Search

Wildfires in transmission line passages are a severe threat to power security. Two wildfire detection models, which are based on the YOLOv5, are proposed in this paper. Due to the limited computing power of embedded terminals, the proposed models simplify the network structure of YOLOv5. Specifically, one is that only the overall structure of the neck and head parts in the original network structure is simplified, and the second method is based on the first method to delete the modules of backbone, which greatly reduces the number of parameters of the model. The experimental results show that the proposed lightweight models can achieve real-time monitoring in embedded devices while the accuracy and recall remain high.

Because the existing finger vein segmentation networks are too large and not suitable for implementation in mobile terminals, the reduction of the parameters of the lightweight network leads to the reduction of the segmentation index, and the long-running time of deep network on hardware platforms; this paper proposes a lightweight real-time segmentation method for finger veins based on embedded terminal technique. In the preprocessing stage of the algorithm, the data is greatly expanded by randomly selecting the center to obtain sub-blocks on each image of the training set. The network first uses deep separable convolution to greatly reduce the U-Net parameters of a basic network and introduces an attention module to reorder the features to improve network performance, followed by a preliminary lightweight network Dinty-NetV1. Second, the Ghost module is added to the deep separable convolution, and the feature map of the network part is obtained through a cheap operation so that the network is further compressed to obtain Dinty-NetV2. After adding channel shuffle, all the characteristic channels are evenly shuffled and reorganized to obtain Dinty-NetV3. Finally, a study of the filter norm yields the distribution characteristics of the finger vein picture features. By using the geometric median pruning method, the network models for each stage of the algorithm proposed in this paper achieved better segmentation performance and shorter split time after pruning. The overall Dinty-NetV3 model size is only less than 9% of the U-Net and Mult-Adds is less than 2% of the U-Net with the same structure. After testing on two-finger vein datasets SDU-FV and MMCUBV-6000, we confirm that the performance of Dinty-NetV3 surpasses all previously proposed classic compression model algorithms and it is not inferior to more complex and huge networks such as U-Net, DU-Net, and R2U-Net. The proposed algorithm has advantages in terms of time needed to train the network, and we verify its universality using NVIDIA's full range of embedded terminals.

Copyright of Journal of Signal Processing is the property of Journal of Signal Processing and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Crowd counting plays an important role in urban management and public security. Recently, deep learning has shown a great advantage in making the quality of crowd counting more accurate. However, how to apply deep learning models to embedded terminals is still a challenging issue. The main contradiction of the problem lies in high demand for computing resources for deep learning and strict limitation of power consumption from embedded devices. In order to achieve the crowd counting by edge computing (in embedded terminals), we propose a tiny model based on convolutional neural networks. The model can be switched into other two forms to adapt tradeoff between the accuracy and efficiency. Especially, different forms share main parameters so as to save storage and avoid retraining. What is more, this model supports variant sizes of input images, which benefits the applications to a variety of embedded devices. The experiments on two different embedded terminals have shown the effectiveness, efficiency, and flexibility of our proposed model.