Localized burn damage on bearing raceways can alter the near surface microstructure and thermophysical properties, thereby reducing bearing service life. However, such defects usually show weak visual features, making rapid and reliable identification difficult. To address this problem, an active infrared thermal wave detection method based on modulated laser excitation is proposed in this study. First, the burned region on the bearing raceway is equivalent to a two layer structure consisting of a surface burn layer and a sound substrate. A one dimensional multilayer heat conduction model is established to analyze the effects of burn layer thickness, substrate thickness, thermophysical properties, excitation intensity, and modulation frequency on the photothermal amplitude and phase responses. Then, a modulated laser infrared thermography system is developed to acquire data from four circumferential directions of the bearing raceway. The infrared temperature sequences are processed by detrending using a fourth order polynomial and lock-in demodulation, and amplitude and phase maps are obtained at different modulation frequencies. The experimental results show that the burned region exhibits a stronger periodic thermal response than the sound region and can be distinguished in the time domain temperature curves, lock-in amplitude, and phase features. Frequency selection based on the contrast-to-noise ratio shows that, under the present experimental conditions, 0.2 Hz provides a good balance between amplitude contrast and phase stability. The comparison of different feature extraction methods indicates that the lock-in amplitude and FFT amplitude have good stability and clear physical interpretability, while phase, PCA, and PLSR features can serve as auxiliary criteria for burn identification. These results verify the feasibility of the modulated laser active infrared thermal wave method for detecting localized burn damage on bearing raceways.
Hyperspectral image (HSI) classification remains challenging in real-world scenarios due to severe class imbalance and ambiguous boundaries caused by mixed pixels. Existing methods typically address class imbalance at the data or loss level, while overlooking its intrinsic coupling with spectral–spatial representation learning. This limitation often leads to the degeneration of discriminative structures and the dominance of majority classes in feature space. To tackle this issue, this paper proposes a multi-scale spectral–spatial selective network (M3S-Net) that jointly enhances feature representation and enables distribution-aware feature learning and optimization. Specifically, a high-frequency residual spatial attention (HRSA) module is designed to learn a smooth background-response suppression mask inspired by high-pass residual modeling, thereby mitigating the inherent over-smoothing effect of convolutional operations and enhancing boundary-aware representations. Furthermore, a dynamic selective fusion module (DSFM) is developed to adaptively aggregate shallow spatial details and deep semantic features, effectively bridging the semantic gap across different scales. To address long-tail distribution and improve generalization, a distribution-balanced regularization strategy is proposed by integrating inverse class weighting and label smoothing, thereby jointly alleviating distribution bias and suppressing overfitting. In addition, a test-time augmentation (TTA) scheme is employed to enhance inference robustness without introducing additional training cost. Extensive experiments on four widely used benchmark datasets demonstrate that the proposed method achieves top or highly competitive performance under limited-label and imbalanced evaluation protocols. For instance, M3S-Net obtains 85.69 ± 1.49% overall accuracy using only 1% of the training samples on Indian Pines, exceeding CLSJE by 0.31 percentage points and maintaining balanced average accuracy.
High dynamic range (HDR) reconstruction from multi-exposure sequences is essential for imaging scenes with extreme illumination, especially for short-wave infrared (SWIR) imaging, where single exposures often fail to preserve details in both highlights and shadows. Existing deep learning HDR methods are primarily developed for visible-spectrum images and do not account for SWIR-specific characteristics, including single-channel sensing and near-linear radiometric response. This paper introduces a Swin Transformer-based framework with a Feature Enhancement Encoder and Feature Propagation blocks. We proposed the Enhanced Frame Attention module to improve multi-exposure feature aggregation and HDR reconstruction quality. To address limited SWIR training data, a two-stage training strategy is adopted, consisting of grayscale pre-training followed by SWIR domain-adaptive fine-tuning. To support training and evaluation on real SWIR data, we introduce the SWIR-HDR dataset, containing InGaAs-captured image sequences paired with physics-based ground truth. For fair evaluation, representative state-of-the-art visible-light HDR methods are retrained on the proposed dataset. Experimental results demonstrate that the proposed approach achieves superior reconstruction performance.
Outdoor object detection is essential for remote sensing and intelligent surveillance. Traditional single-modality sensors exhibit severe performance degradation under adverse weather and night-time conditions. Visible-infrared fusion methods significantly enhance object perception. However, heterogeneous imaging mechanisms induce inter-modal intensity inconsistencies and spatial misalignment, which degrade feature reliability and amplify modality imbalance during fusion. These factors jointly disrupt feature representation consistency and substantially impair small target detection, thereby making detection highly sensitive to background interference and localization deviations. To address these issues, CBF-Net, a YOLOv8-based visible-infrared small-target detection framework, has been developed that incorporates task-oriented fusion, contextual enhancement, and localization-stable regression. In the MIBF module, a learnable channel-wise bias is combined with spatial-channel attention to recalibrate feature-level modality responses during RGB-IR fusion. The MSE module uses parallel depthwise separable convolutions with multiple receptive fields to incorporate local and contextual cues for small target representation. In addition, a Consensus-NWD loss is formulated to jointly optimize detection by enforcing inter-modal distribution consistency and adopting normalized Wasserstein distance for stable regression of small targets. Experiment results on the VEDAI dataset show that CBF-Net improves mAP@50 and mAP@50:95 by 5.2 % and 5.0 % over the YOLOv8 baseline. Supplementary evaluation on LLVIP suggests the potential applicability of CBF-Net to another visible-infrared detection scenario, achieving gains of 1.8 % in mAP@50 and 4.7 % in mAP@50:95. The proposed method provides an effective solution for mitigating modality-related feature degradation and enhancing small target detection in the evaluated RGB-IR scenarios.
Breast cancer remains one of the most prevalent malignancies among women worldwide, and early detection is essential for improving survival and treatment outcomes. Although mammography, ultrasound, and magnetic resonance imaging are widely used clinically, their limitations have increased interest in complementary approaches. Depending on the modality, these include exposure to ionizing radiation, reduced sensitivity in dense breast tissue, operator dependence, and high cost. Infrared Thermal Imaging (ITI) has therefore emerged as a complementary modality because it is non-invasive, non-contact, radiation-free, and potentially cost-effective. Unlike anatomical imaging, ITI provides functional information by analyzing surface temperature distributions associated with physiological alterations in breast tissue.This review systematically examines ITI-based breast cancer research by organizing the literature into three complementary research environments—in silico, in vitro, and in vivo—thereby highlighting the distinct role of each stage in the development, validation, and clinical translation of ITI. In silico studies use mathematical models and computational simulations to investigate heat transfer in breast tissue and tumor-related temperature changes. In vitro studies use tissue-mimicking phantoms to test thermal models, imaging systems, and analysis methods under controlled laboratory conditions. In vivo studies represent the clinical application stage, where thermograms acquired from human subjects are analyzed using computer-aided methods for screening, diagnosis, and risk assessment. By integrating evidence across these environments, this review traces ITI from computational modeling through experimental validation to clinical application, clarifies its potential as a functional imaging technique that complements established breast imaging modalities, and provides a framework for future research.
