) exceeding 0.97 for major fatty acids, root mean square errors of prediction (RMSEP) below 0.35 <em>g/100 g</em>, and residual prediction deviation (RPD) values greater than 2.5. External validation using independent samples confirmed the robustness of the method, with relative prediction errors remaining under 10 % for all analytes. The CARS-PLS framework, known to resist overfitting and its adaptability to complex spectral datasets, provides a transformative tool for real-time quality control in peanut processing chains, with potential applications extending to other edible oil analyses.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106316"},"PeriodicalIF":3.4,"publicationDate":"2025-12-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145787448","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Second harmonic signal reconstruction and noise elimination in methane detection via Quartz-Enhanced photoacoustic spectroscopy using an ESMD-AE-WTD model 基于ESMD-AE-WTD模型的石英增强光声光谱甲烷检测中的二次谐波信号重建和噪声消除

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-11 DOI: 10.1016/j.infrared.2025.106312

Tingting Zhang, Minghe Wu, Qinduan Zhang, Jiqiang Wang, Yefeng Gu, Wei Wang, Chunsheng Li, Hongzhen Yu, Guo Sun, Dong Li

Quartz-enhanced photoacoustic spectroscopy (QEPAS) enables highly sensitive and selective trace gas detection. This study proposes an ESMD-AE-WTD model for second harmonic signal reconstruction and noise elimination in QEPAS sensing systems. The model integrates Extreme-point Symmetric Mode Decomposition (ESMD), Approximate Entropy (AE), and Wavelet Threshold Denoising (WTD) to adaptively separate and suppress noise. Specifically, ESMD decomposes the noisy signal and extracts dominant components, AE quantitatively evaluates and selects effective modes, and WTD applies threshold filtering to remove residual noise and reconstruct the signal. Experimental results show that for methane (CH₄) detection at 800 ppm, the signal-to-noise (SNR) ratio increases from 69.90 to 443.22, corresponding to a 6.34-fold improvement. Compared with conventional denoising techniques such as WTD, Savitzky-Golay(S-G) filtering, and the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN)-S-G algorithm, the proposed approach achieves superior performance and significantly enhances the accuracy of QEPAS-based gas sensing. These findings demonstrate that the ESMD-AE-WTD model has strong potential for CH₄ detection and provides a solid foundation for extending photoacoustic spectroscopy (PAS) to environmental monitoring, industrial control, safety, and medical diagnostics.

石英增强光声光谱（QEPAS）实现高灵敏度和选择性的痕量气体检测。本文提出了一种用于QEPAS传感系统中二次谐波信号重构和噪声消除的ESMD-AE-WTD模型。该模型结合了极值对称模态分解（ESMD）、近似熵（AE）和小波阈值去噪（WTD），实现了对噪声的自适应分离和抑制。其中，ESMD分解噪声信号并提取优势分量，AE定量评价并选择有效模式，WTD采用阈值滤波去除残余噪声，重构信号。实验结果表明，在800 ppm下检测甲烷（CH4）时，信噪比从69.90提高到443.22，提高了6.34倍。与WTD、Savitzky-Golay（S-G）滤波、改进的全集成经验模态分解与自适应噪声(ICEEMDAN)-S-G算法等传统去噪技术相比，该方法具有优越的性能，显著提高了基于qepas的气体传感精度。这些发现表明，ESMD-AE-WTD模型具有很强的CH4检测潜力，为将光声光谱（PAS）扩展到环境监测、工业控制、安全和医疗诊断等领域提供了坚实的基础。

