Abstract:Accurate 3D object detection and localization are critical for UAV-based inspection and obstacle avoidance. Traditional methods often integrate detection and localization within a unified network, resulting in complex architectures, high computational costs, and challenges in real-time deployment. To address these issues, we propose a lightweight 3D object detection and localization framework based on network decoupling. First, a lightweight 2D object detection network is designed, incorporating efficient feature extraction and an enhanced attention mechanism, which significantly reduces the number of parameters while improving generalization across diverse target types. Second, we introduce a visual/LiDAR fusion-based depth completion network with cross-layer connections and auxiliary loss functions to achieve high-precision dense depth map estimation. Finally, a pixel/depth alignment scheme is developed to accurately compute 3D spatial positions of detected objects via coordinate transformation. Experimental results demonstrate that, compared to the YOLOv9 detection algorithm, the proposed method improves object detection accuracy by 14%, and enhances 3D localization accuracy by 45% over the AVOD framework. Moreover, the proposed approach achieves a processing rate of 36 frames per second on UAV edge devices, representing a 90% increase over AVOD, highlighting its practical value for real-time UAV-based object detection applications.

Abstract:To address the issues of slow convergence, high parameter sensitivity, and low computational efficiency in robot path planning within dynamic unknown environments, a novel algorithm named SSA-Qlearning was proposed by integrating the Sparrow Search Algorithm (SSA) with Quality-learning(Q-Learning). The method optimized the learning rate and decay factor of Q-Learning by introducing the collaborative mechanism among discoverers, followers, and scouts in SSA, and designed a dynamic weight adjustment strategy to adaptively explore the parameter space, thus eliminating the bias in phase-based optimization of traditional Q-Learning. The algorithm quantifies environmental dynamics by introducing a dynamic environmental factor to achieve a dynamic balance between exploration and safety, maintained the lightweight characteristics of Q-Learning, and avoided the high computational cost of Double Deep Q-Network (DDQN). The experimental results indicate that SSA-Qlearning significantly improves the path success rate in 5×5, 10×10, and 15×15 dynamic grid environments, with training times being only 8.07%, 3.4%, and 3.03% of DDQN, respectively, achieving a lightweight reinforcement learning effect close to the performance of DDQN.

Abstract:In the field of traffic sign detection, challenges arise due to the small area coverage of distant traffic signs in the scene and the diverse scales of the signs. To overcome the above challenges, this paper presents an improved YOLOv8s-based traffic sign detection algorithm, YOLOv8s-REMN. First, the method introduces the RFAConv into the backbone network to enhance the receptive field and feature extraction capability of the network. Second, the EAGFM module is added to the neck network to optimize multi-scale feature fusion. Then, the MSDEF module is incorporated into the detection head to increase the small object detection head, improving the detection of small targets. Finally, the NWD loss function replaces the CIOU loss function to optimize the bounding box regression and improve the precision of small object localization. Experimental results show that YOLOv8s-REMN achieves significant performance improvements on the TT100K dataset. Compared to the original YOLOv8s, mAP@0.5 increases by 6.6%, mAP@0.5:0.95 increases by 5.1%. The effectiveness of the algorithm is also validated on the Chinese Traffic Sign Detection dataset, CCTSDB2021, where YOLOv8s-REMN outperforms YOLOv8s with a 2.9% increase in mAP@0.5, a 2.9% increase in mAP@0.5:0.95.

Abstract:To address issues in sand-dust images, such as color shift, low clarity, and he poor performance of the dark channel prior method in handling sky regions, a sand-dust image enhancement method based on color cast correction and sky segmentation is proposed. First, the color cast in sand-dust images is corrected using a combination of channel compensation and the gray-world algorithm. Second, a dehazing method based on sky segmentation is proposed. The segmentation threshold is determined using information entropy, which separates the image into sky and non-sky regions. The dark channel is optimized using a fusion window. Then, an adaptive adjustment factor is introduced to refine the transmission map, and the atmospheric scattering model is employed to restore the image. Finally, in the HSV color space, an adaptive saturation enhancement algorithm and adaptive gamma correction are applied to adjust the image's saturation and brightness. Experimental results show that the proposed method can correct the color cast in sandstorm images, enhance image clarity, and improve restoration performance in sky regions. The method achieves improvements of 2.27%, 4.34%, and 0.25% in terms of average gradient, standard deviation, and information entropy, respectively.

Abstract:To capture the complex spatiotemporal dependencies in pedestrian trajectories, this paper proposes a trajectory prediction model that combines a bidirectional temporal learning module and a spatiotemporal interaction learning module. The model leverages bidirectional temporal feature modeling and self-supervised learning to extract spatiotemporal interaction features. In the bidirectional temporal learning module, a bidirectional temporal convolutional network is utilized to simultaneously model both historical and future trajectory information, enabling the capture of dynamic trajectory changes. In the spatiotemporal interaction learning module, the Test-Time Training (TTT) layer is employed with a self-supervised learning mechanism to dynamically adjust feature representations during the inference stage, thereby modeling spatiotemporal correlations. Finally, an adaptive fusion strategy is used to combine the features extracted by the two modules, focusing on key features while suppressing irrelevant information. Experimental results demonstrate that the proposed model achieves competitive prediction performance on the ETH and UCY datasets.

