Salient object detection (SOD) is a fundamental component of unmanned aerial vehicle (UAV) vision systems, as it enhances autonomous perception and facilitates precise target localization. However, achieving an optimal balance between model efficiency and detection accuracy remains a significant challenge, particularly in resource-constrained scenarios such as UAVs. To address this issue, a lightweight SOD framework enhanced with edge information is proposed, aiming to achieve both high efficiency and high precision under limited computational resources. First, an edge-aware depthwise separable convolution module is introduced to replace conventional encoder operations, thereby enhancing spatial detail preservation while reducing computational cost. Second, a novel Edge Feature Enhancement Module is designed to extract fine-grained edge features from encoder outputs and to guide the decoder in reconstructing object with higher accuracy. Third, an edge-integrated loss function is developed by incorporating edge supervision into the standard cross-entropy loss, thereby further improving the model’s capability to delineate salient object boundaries. Extensive experiments conducted on multiple benchmark SOD datasets demonstrate that the proposed method achieves competitive, and in some cases superior, performance compared to existing state-of-the-art lightweight models, while maintaining low computational complexity. The proposed framework provides a promising solution for real-time and energy-efficient SOD in low-power computing environments.

Object detection plays a pivotal role in intelligent transportation systems. In recent years, the rapid advancements in deep learning have propelled object detection to a new level. However, the complexity of traffic environments and the inherent limitations of Convolutional Neural Networks (CNNs) in dealing with feature uncertainty have collectively resulted in a stagnation in the development of traffic object detection. To surmount these challenges, this study explores the combination of fuzzy theory with deep learning object detectors. Specifically, the features are first transformed into the fuzzy domain through fuzzification, and then the improved lightweight fuzzy inference method is employed for information interaction and selection, effectively managing and mitigating feature uncertainty. By merging the fuzzy strategy with YOLOv6 framework, a novel Fuzzy-based YOLO-style detector (F-YOLO) was proposed. F-YOLO is distinguished by three key innovations: the fuzzy fusion module, the fuzzy attention mechanism and the GPBblock. These components respectively bolster multi-scale feature fusion in the neck, refine prediction efficiency in the head, and enhance the global contextual learning of the backbone. Comprehensive experiments conducted across various datasets, including those tailored for traffic object detection, establish F-YOLO's state-of-the-art (SOTA) performance in YOLO-style detectors. F-YOLO demonstrates robustness and wide applicability, notably surpassing comparison models by 20.7% on VisDrone2019 dataset, 1.3% on BDD100K, and 6.2% on UA-DETRAC. Furthermore, F-YOLO also achieved competitive results on the MS COCO dataset.

As a key enabler of Industry 4.0, the Industrial Internet of Things (IIoT) has been rapidly advancing, driving the increasingly widespread adoption of artificial intelligence (AI) in industrial production. However, the high computational demands of AI tasks contrast sharply with the limited computing resources available in industrial environments, highlighting the need for efficient task offloading strategies. This paper addresses AI task offloading under resource-constrained IIoT scenarios by proposing LSTM-Enhanced Hierarchical-Classification Offloading with DQN (LHC-DQN). The proposed framework adopts a two-layer offloading architecture: the first layer performs global scheduling by assigning tasks to cloud, edge, or device resources, while the second layer refines the allocation. Cloud and edge nodes apply mathematical optimization, whereas device nodes leverage DRL agents for autonomous decision-making. To handle AI task heterogeneity, a classification-aware mechanism is introduced in the first layer, deploying separate DQN agents for inference and training tasks to improve adaptability and efficiency. Furthermore, an LSTM module is integrated into the DQN backbone to capture temporal dependencies in task states. Experimental results in a simulated environment demonstrate that LHC-DQN significantly outperforms traditional methods, increasing task completion rates from approximately 47% to 68%. Ablation and generalization tests further confirm the robustness and effectiveness of the proposed method. Overall, LHC-DQN offers a practical and efficient solution for intelligent task offloading and resource scheduling in IIoT environments.