{"title":"Second harmonic signal reconstruction and noise elimination in methane detection via Quartz-Enhanced photoacoustic spectroscopy using an ESMD-AE-WTD model","authors":"Tingting Zhang, Minghe Wu, Qinduan Zhang, Jiqiang Wang, Yefeng Gu, Wei Wang, Chunsheng Li, Hongzhen Yu, Guo Sun, Dong Li","doi":"10.1016/j.infrared.2025.106312","DOIUrl":"10.1016/j.infrared.2025.106312","url":null,"abstract":"<div><div>Quartz-enhanced photoacoustic spectroscopy (QEPAS) enables highly sensitive and selective trace gas detection. This study proposes an ESMD-AE-WTD model for second harmonic signal reconstruction and noise elimination in QEPAS sensing systems. The model integrates Extreme-point Symmetric Mode Decomposition (ESMD), Approximate Entropy (AE), and Wavelet Threshold Denoising (WTD) to adaptively separate and suppress noise. Specifically, ESMD decomposes the noisy signal and extracts dominant components, AE quantitatively evaluates and selects effective modes, and WTD applies threshold filtering to remove residual noise and reconstruct the signal. Experimental results show that for methane (CH<sub>4</sub>) detection at 800 ppm, the signal-to-noise (SNR) ratio increases from 69.90 to 443.22, corresponding to a 6.34-fold improvement. Compared with conventional denoising techniques such as WTD, Savitzky-Golay(S-G) filtering, and the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN)-S-G algorithm, the proposed approach achieves superior performance and significantly enhances the accuracy of QEPAS-based gas sensing. These findings demonstrate that the ESMD-AE-WTD model has strong potential for CH<sub>4</sub> detection and provides a solid foundation for extending photoacoustic spectroscopy (PAS) to environmental monitoring, industrial control, safety, and medical diagnostics.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106312"},"PeriodicalIF":3.4,"publicationDate":"2025-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145733683","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Infrared target detection model based on global edge feature extraction 基于全局边缘特征提取的红外目标检测模型

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-11 DOI: 10.1016/j.infrared.2025.106313

Mei Da , Yixuan Shen , Lin Jiang , Jue Hu , Zhijian Zhang

Infrared detection missions are of critical importance for night surveillance, search and rescue operations and border security, especially in low-light or total darkness environments. In such conditions, infrared imaging can capture the thermal radiation signature of a target, thus guaranteeing the continuity and effectiveness of surveillance activities. However, the detected signal is weak, the signal-to-noise ratio is low, and it is susceptible to being drowned by the background. Furthermore, the number of features is limited, and they lack specific shapes, thus preventing the acquisition of their texture information. To address the aforementioned issues, a novel infrared target detection model with global edge feature extraction is proposed. First, we design a global edge feature extraction backbone network (GENet) to improve the feature extraction efficiency and achieve accurate real-time feature extraction. Furthermore, to further enhance the learning ability of the network by fusing multi-dimensional feature information, a new C3_DWR module is designed to enhance the feature expressiveness and context-awareness to improve the detection performance of the target. Subsequently, a multi-scale detection head is designed to capture fine-grained details and high-level semantic information. This is crucial for accurately identifying and localising objects in diverse and complex scenes. Finally, the SIoU loss function is used to increase the convergence speed of the model and reduce losses to enhance the robustness of the model. We conducted comparative experiments on the infrared dataset (IData) used in this paper, as well as the publicly available FLIR, HIT-UAV, and IVdata infrared datasets. The results showed that GDM-YOLO achieved an average accuracy rate of 90.3% on the IData dataset ([email protected]) and an accuracy rate of 85.1% on the FLIR dataset. the average accuracy on the HIT-UAV dataset ([email protected]) reached 80.6%, and on the IVdata maritime infrared dataset it reached 89.8%, validating the feasibility of the algorithm proposed in this paper.

