The YOLOv5 algorithm is widely used in edge computing systems for object detection. However, the limited computing resources of embedded devices and the large model size of existing deep learning-based methods increase the difffculty of real-time object detection on edge devices. To address this issue, we propose a smaller, less computationally intensive and more accurate algorithm for object detection. Multi-scale feature fusion-YOLO (MFF-YOLO) is built on top of the YOLOv5s framework, but it contains substantial improvements to YOLOv5s. First, we design the MFF module to improve the feature propagation path in the feature pyramid, which further integrates the semantic information from different paths of feature layers. Then, a large convolutionkernel module is used in the bottleneck. The structure enlarges the receptive ffeld and preserves shallow semantic information, which overcomes the performance limitation arising from uneven propagation in feature pyramid networks (FPN). In addition, a multi-branch downsampling method based on depthwise separable convolutions and a bottleneck structure with deformable convolutions are designed to reduce the complexity of the backbone network and minimize the real-time performance loss caused by the increased model complexity. The experimental results on the PASCAL VOC and MS COCO datasets show that, compared with YOLOv5s, MFF-YOLO reduces the number of parameters by 7% and the number of FLOPs by 11.8%. The mAP@0.5 has improved by 3.7% and 5.5%, and the mAP@0.5:0.95 has improved by 6.5% and 6.2%. Furthermore, compared with YOLOv7-tiny, PP-YOLO-tiny, and other mainstream methods, MFF-YOLO has achieved better results on multiple indicators.
- Article type
- Year
- Co-author


Edge computing nodes undertake an increasing number of tasks with the rise of business density. Therefore, how to efficiently allocate large-scale and dynamic workloads to edge computing resources has become a critical challenge. This study proposes an edge task scheduling approach based on an improved Double Deep Q Network (DQN), which is adopted to separate the calculations of target Q values and the selection of the action in two networks. A new reward function is designed, and a control unit is added to the experience replay unit of the agent. The management of experience data are also modified to fully utilize its value and improve learning efficiency. Reinforcement learning agents usually learn from an ignorant state, which is inefficient. As such, this study proposes a novel particle swarm optimization algorithm with an improved fitness function, which can generate optimal solutions for task scheduling. These optimized solutions are provided for the agent to pre-train network parameters to obtain a better cognition level. The proposed algorithm is compared with six other methods in simulation experiments. Results show that the proposed algorithm outperforms other benchmark methods regarding makespan.

Service recommendation provides an effective solution to extract valuable information from the huge and ever-increasing volume of big data generated by the large cardinality of user devices. However, the distributed and rich multi-source big data resources raise challenges to the centralized cloud-based data storage and value mining approaches in terms of economic cost and effective service recommendation methods. In view of these challenges, we propose a deep neural collaborative filtering based service recommendation method with multi-source data (i.e., NCF-MS) in this paper, which adopts the cloud-edge collaboration computing paradigm to build recommendation model. More specifically, the Stacked Denoising Auto Encoder (SDAE) module is adopted to extract user/service features from auxiliary user profiles and service attributes. The Multiple Layer Perceptron (MLP) module is adopted to integrate the auxiliary user/service features to train the recommendation model. Finally, we evaluate the effectiveness of the NCF-MS method on three public datasets. The experimental results show that our proposed method achieves better performance than existing methods.