English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The use of deep learning algorithms to intelligently identify objects from video has a wide range of applications. The more advanced system based on the tensorflow framework is proposed for deep neural network recognition of objects in this paper. Our work is mainly the following: (1) Using the vgg deep convolutional network architecture as the infrastructure. (2) Real-time recognition of objects in video. (3) The system framework can be used in substations to realize real-time identification of objects with security threats. Introduction Deep learning has developed rapidly in recent years and is active in many research fields. Deep neural networks show powerful forces in the field of computer vision, and their recognition accuracy is also getting higher and higher. Substations are an important part of power transmission. Because the distributed local environment is different and unattended, its equipment will be damaged or stolen. This situation will cause great trouble to the security of the power grid. The existing video surveillance technology requires the on-duty personnel to manually monitor the display, which not only requires a lot of manpower to judge, but also has a slow processing speed. Based on this situation, there is an urgent need for an intelligent video surveillance analysis technology for real-time detection and identification of objects[1] that affect the safety of substations . Identifying objects from surveillance video Generating natural language descriptions is a process of intelligent detection processing. At present, object recognition based on computer vision is still in an immature stage. Many models have been established and proposed, but there are some common problems: low recognition of objects, large number of training samples, and inability to detect in real time. Among them, the convolutional neural network (CNN) overcomes the subjectivity and limitations of the artificial extraction feature. In recent years, a variety of convolutional neural network frameworks have been proposed. For example, some advanced deep CNN frameworks use Caffe[2], Keras[3], etc. The performance and model design have drawbacks. The Tensorflow framework has a flexible portability. It is helpful for observing complex network structures and monitoring long-time and large-scale training. It is more flexible in implementing deep CNN and can adapt to different requirements and constraints. This article uses a deep CNN framework for real-time recognition of video objects, which can be applied in many research areas. Convolutional Neural Network Theory Basis Introduce In recent years, the convolutional neural network structure applied in the field of computer vision is mostly based on the LeNet convolution model structure proposed by Lecun in 1998 for handwritten letter recognition, until 2012 [4], with the release of large ImageNet data sets. With the increase in computing power, people regained their interest in research on convolutional neural networks. The structure of the convolutional neural network was originally inspired by the biological vision mechanism, just as the working principle of the human visual cortex described by Hube in his pioneering research. Subsequently, Neocognitron, proposed by Fukushima, is the predecessor of the convolutional neural network [5]. This alternate form is designed to imitate the processing mechanism in simple cells of biology. The schematic diagram of the structure is shown in the figure below. Figure 1. Neuro-sensory architecture. In addition, after the convolution operation, it will follow a non-linear change unit. The common nonlinear function is the correct linear unit ReLu. Its mathematical expression is: ; 0 = 0 ; 0 x ifx x x φ ≥   <  ( ) (1) Neural network structure is usually composed of input layer x, y, and the output layer stack multiple hidden layer h. Each layer is also composed of multiple units. Generally, each hidden unit j h receives input from all units at the previous layer and combines them in weighted combinations. The mathematical form of its nonlinear combination is as follows: ) j j ij i i h F b X ω = +∑ ( (2) Advances in Convolution Neural Networks CNN is a set of multilayer neural networks dedicated to two-dimensional data such as images and video[6]. CNN was the first truly successful deep learning method in which many layers of the hierarchy were successfully trained in a robust manner[7]. CNN has been proposed as an in-depth learning framework. There are two factors that promote the use of convolutional neural networks in computer vision. The first is that convolutional neural networks can use the 2D structure of the image and the high degree of correlation between adjacent pixels of the image. The second is that the convolutional neural network also introduces a pooling step to ensure the translation invariance of the image to a certain extent, which makes the model unaffected by the position change [8]. For example, for the target recognition task, the shallow layer in the convolutional network will learn some edge and corner features of the image, and at the deeper level, the features of the entire target can be learned [9]. CNN two cores refer to CNN's two core operations: convolution and pooling [10]. The random combination of convolution and pooling gives CNN a lot of flexibility, and so many familiar classic networks have also been born: AlexNet, VGGNet, Google Inception Net, ResNet, and the four networks are increasing in depth and complexity. AlexNet won the 2012 ILSVRC (ImageNet Large Scale Visual Recognition Challenge) competition, Google Inception Net won the 2014 ILSVRC championship, VGGNet won the 2014 ILSVRC runner-up, and ResNet won the 2015 championship. Visual Geometry Group Deep Convolutional Network Model Introduction to the Vgg Model Visual Geometry Group is referred to as Vgg. The VGG-16 convolutional network has 21 layers, including 13 convolutional layers, 5 pooled layers, and 3 fully connected layers. The specific network structure is shown in Figure 2. Among them, VGG-16 uses multiple convolutional layers of smaller convolution kernels instead of one convolutional core with larger convolutional layers. On the one hand, the parameters can be reduced. On the other hand, more nonlinear mappings can be made to increase the network[11]. The ability to express. We wrote it in TensorFlow in the vgg16.py file. Here we include a preprocessing layer that takes an RGB image with a pixel value in the range 0-255 and subtracts the average image value. Figure 2. Network structure for Vgg-16. Model Improvement After testing several classic models, we selected the classic model based on Vggnet[12]. We modified the model and added a sigmoid layer before the softmax layer in the original model. Because we care about the marginal distribution of each prediction parameter, not the joint distribution of all variables[13]. We used the tensorflow framework to run our improved model. Vgg-16 has 13 convolutional layers, and we trained it more on classic data sets. Experimen Results We implemented the code based on the tensorflow platform. The classical Vgg network and the improved Vgg network were respectively used in the experiment. The experimental results show that the algorithm converges quickly. We selected up to 400,000 iterations, with 64 images for each training iteration. We noticed that the model converges in the 280,000th iteration, so we can test with this model. In the test phase, we can identify objects according to the front camera, and the effect is very good. Figure 3 is our recognition rate for objects that appear in substations and pose safety threats. The accuracy of the improved model to some objects is shown in the figure below. Figure 3. The accuracy of the improved model in identifying some objects. We trained with classic VGG models and improved models, each model for the 3 biggest iterative training, and testing on 30000 images of validation set per 10000 iterations the generated model, finally identification error of the comparison results as shown. Figure 4. Comparison of Vgg and improved Vgg test phase. The Application Prospect of Deep Learning in Object Recognition Compared with image recognition, the application of deep learning in video classification is far from mature. The image features obtained from ImageNet training can be directly and effectively applied to various image-related recognition tasks and other different image test sets, with good generalization performance. For example, this paper uses deep learning to identify substation safety threats. In order to achieve the features that can be used for video analysis, deep learning must not only establish large-scale training data sets, but also need to study new depth models suitable for video analysis. Although deep learning has achieved great success in practice, the characteristics of the depth model obtained through big data training, such as the robustness, sparseness, and selectivity of occlusion and selectivity have attracted attention. For example, what is the beneficial invariability of recognition for each layer of transformation and what information has been lost? Recently, Mallat used wavelets to quantify the deep network structure, which is an important exploration in this direction. The internal structure of the depth model cannot be exposed, Various applications related to image recognition are also promoting the rapid development of deep learning in all aspects of network architecture, layer design, and training methods. Deep convolutional neural network is a major contributor behind the recent wave of AI. It can be predicted that in the coming years, deep learning will enter a period of rapid development in terms of theory, algorithms, and applications. Acknowledgement This paper was supported by National Natural Science Foundation of China (61401269,61572311), Shanghai Technology Innovation Shanghai Technology Innovation P

Research on Substation Real-Time Object Recognition Algorithm Based on Deep Learning