红外探测任务对于夜间监视、搜救行动和边境安全至关重要，特别是在低光或完全黑暗的环境中。在这种情况下，红外成像可以捕获目标的热辐射特征，从而保证监视活动的连续性和有效性。但检测到的信号较弱，信噪比较低，容易被背景淹没。此外，特征的数量有限，缺乏特定的形状，从而阻碍了纹理信息的获取。针对上述问题，提出了一种基于全局边缘特征提取的红外目标检测模型。首先，设计了一个全局边缘特征提取骨干网（GENet），提高了特征提取效率，实现了精确的实时特征提取；此外，为了通过融合多维特征信息进一步增强网络的学习能力，设计了新的C3_DWR模块，增强特征表达能力和上下文感知能力，提高目标的检测性能。随后，设计了一个多尺度检测头来捕获细粒度细节和高级语义信息。这对于在各种复杂场景中准确识别和定位物体至关重要。最后，利用SIoU损失函数提高模型的收敛速度，减小损失，增强模型的鲁棒性。我们在本文使用的红外数据集（IData）以及公开的FLIR、HIT-UAV和IVdata红外数据集上进行了对比实验。结果表明，GDM-YOLO在IData数据集（[email protected]）上的平均准确率为90.3%，在FLIR数据集上的平均准确率为85.1%。在HIT-UAV数据集（[email protected]）上的平均精度达到80.6%，在IVdata海上红外数据集上的平均精度达到89.8%，验证了本文算法的可行性。

{"title":"Infrared target detection model based on global edge feature extraction","authors":"Mei Da , Yixuan Shen , Lin Jiang , Jue Hu , Zhijian Zhang","doi":"10.1016/j.infrared.2025.106313","DOIUrl":"10.1016/j.infrared.2025.106313","url":null,"abstract":"<div><div>Infrared detection missions are of critical importance for night surveillance, search and rescue operations and border security, especially in low-light or total darkness environments. In such conditions, infrared imaging can capture the thermal radiation signature of a target, thus guaranteeing the continuity and effectiveness of surveillance activities. However, the detected signal is weak, the signal-to-noise ratio is low, and it is susceptible to being drowned by the background. Furthermore, the number of features is limited, and they lack specific shapes, thus preventing the acquisition of their texture information. To address the aforementioned issues, a novel infrared target detection model with global edge feature extraction is proposed. First, we design a global edge feature extraction backbone network (GENet) to improve the feature extraction efficiency and achieve accurate real-time feature extraction. Furthermore, to further enhance the learning ability of the network by fusing multi-dimensional feature information, a new C3_DWR module is designed to enhance the feature expressiveness and context-awareness to improve the detection performance of the target. Subsequently, a multi-scale detection head is designed to capture fine-grained details and high-level semantic information. This is crucial for accurately identifying and localising objects in diverse and complex scenes. Finally, the SIoU loss function is used to increase the convergence speed of the model and reduce losses to enhance the robustness of the model. We conducted comparative experiments on the infrared dataset (IData) used in this paper, as well as the publicly available FLIR, HIT-UAV, and IVdata infrared datasets. The results showed that GDM-YOLO achieved an average accuracy rate of 90.3% on the IData dataset ([email protected]) and an accuracy rate of 85.1% on the FLIR dataset. the average accuracy on the HIT-UAV dataset ([email protected]) reached 80.6%, and on the IVdata maritime infrared dataset it reached 89.8%, validating the feasibility of the algorithm proposed in this paper.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106313"},"PeriodicalIF":3.4,"publicationDate":"2025-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145733681","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Improving joint localization of humans in infrared images via domain adaptation and transformer-GCN based pose estimation 利用域自适应和基于变换gcn的姿态估计改进红外图像中人体关节的定位

