Abstract：In recent years, airborne infrared small target detection technology has become a research hotspot in the military and civilian fields. However, in practical applications, the influence of factors such as complex background and low signal-to-noise ratio still make infrared small target detection a challenge. Therefore, this paper proposes an improved airborne infrared small target detection algorithm AIR-YOLOv7, and uses the example transfer learning method to analyze the characteristics of small infrared targets, expand the data set, and further improve the performance of the algorithm. The experimental results show that the AIR-YOLOv7 algorithm has a better performance in infrared small target detection in airborne complex scenes, with a mAP value of 97.09% and an FPS of 102.09. With only a small amount of expansion of the data set in this paper, the instance migration method increases the mAP value of the algorithm by 0.96 percentage points, which provides a theoretical basis for the subsequent hardware platform edge computing transplantation.

Abstract：To address the problems of low target recognition accuracy and poor real-time performance of unmanned aerial vehicles (UAVs) in complex environments, this paper proposes a UAV target recognition algorithm based on adaptive neural networks and multi-scale feature fusion. The algorithm employs an improved convolutional neural network to extract multi-level features, combines attention mechanisms to adaptively adjust feature weights, and enhances target representation capability through multi-scale feature fusion. Experiments on the DroneCrowd dataset demonstrate that compared to algorithms such as ResNet-50, YOLOv11 and EfficientNet, the proposed method achieves an average recognition accuracy of 94.3%, representing an improvement of 8.7 percentage points over ResNet-50 and 2.0 percentage points over YOLOv11; an F1 score of 93.9%; and a processing speed of 61.0 FPS, representing a 35% improvement over ResNet-50. The method exhibits excellent robustness, providing an effective solution for UAV target recognition.

Abstract：Cellular-connected unmanned aerial vehicles (UAVs) show great potential in 5G networks. To address the challenge of maintaining stable connections with ground base stations during communication tasks, this paper investigates an UAV trajectory optimization problem. The objective is to jointly minimize the task completion time and communication outage duration while maximizing communication throughput as the UAV travels from the starting point to the destination within a given area. Considering the non-convex nature of the problem, a multi-step learning-based Dueling Double Deep Q Network (D3QN) algorithm is adopted to achieve adaptive trajectory optimization through interactive learning between the UAV and ground base stations. Simulation results show that, compared with the direct flight strategy, this method reduces the task completion time by 28%, cuts down the communication outage duration by 42%, and increases the average system throughput by 35%, achieving significant improvements in task efficiency, communication stability and system throughput.

Abstract：In response to the scheduling complexities introduced by the integration of distributed generation and the limitations of existing single-dimensional evaluation methods, this study constructs a comprehensive evaluation index system for Artificial Intelligence (AI)-enabled distributed power scheduling. The system encompasses three dimensions: economy, environmental friendliness, and stability. The research employs a combination of the Analytic Hierarchy Process (AHP) and the entropy method, based on cluster analysis, for comprehensive weighting. Additionally, a multi-task deep learning algorithm incorporating an attention mechanism is designed to synchronously predict multi-dimensional indicators. Simulation results demonstrate that, compared to traditional scheduling schemes, the AI-enabled scheduling approach based on this evaluation system achieves a better overall balance across key indicators such as total cost, clean energy penetration rate, and power supply reliability. This provides effective theoretical and methodological support for the coordinated optimization of economy, environmental protection, and stability in power systems.

Abstract：To address the conflict between the limited edge computing capacity of Unmanned Aerial Vehicles （UAVs） and the demand for high-precision multi-modal 3D detection， this paper proposes a lightweight detection network tailored for airborne platforms， termed Lite-VAFNet， building upon VAF-Net. A grid dimensionality-reduction detection head is constructed to compress the parameter volume by approximately 63%， thereby alleviating device storage constraints. A Linear-Bottleneck fusion module is designed to execute feature interaction with linear complexity， effectively eliminating the peak memory bottleneck. Furthermore， a collaborative acceleration framework integrating spatial resampling and logit distillation is introduced to overcome memory access limitations while compensating for the accuracy degradation induced by quantization. Experiments on the KITTI benchmark demonstrate that Lite-VAFNet achieves a 3D AP （Mod.） of 85.24% with merely 14.60 M parameters， significantly outperforming state-of-the-art models such as BEVFusion. This research strikes an optimal balance between accuracy and efficiency while substantially reducing resource consumption， exhibiting exceptional potential for edge deployment.

Abstract：In vessel detection from low-altitude UAV perspectives in inland rivers, traditional algorithms struggle to accurately detect small vessels due to issues such as small target size, vessel occlusion, complex backgrounds, light reflection, and wave disturbances. To address these problems, this study proposes an improved algorithm based on YOLOv11n—YOLO11-FFW (YOLO11—FEM FFM_Concat WIoUv2). To enhance the feature extraction ability for small vessel targets, the Feature Enhancement Module (FEM) is introduced, which expands the receptive field through multi-branch atrous convolution and integrates multi-scale contextual information. To improve multi-scale feature expression in complex backgrounds, the Feature Fusion Module Concat (FFM_Concat) is introduced, incorporating a learnable weight recalibration mechanism on top of the BiFPN structure, achieving adaptive fusion of high- and low-level features. To increase the model's robustness in scenarios with water surface reflection, occlusion, and dense targets, the loss function is improved to WIoUv2, dynamically balancing localization and classification losses. Experimental results show that compared to YOLOv11, YOLO11-FFW achieves a 1.4% increase in mAP@0.5, a 0.8% increase in precision, and a 2.4% increase in recall, which is verified to be effective in detecting small vessels in inland river scenarios from complex UAV perspectives.