Lina Wu

High-quality light field (LF) acquisition involves a trade-off between spatial and angular resolution. Hybrid camera systems offer a practical solution but present challenges for dense reconstruction due to sparse angular sampling and complex occlusions. To address these issues, this paper proposes a geometry-aware implicit neural representation (INR) framework for LF reconstruction. Distinct from traditional discrete representations or generic ray-space embeddings, we introduce a compact, continuous representation that combines a radiance field with a geometry module, where scalar disparity provides an ideal geometric interpretation and a ray-displacement field serves as the practical realization for occlusion-aware reconstruction. This framework leverages the global epipolar geometry of the 2D LF grid while utilizing the displacement field to correct local geometric inconsistencies caused by occlusions and non-Lambertian effects. By coupling these fields with a radiance network, our method enables end-to-end differentiable optimization from sparse, multi-resolution inputs without relying on large-scale external training datasets. Experiments on hybrid camera data fusion and spatial-angular super-resolution tasks demonstrate that our approach preserves high-frequency details and geometric consistency.

Path loss (PL) is a key channel characteristic that provides theoretical guidance for wireless network planning and optimization. Traditional channel models face an inherent trade-off: empirical models are limited in accuracy due to the omission of environmental factors, while deterministic models suffer from low efficiency due to their reliance on manual environmental modeling. Although deep learning (DL) has gained attention in PL prediction, existing DL-based models still lack expressive environmental representations, limiting improvements in prediction results. To address this limitation, an electromagnetic-geometric (E-G) coupled environmental representation is proposed, which describes the environment by integrating the types of land covers with their elevation information. An intelligent PL model is developed based on E-G environmental representation. Experimental results show that, without cumbersome environmental modeling, E-G features significantly improve prediction accuracy, highlighting the potential of the proposed model to overcome the accuracy-efficiency trade-off in traditional PL models.

Reconfigurable intelligent surface (RIS) is a promising technology to enhance the performance of mobile edge computing (MEC) network. Nevertheless, the RIS-enabled MEC network design is a non-trivial problem. This paper investigates the RIS-enabled MEC network, where the Internet of Things devices (IoTDs) with limited energy budgets can offload their partial computation tasks to the base station (BS). We first formulate a sum computation rate maximization problem by jointly designing the RIS phase shifts, and the IoTDs' energy partition strategies for local computing and offloading. Then, to handle the non-convex optimization problem, we propose a deep reinforcement learning (DRL)-based algorithm, in which the twin delayed deep deterministic policy gradient (TD3) algorithm is adopted to optimize the RIS phase shifts and the IoTDs' energy partition strategies. Simulation results show that the proposed TD3 solution can reach a better sum computation rate than the benchmark algorithms.

Cell-free multiple-input multiple-output (MIMO) and reconfigurable intelligent surface (RIS) have been envisioned as two promising techniques to enhance the data transmission rate of high-speed railway (HSR) networks. This paper considers the HSR cell-free MIMO system empowered by RIS with finite discrete phase shifters to pursue performance improvement. Particularly, the RIS phase shift optimization problem is formulated, aiming at maximizing the achievable rate. To deal with the complicated control problem, a deep reinforcement learning (DRL)-based scheme is proposed, where double deep Q-network (DDQN) method is invoked for designing phase shifts. Simulation results demonstrate that compared with the existing optimization-based baseline scheme, the proposed scheme can obtain the comparable achievable rate with much shorter time consumption.

Satellite image based land cover classification, which falls under the category of semantic segmentation, is critical for many global and environmental applications. Deep learning has been proven to be excellent in semantic segmentation. However, mainstream neural networks formed by connecting high-to-low convolutions in series are prone to losing image information, which affects the accuracy of semantic segmentation. Besides, it is difficult to distinguish adjacent land cover classes with similar colors using only RGB information presented by satellite images. Striven to maintain high-resolution representations and improve the inter-class distinguishability, a Canny enhanced high-resolution neural network (C-HRNet) is proposed based on Canny edge features and the high-resolution neural network (HRNet) that maintains high-resolution representations throughout the process. Meanwhile, we construct a novel dataset for model evaluation, which provides an automatic dataset construction method, making the dataset more efficient in construction and richer in samples. Extensive experiments are conducted on datasets at different granularities. Quantitative results demonstrate that for large-scale fine-grained scenarios, C-HRNet outperforms state-of-the-art semantic segmentation networks due to the accurate spatial localization ability of Canny edge features. For small-scale coarse-grained scenarios, Canny extracts a large number of edge features that highlight the positional differences between adjacent instances belonging to the same land cover class, which slightly degrades the performance of C-HRNet, but it can still provide reliable land cover classification results. Based on this conclusion, we apply C-HRNet to large-scale wireless channel simulations that are location-sensitive and require fine-grained semantic segmentation, which are proven to be accurate and effective.

In this paper, we apply reconfigurable intelligent surface (RIS) technique to aid the computation offloading of mobile edge computing (MEC) network and investigate how it can be exploited to reduce computation offloading delay. In order to minimize the long-term computation offloading delay, we formulate an optimization problem, which jointly optimizes the power control, computation offloading volume, the edge computing resource assigned to each user equipment (UE), as well as the RIS phase shift. To tackle this problem, we first convert it into a Markov decision process (MDP), then propose an efficient algorithm based on deep reinforcement learning (DRL), namely deep deterministic policy gradient (DDPG). Numerical results demonstrate that 1) compared to the MEC network without RIS, the RIS aid MEC network can achieve lower delay; 2) the proposed DDPG-based scheme can learn from the RIS aided environment to effectively reduce the computation offloading delay.

Accurate path loss (PL) models are critical for the design and optimization of wireless communication systems. The selection between empirical models and deterministic models are usually a tradeoff between accuracy and efficiency. In this letter, a novel environmental characterization method is defined by considering clutter information in the vector-based and image-based structures. Three PL models are proposed with different artificial neural network (ANN) architectures and two environmental characterizations. Comparative analysis shows that an image-based environmental characterization, which can reduce the influence of interference types, is better than a vector-based characterization. The combination of an image-based environmental characterization method and ANNs can enhance PL prediction in both accuracy and efficiency.

Accurate path loss (PL) prediction is essential for predicting transmitter coverage and optimizing wireless network performance. Traditional PL models are difficult to cope with the development trend of diversity, time-varying and mass wireless channels. In this work, the most widely used multilayer perceptron (MLP) neural network in artificial neural network (ANN) is employed to accurately predict PL. Three types of environmental features are defined and extracted, which describe the propagation environment only by considering limited environmental types instead of complex 3D environment modeling. Principal component analysis (PCA) is used to generate the low-dimensional environmental features, and eliminate redundant information among similar environmental types. Moreover, the information of base station (BS) and the receiver (Rx), including 3D locations, frequency, the transmitted power of BS, the antenna information, the feeder loss, and the received power of all the locations are obtained from the measurements. Different environmental features are combined with the information of BS and Rx to construct seven datasets for PL prediction models based on MLP neural networks. The impacts of the number of neurons in the hidden layer, the number of hidden layers, the number of training samples, and environmental features on PL prediction models are explored by considering the absolute value of mean error (AME), the mean absolute error (MAE), the standard deviation (STD) of error, the correlation coefficient, and the time ratio, respectively. This work aims to understand the propagation characteristics of radio waves, which can provide a theoretical basis for wireless network optimization and communication system design.