IF 3.4 3区物理与天体物理 Q2 INSTRUMENTS & INSTRUMENTATION

Infrared Physics & Technology

Pub Date : 2025-12-10 DOI: 10.1016/j.infrared.2025.106299

Avinash Upadhyay, Ankit Shukla, Manoj Sharma

Infrared (IR) images pose significant challenges for human pose estimation due to their limited interpretability, where edge information is often obscured in thermal signatures. Furthermore, IR images exhibit high non-uniformity resulting from varying emissivity and temperature differences across different materials. These characteristics make it particularly difficult to extract relevant features from IR images, which are essential for accurately estimating body joint locations. Additionally, occlusions of body joints are prevalent in IR images, as thermal signatures from different body parts can appear similar, causing joints to be perceived as a single entity. To address these challenges, we propose a two-fold solution. Firstly, we introduce a domain adaptation network-based feature extraction method, designed to capture robust edge features within IR images. Secondly, we present a novel Thermal-TransGCN network that processes these features to learn both global joint relationships by using self-attention mechanism and inter-joint relationships by considering each joint as node of the graph and then learning the node representations, effectively localizing joints in humans captured via infrared sensors. Our proposed approach enables precise joint localization, even in the presence of occlusions, and significantly improves the robustness of feature extraction and accuracy of pose estimation. Experimental results demonstrate that our method outperforms existing techniques in terms of joint localization accuracy, setting a new benchmark for pose estimation tasks on the LWIRPOSE dataset in IR imagery. This approach has potential applications in surveillance, sports analysis, and human–computer interaction.

红外（IR）图像由于其有限的可解释性而对人体姿态估计提出了重大挑战，其中边缘信息通常在热特征中被掩盖。此外，由于不同材料的发射率和温度差异不同，红外图像表现出高度的不均匀性。这些特征使得从红外图像中提取相关特征变得特别困难，而这些特征对于准确估计人体关节位置至关重要。此外，身体关节的闭塞在红外图像中很普遍，因为来自不同身体部位的热特征可能看起来相似，导致关节被视为一个单一的实体。为了应对这些挑战，我们提出了一个双重解决方案。首先，我们引入了一种基于域自适应网络的特征提取方法，旨在捕获红外图像中的鲁棒边缘特征。其次，我们提出了一种新的Thermal-TransGCN网络，该网络对这些特征进行处理，利用自关注机制学习全局关节关系，并通过将每个关节作为图的节点学习节点表示来学习关节间关系，从而有效地定位红外传感器捕获的人体关节。我们提出的方法能够在存在咬合的情况下实现精确的关节定位，并显著提高特征提取的鲁棒性和姿态估计的准确性。实验结果表明，该方法在联合定位精度方面优于现有技术，为红外图像中LWIRPOSE数据集的姿态估计任务设定了新的基准。这种方法在监视、运动分析和人机交互方面有潜在的应用。

{"title":"Improving joint localization of humans in infrared images via domain adaptation and transformer-GCN based pose estimation","authors":"Avinash Upadhyay, Ankit Shukla, Manoj Sharma","doi":"10.1016/j.infrared.2025.106299","DOIUrl":"10.1016/j.infrared.2025.106299","url":null,"abstract":"<div><div>Infrared (IR) images pose significant challenges for human pose estimation due to their limited interpretability, where edge information is often obscured in thermal signatures. Furthermore, IR images exhibit high non-uniformity resulting from varying emissivity and temperature differences across different materials. These characteristics make it particularly difficult to extract relevant features from IR images, which are essential for accurately estimating body joint locations. Additionally, occlusions of body joints are prevalent in IR images, as thermal signatures from different body parts can appear similar, causing joints to be perceived as a single entity. To address these challenges, we propose a two-fold solution. Firstly, we introduce a domain adaptation network-based feature extraction method, designed to capture robust edge features within IR images. Secondly, we present a novel Thermal-TransGCN network that processes these features to learn both global joint relationships by using self-attention mechanism and inter-joint relationships by considering each joint as node of the graph and then learning the node representations, effectively localizing joints in humans captured via infrared sensors. Our proposed approach enables precise joint localization, even in the presence of occlusions, and significantly improves the robustness of feature extraction and accuracy of pose estimation. Experimental results demonstrate that our method outperforms existing techniques in terms of joint localization accuracy, setting a new benchmark for pose estimation tasks on the LWIRPOSE dataset in IR imagery. This approach has potential applications in surveillance, sports analysis, and human–computer interaction.</div></div>","PeriodicalId":13549,"journal":{"name":"Infrared Physics & Technology","volume":"153 ","pages":"Article 106299"},"PeriodicalIF":3.4,"publicationDate":"2025-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145733678","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"物理与天体物理","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0