Linqing Gui

Hyperspectral image (HSI) restoration is a critical preprocessing step to improve data quality for downstream applications. Existing model-driven methods depend on hand-crafted priors that are often inadequate to capture the highly coupled spatial–spectral characteristics of HSIs, while deep learning-based approaches frequently exhibit limited robustness and generalization beyond the training distribution. To tackle these challenges, we propose a unified restoration framework termed Robust Deep Low-Rank Tensor Factorization (RDLTF) for HSI restoration. First, to mitigate the computational burden and exploit the global spectral correlation, we introduce a robust deep low-rank subspace representation strategy, where the unsupervised deep network generates reduced low-dimensional components that are subsequently projected onto the original space via a reliable spectral basis. Second, a flexible truncated-Cauchy loss function to enhance robustness. Furthermore, to improve restoration accuracy, an explicit spatial-spectral total generalized variation (SSTGV) regularizer is integrated into the restoration model, which simultaneously preserves piecewise smoothness and suppresses spectral distortions. An efficient ADMM algorithm is developed to solve the resulting non-convex optimization problem. We evaluate the proposed method on several datasets for image denoising, inpainting, and super-resolution. The results demonstrate that the proposed method outperforms existing unsupervised and supervised techniques in both quantitative metrics and visual quality.

Sedentary lifestyle has become a major health risk in modern society. Long sitting time can be detected by accurate recognition of sitting and standing (sit-stand) activities. WiFi-based sitting time detection has the remarkable advantage of low cost, noncontact, and privacy-protection. However, accurate recognition of sit-stand activities via WiFi signal is still facing two challenges. The first and also tougher challenge is inevitable mistakes in recognition results of traditional machine learning methods, while the second challenge is the difficulty of accurate activity segmentation before activity recognition. To the best of the authors' knowledge, few work addresses the above challenges, particularly the first challenge. A new contactless sitting time detection system is designed accordingly. The system first accurately segments all activities and removes in-seat activities. Then, the mistakes in sit-stand activity recognition are effectively corrected by a new recognition error correction method. The proposed method first creates and updates a correction benchmark that can satisfy both successive correlation and waveform symmetry between sit-stand activities. The recognition results of traditional machine learning methods are then corrected based on the latest correction benchmark. Extensive experiment results demonstrate that compared to related work, the designed system has much better accuracy on both sit-stand activity recognition and sitting time estimation. The experiment results also demonstrate the robustness of the designed system.

As a critical preprocessing step, hyperspectral image (HSI) denoising aims to improve the HSI quality for subsequent applications. While unsupervised HSI denoising methods based on Deep Image Prior (DIP) have garnered attention due to their pre-training-free advantage, existing DIP-based approaches typically utilize L₂-norm as data fidelity, making them inefficient in handling complex mixed noise. Moreover, such unsupervised methods only focus on spatial domain priors, lacking a comprehensive characterization of the spatial-spectral correlations inherent in HSIs. To tackle these limitations, we propose a robust deep recovery (RDR) model for HSI denoising with spatial-spectral total generalized variation (SSTGV) prior. Specifically, the truncated-Cauchy loss function is adopted to suppress the interference of outliers and enhance the robustness against sparse noise. Moreover, the SSTGV prior is integrated into the unsupervised RDR model, resulting in complementary effect of deep prior and handcraft prior. To solve the resulting optimization problem, an efficient ADMM algorithm is developed with convergence guarantee. Experimental results demonstrate the significant advantages of our approach in both noise suppression and detail preservation, highlighting its robustness and adaptability for varied HSI denoising applications.

Leveraging ubiquitous WiFi infrastructures, intrusion detection methods based on Received Signal Strength (RSS) offer compelling advantages, including cost-effectiveness and privacy protection. However, existing RSS-based intrusion detection solutions fall short of accurately estimating and extending the WiFi sensing bound. In this paper, we propose a novel model of motion-disturbed RSS and design an effective R-ratio indicator to extend the intrusion detection bound. Specifically, we first establish a general model of motion-disturbed RSS and derive the blocked area and reflection area in this RSS model. Then, we define the WiFi intrusion detection bound and propose a performance indicator called R-ratio to extend the bound with RSS. Furthermore, based on the statistical properties of noise, we design an efficient filter to further weaken the noise. We also propose two new methods to further extend intrusion detection bound. Extensive experimental results demonstrate that the proposed power sum ratio based intrusion detection method can approximately double the WiFi intrusion detection bound compared to other methods with raw RSS data, and our developed motion-disturbed RSS model can provide valuable insights and guidance to the intrusion detection system.

Integrated Sensing and Communications (ISAC) integrates sensing and communication functions through ubiquitous wireless signals, providing a seamless and flexible interaction experience, making it an ideal choice for intelligent Human-Computer Interaction (HCI). Among various interaction methods, gesture recognition has garnered widespread attention. However, current RF-based gesture recognition methods within ISAC are constrained by single-target sensing and insufficient robustness. In this paper, we propose MultiGes, a real-time multi-user gesture recognition system designed to support ISAC-driven scenarios. MultiGes employs dual commercial Impulse Radio Ultra-Wideband (IR-UWB) devices to create multiple wireless links, capturing dynamic motion features from multiple targets. First, a human energy map is constructed based on the reflected signal energy to determine multi-target coordinates. Then, a Differential Human Relative Velocity (DHRV) matrix is extracted to capture fine-grained motion information. Finally, we design a lightweight STNet model to extract spatiotemporal gesture features from the DHRV matrix, enabling real-time multi-target gesture recognition. We implement the MultiGes system prototype and conduct extensive experiments on ten common gestures in HCI scenarios. Experimental results demonstrate that MultiGes achieves efficient recognition for 2 to 5 users, with an average accuracy of over 90%, providing a robust, scalable, and real-time solution for multi-target gesture recognition in ISAC-driven smart environments.

Gesture recognition provides an effective human-computer interaction that makes device control more intuitive and convenient. Although the research on mmWave radar-based gesture recognition has demonstrated promising results, existing studies have exclusively addressed the cross-domain challenge or the human activity interference problem, and no attention has been paid to the cross-domain problem in the presence of human activity interference. To address these issues, we propose a novel mmWave radar-based gesture recognition system, named GestSAM, which leverages a parameter-free attention mechanism to effectively extract gesture features that are less affected by environmental noise. By integrating this mechanism with deep learning techniques, GestSAM significantly reduces the impact of human activity interference while maintaining robust cross-domain gesture recognition performance. This approach ensures robust, high-accuracy recognition of gestures. In order to evaluate the performance of our system, we construct a dataset containing six different gesture types performed by fifteen volunteers in seven different scenarios and simulate three interference conditions. The experimental results show that under human activity interference, the model achieves average recognition accuracies of 92.79% and 94.62% in cross-user and cross-scenario, respectively.