With innovative applications of unmanned aerial vehicle (UAV) ad hoc networks in various areas, their demands on broad bandwidth, large capacity and low latency become prominent. The combination of millimeter wave, directional antenna and time division multiple access techniques, which enables concurrent transmissions, is promising to deal with it. In this paper, we study the resource allocation problem in UAV ad hoc networks. Specifically, the slot assignment, antenna boresight and transmit power are jointly optimized to promote the network capacity. First, we formulate the optimization problem as the maximization of the fairness-weighted network capacity, subject to the constraint on priority guarantee. Then, because the formulated problem is a mixed integer non-linear programming problem (MINLP), which is NP-hard, two algorithms called dual-based iterative search algorithm (DISA) and sequential exhausted allocation algorithm (SEAA) are respectively proposed to efficiently solve it with acceptable complexity. DISA slacks the MINLP into a continuous-variable optimization problem and solves it with the Lagrangian dual method in an iterative manner. As a heuristic method, SEAA schedules links sequentially, i.e., from high-priority to low-priority ones. Numerical results demonstrate that both DISA and SEAA can efficiently allocate resources for UAVs, while guaranteeing the fairness and priority of links.
12 April 2023 in IEEE Transactions on Geoscience and Remote Sensing
In the process of downsampling, the detection of small infrared targets encounters problems such as imaging area, missing texture features, and disappearance of targets. This study proposes a local contrast attention guide network (LCAGNet), which implements and uses a SPD-CSP Resblock (i.e., CSPResblock based on sub-pixel DownSample), a MALC (i.e., multiple attention local contrast) module, and MCAF (i.e., multi-scale cooperative attention fusion) module, as well as e-maxmin (i.e., natural logarithm normalization) loss function for reconstructing images. First, by segmenting the gradient flow, the SPD-CSP Resblock propagates it in different paths. Second, under the guidance of the MALC module, the accuracy and convergence speed of the network are improved by constructing local nonlinear feature layers. Third, the weighted low-level detail feature channel is integrated with the deep semantic feature by MCAF to enhance the semantic information of the target. By standardizing the NL(i.e., natural logarithm) probability of the network output, the loss of target shape and position could be calculated accurately. Also, in the upsampling stage, the sub-pixel conv method reconstructs features with high resolution and further boosts detection accuracy. A series of detailed ablation experiments were conducted on each module in the network architecture to verify the effectiveness of the module and compare LCAGNet with other advanced data- and model-driven methods. The experimental results showed that LCAGNet achieved better results than other advanced benchmarking methods on two datasets of infrared small-dim target detection and tracking in both single-frame infrared small target benchmark (SIRST) and ground-sky background (KBT).
At present, the applications of multiple unmanned aerial vehicles (UAVs) are becoming more and more widespread, covering many civil and military fields. When performing tasks, UAVs will form a flying ad hoc network (FANET) to communicate to each other. However, subject to high mobility, dynamic topology, and limited energy of FANETs, maintaining stable communication performance is a challenging task. As a potential solution, the clustering routing algorithm divides the entire network into multiple clusters to achieve strong network performance. Meanwhile, the accurate localization of UAV is also strongly required when FANETs are applied in the indoor scenario. In this paper, we propose a firefly swarm intelligence based cooperative localization (FSICL) and automatic clustering (FSIAC) for FANETs. Firstly, we combine the firefly algorithm (FA) and Chan algorithm to better cooperative locate the UAVs. Secondly, we propose the fitness function consisting of link survival probability, node degree-difference, average distance, and residual energy, and take it as the light intensity of the firefly. Thirdly, the FA is put forward for cluster-head (CH) selection and cluster formation. Simulation results indicate that the proposed FSICL algorithm achieves the higher localization accuracy faster, and the FSIAC algorithm achieves the higher stability of clusters, longer link expiration time (LET), and longer node lifetime, all of which improve the communication performance for indoor FANETs.
Routing in flying ad hoc networks (FANETs) is hindered by high mobility, trajectory-induced topology dynamics, and energy constraints. Conventional topology-based or position-based protocols often fail due to stale link information and limited neighbor awareness. This paper proposes a trajectory-informed routing protocol enhanced by Q-learning: Traj-Q-GPSR, tailored for mission-oriented UAV swarm networks. By leveraging mission-planned flight trajectories, the protocol builds time-aware two-hop neighbor tables, enabling routing decisions based on both current connectivity and predicted link availability. This spatiotemporal information is integrated into a reinforcement learning framework that dynamically optimizes next-hop selection based on link stability, queue length, and node mobility patterns. To further enhance adaptability, the learning parameters are adjusted in real time according to network dynamics. Additionally, a delay-aware queuing model is introduced to forecast optimal transmission timing, thereby reducing buffering overhead and mitigating redundant retransmissions. Extensive ns-3 simulations across diverse mobility, density, and CBR connections demonstrate that the proposed protocol consistently outperforms GPSR, achieving up to 23% lower packet loss, over 80% reduction in average end-to-end delay, and improvements of up to 37% and 52% in throughput and routing efficiency, respectively.
Knowledge distillation has proven to be an effective model compression method that exploits the knowledge from a teacher model for supervising a student model by minimizing the distribution difference between the knowledge and the prediction produced by the student model. In this work, we focus on improving its performance from the perspective of divergence measures. A general form that represents a family of divergence measures is introduced and a novel learning paradigm that jointly optimizes multiple measures is proposed by formalizing it as a multi-objective learning problem. Conditioned on Pareto optimality, the weights of different divergences are tuned in an automated way during training. Extensive experiments on multiple datasets show the proposed method can significantly improve the performance of student network compared to the state-of-the-art methods for knowledge distillation.
1 August 2021 in Journal of Physics: Conference Series
Small object detection is a challenging research direction in the field of computer vision, due to the low resolution and restricted information of small objects. At present, the general detectors only use appearance features to classify and locate objects, but they are prone to failure under the interference of background noise. On the other hand, the detector based on deep neural network has excellent performance on large scale, but it is difficult to extract enough information of small objects. This paper proposes a feature enhancement network (FENet), which contains two modules. The Residual feature enhancement (RFE) module combines residual learning and sub-pixel convolution to improve the resolution of input small objects and remove image noise. The Attention Feature Pyramid (AFP) module integrates the feature pyramid and attention mechanism, which can extract context information and filter redundant context information. At the same time, considering the imbalance of the contribution of large and small objects to the loss function during the training process, a feedback-driven function is introduced to solve the problem of uneven loss under multiple scales. Experimental results show that compared with the existing small object detection methods, our method has better performance.