Research On Dynamic Number Detection With Tree-Fork Convolutional Neural Network

With the continuous development of Internet technology and the continuous improvement of people's living standards,the application of the Internet is also becoming more and more widespread.And the demands are getting higher and higher.From the initial access to information to immersion in life,food,clothing,housing,transportation,and other aspects,For example,it is applied to face recognition,intelligent driving,target detection and target tracking.In the application,the performance requirements are also getting higher and higher.It not only needs high accuracy,but also needs timeliness.Improving accuracy often results in a steep increase in computation.This will inevitably require more computer hardware and so on.But also easy to cause its delay,reduce the timeliness.Computer hardware has come a long way.The advent of 5G era is undoubtedly a window period for the development of artificial intelligence.Deep learning is an important branch of artificial intelligence.Its importance is self-evident.As an outstanding representative in the field of deep learning,it is also sought after by relevant researchers.This paper is based on the principle of convolutional neural network.By designing and improving the structure of the convolutional neural network to improve its network accuracy.And minimize network parameters.Make it more practical.The network performance is verified by image classification and recognition.On this basis,the target detection is studied and improved.In this paper,it is finally applied to dynamic number detection.The main research contents of this paper include the following four aspects:(1)Design tree-fork convolutional neural network.Cross convolution using multiple convolution kernel.Split-transform-split-merge method is adopted.Not only does it increase the complexity of the network,conducive to feature screening.Improve the generalization ability of the network.It can also extract more recessive features from the feature map.And the basic architecture of the network remains the same.The traditional convolution module in the middle convolutional layer of the network is replaced with a tree-fork module.By training in several open data sets.The performance of tree-fork convolutional network is compared with that of traditional convolutional network.In this paper,the network accuracy of tree-fork module in101_food,Caltech256,GTSRB and cifar10 is improved by 4.1 percentage points,4.7 percentage points,1.8 percentage points and 1.9 percentage points respectively compared with traditional CNN.It can be seen from the experimental results that the performance of the network in recognition accuracy is improved.(2)This paper uses the tree-fork module to improve the classic Darknet53.The residual structure is integrated into the tree fork module.It is beneficial to optimize tree-fork network.Take advantage of the robustness of the residual structure.Thus,overfitting caused by convolutional layer deepening can be alleviated.On the other hand,realize the optimization of the tree cross module to the deep network Darknet53.It is helpful to further verify the feasibility of tree-fork module.Improved Darknet53 has 2.4 percentage points more accuracy than Darknet53.(3)The target detection in this paper is accomplished by using a multi-task network.In other words,target location and target classification are realized through a network.Darknet53 and improved Darknet53 in chapter 4 are used as two basic networks for target detection.Bounding Box Prediction is adopted for target position positioning.In order to realize the detection of small objects,FPN is used for multi-scale detection.Multi-target detection is the detection of multiple target categories.For example,the COCO2014 dataset used in this article has 80 classes.After the training of target detection,80 kinds of targets can be detected.Three different scale feature maps were used for detection.For each scale,255 feature map are output.Includes target location,target confidence,and target category prediction.The mean Average Precision of the improved target detection network increased by 2.7 percentage points,and the detection time was shortened by 3 ms.(4)Dynamic number detection is realized.Dynamic number detection can detect pedestrians in each frame in real-time video stream.And count the number of pedestrians.In this paper,the convolution neural network method in deep learning is adopted.Dynamic number detection is a single target detection based on multi-target detection.The detection principle is basically the same.Dynamic number detection is also a hot and difficult point in the field of computer vision.It is not only necessary to detect pedestrians in video frames.It is also necessary to determine its location and size.In this paper,rectangular box is used.This is also similar to face detection.It is a typical target detection problem.On the one hand,dynamic number detection simplifies the calculation of multi-target detection.Make the detection more real-time.On the other hand,the convolution neural network method is more suitable for the actual needs.It can be used in intelligent robot,intelligent video surveillance and automobile driverless system(ADAS).So it is more practical.