Abstract:The conventional Cycle Generative Adversarial Network (Cycle Generative Adversarial Network,CycleGAN), as an image style transfer model that does not require paired datasets, in the task of low-light domain image enhancement, there are problems such as the loss of details in the generated images, color distortion, and poor adaptability in complex scenarios. Based on this, this paper proposes an improved CycleGAN model which aims to enhance the effect of low - light image enhancement.First, in the design of the generator, a two - stage color correction module is integrated during the upsampling phase to alleviate the problems of color distortion and quality degradation in low - light environments. Second, the channel - spatial hybrid attention mechanism is embedded in the skip connection layer to achieve the adaptive strengthening of key information during the feature fusion process. Then, in the design of the discriminator, a global - local discrimination mode is adopted, enabling it to take into account the discrimination ability of both global information and local details. Finally, in the design of the loss function, perceptual style loss and content loss are introduced on the basis of adversarial loss to further improve the structural fidelity and visual naturalness of the generated images.Through the subjective and objective experimental evaluations and comparison with various representative image enhancement models, the experimental results show that this model has a good enhancement effect on low - light images. It can effectively enhance the overall brightness and local details of the images without causing color distortion, thus it improves the quality of the generated images.

Abstract:To address the issues of spatial distribution imbalance and low utilization rates in electric taxi charging facility siting, this paper proposes a multi-objective particle swarm optimization algorithm (FMPPSO) integrating epsilon constraint and fuzzy mathematical programming. By constructing a multi-constraint siting model incorporating land costs, passenger pickup rates, and battery degradation, we design an adaptive objective weight allocation strategy based on fuzzy membership functions to resolve the premature convergence challenge of traditional evolutionary algorithms in multi-objective optimization. An epsilon constraint mechanism introduced dynamically balances convergence and solution set diversity, generating high-quality Pareto frontier solution sets.Finally, simulation experiments and comparative analyses are conducted to validate the effectiveness of FMPPSO in solving the electric taxi charging station placement problem.

Abstract:Whole Slide Images(WSIs) are generally considered the golden standard for pathological diagnosis. Accurate classification of WSIs provides detailed information on tumor type, grade, and stage, which is crucial for cancer prognosis and treatment strategy selection. Currently, in the field of computational pathology, analysis methods based on Multi-Instance Learning (MIL) are becoming the mainstream approach for the classification of Whole-Slide Pathological Images. However, these methods mostly focus on single-scale pathological images, which limits the understanding of the mechanisms of cancer development and progression at different levels. Additionally, the high-resolution nature of pathological images and the information discrepancies across different scales pose challenges for efficiently integrating and analyzing pathological image patches within a single scale as well as across multiple scales. To address these issues, this paper proposes a whole-slide pathology image classification method based on deformable attention and multi-scale multi-instance learning (DMSMIL). Specifically, this method enhances the efficiency of attention computation by designing a deformable attention branch to learn the associations among image patches within the same scale. Meanwhile, an association algorithm based on Optimal Transport (OT) is designed to integrate pathological images across different scales, achieving efficient alignment of multi-scale pathological information. Experimental results on breast cancer subtype classification and lung cancer subtype classification tasks show that the proposed method achieved classification accuracy of 85.39% and 92.00% respectively, showing improved performance compared to mainstream whole slide image classification methods.

Abstract:In recent years, advances in computer vision have made it possible to reconstruct images based on EEG information, which is of great significance in fields such as medical image reconstruction and brain-computer interfaces. However, due to the complexity and temporal characteristics of EEG signals, existing models face many challenges in feature extraction and image generation tasks. To this end, this paper proposes an EEG signal-driven visual image reconstruction model based on double residual LSTM and DCGAN. The model introduces a long and short-term memory network based on an attentional residual network and a Triplet loss function to enhance the quality of EEG signal feature extraction. ARTLNet integrates residual network, long and short-term memory network and attention mechanism, which improves deep network training through residual con-nection, long and short-term memory network captures time-series features, and attention mecha-nism enhances the focus on key features; it also combines batch normalization and global average pooling to ensure stable signal delivery. In the image generation stage, the model introduces self-designed Deep Convolution Generative Adversarial Networks (DCGAN) with feature fusion strategy, which effectively improves the quality and diversity of the generated images. Experi-mental results show that the improved ARTLNet achieves higher accuracy on both the Characters and Objects datasets with different classification and clustering algorithms, and the proposed model also performs better in terms of image generation quality, especially in terms of image clarity and category differentiation.

Abstract:With the development of 5G and 6G communication, satellite communication, which can provide the global seamless coverage capability, has become an indispensable and important role. In view of the open channel characteristics of satellite communication, the security of satellite communication has been paid more and more attention, especially the demand for satellite secure communication in national defense and military communication. Based on the architecture of the Multi-Satellite and Terminal Multi-Antenna (MS-TMA) satellite communication system, from the physical layer of security and covert communication theory, it puts forward a satellite covert communication strategy, and the covert communication performance theory analysis and simulation, the research outcomes have significant reference value for the advancement of satellite secure communication theory and the development of its applications.