Object detection in complex traffic scenarios is crucial for Intelligent Transportation Systems (ITS). At present, most real-time traffic object detection methods primarily rely on YOLO-style vision-only detectors, limiting their potential for further improvement. Vision-Language Object Detection (VLOD) has made promising progress currently, yet its adoption in the realm of ITS remains limited. Previous VLOD methods utilize text features in the classification task, without fully exploring their impact on the regression process for object localization. Besides, existing multi-modal fusion approaches fail to fuse text features with multi-scale image features at corresponding scales, which is detrimental to the representation capability of the model. In this work, we dive into the limitations above and introduce Zone-YOLO to improve the VLOD to a new level. Specifically, we propose Scale-Aware Modal Fusion (SAMF) to fully exploit the text and image features and learn to fuse the multi-modal representations seamlessly at different scales with channel- and modal-wise enhancement. Moreover, we present a novel Zone Prompt learning method to introduce text features into regression process and capture the zone-class-entity triple co-occurrence, which significantly improves the localization performance of the model. Extensive experiments show that Zone-YOLO outperforms the comparative methods by a considerable margin, achieving 55.1 AP, 72.1 AP50 and 71.2 APL on COCO. The competitive results on BDD100K and VisDrone2019 further demonstrate the superiority of Zone-YOLO on efficient traffic object detection.

Deep learning-based image stitching pipelines are typically divided into three cascading stages: registration, fusion, and rectangling. Each stage requires its own network training and is tightly coupled to the others, leading to error propagation and posing significant challenges to parameter tuning and system stability. This paper proposes the Simple and Robust Stitcher (SRStitcher), which revolutionizes the image stitching pipeline by simplifying the fusion and rectangling stages into a unified inpainting model, requiring no model training or fine-tuning. We reformulate the problem definitions of the fusion and rectangling stages and demonstrate that they can be effectively integrated into an inpainting task. Furthermore, we design the weighted masks to guide the reverse process in a pre-trained large-scale diffusion model, implementing this integrated inpainting task in a single inference. Through extensive experimentation, we verify the interpretability and generalization capabilities of this unified model, demonstrating that SRStitcher outperforms state-of-the-art methods in both performance and stability.

The swift expansion of the Industrial Internet of Things (IIoT) presents formidable challenges for both data transmission networks and processing units. In this paper, we delve into the issue of accelerating visual data transmission and processing within IIoT systems. Our approach aims to deploy visual saliency detection models in edge data collection devices to empower them with data pre-processing capabilities. To address the challenge of model lightweight deployment on edge devices, we have designed a lightweight saliency detection model. It utilizes our proposed Spatial Difference Attention-based Dynamic Fusion (SADF) module for adaptive detection of salient objects of varying scales. As our model outputs retain only task-relevant information from the images, it maximizes compression on the raw data, significantly reducing the volume of data subsequent tasks and transmission networks need to process. This enhances the speed of data processing. Finally, extensive experimentation across multiple datasets demonstrates that our proposed model achieves a balance between performance and efficiency. Ultimately, this paper presents an effective avenue for efficiently handling data within IIoT systems by combining edge computing nodes with intelligent models.

Emergency prediction and driver attention prediction are fundamental tasks within the realm of self-driving vehicles and assistant driving systems. The utilization of visual saliency detection in these tasks has garnered considerable attention, owing to its inherent advantages. However, current research on traffic saliency detection primarily focuses on emulating the human visual system for attention allocation in traffic scenes, neglecting the detection of complete salient objects. In this paper, we propose the Traffic Salient Object Detection Using a Feature Deep Interaction and Guidance Fusion Network (TFGNet). Different from previous methods, our method detects the complete objects that attract human attention in natural traffic scenes, rather than a certain point without object semantic information, which can provide assistance for target recognition tasks in the domain of intelligent driving. Moreover, we propose a traffic salient object detection framework based on feature interaction and guidance fusion, enabling the detection of salient objects across varying scales. Experimental results on multiple benchmark datasets demonstrate that, compared to the state-of-the-art methods, our method exhibits superior performance in terms of precision, recall, and error rate.