Object detection based on impulse neural network

Resource download address ：https://download.csdn.net/download/sheziqiong/85909649
Resource download address ：https://download.csdn.net/download/sheziqiong/85909649
1 introduction

1.1 Research background

Target detection is a technology that applies a specific computer algorithm to find the desired target in the image . In recent years , With The continuous development of computer hardware , Various algorithms of target detection have also ushered in a huge breakthrough , More and more used in Traffic detection 、 Smart payment 、 Medical imaging and other aspects . In computer vision , Target detection is better than image Class is a more complex problem , It not only needs to know the type of goal , We also need to achieve the positioning of the goal . therefore , Physical examination The test is more difficult , More challenging , The corresponding deep learning model will also be more complex .
There are many algorithms for target detection , Convolutional neural networks （Convolutional Neural Networks, CNN） Is its generation One of the table algorithms . It's a feedforward neural network , With convolution calculation and depth structure . at present , Based on convolutional neural network The target detection algorithm of network can be roughly divided into two modes , namely two­stage Patterns and one­stage Pattern ,two­stage model The detection process of formula is divided into two steps ： First, several candidate boxes are generated by the algorithm , Re pass CNN Check the candidate box classification ;one­stage Mode is end-to-end learning , Directly regress the confidence probability and position of the target , phase There is a loss of accuracy for , But the speed is slow two­stage The algorithm of pattern is faster .[1]
be based on two­stage The algorithm of ：

• R­CNN: By selective search （selective search） To determine the candidate box , Then uniformly compress the candidate boxes To size ; Then use CNN Feature extraction of candidate boxes ; Finally, use multiple support vector machines （SVM） The classifier classifies the output vector , Use boundary regression to generate the target region [2].

• Fast R­CNN： Still use selective search to identify candidate boxes , But input the whole picture into CNN, In volume Use the region of interest on the feature layer （Region of interest pooling,ROI pooling） operation , And from special Extract a feature vector with a specific length from the eigengraph ; Then input the eigenvector into the full connection layer , use softmax Classify them ; Finally, the candidate boxes belonging to the same feature are classified and their positions are regressed [3].

•Faster R­CNN： Use RPN (Region Proposal Network) Not selective search , Greatly reduce The time for extracting candidate boxes . take RPN and Fast R­CNN Combination , First, extract the features of the whole picture ; Then input the feature results into RPN; And then use ROI The pool layer fixes the size of the candidate box ; Finally, it belongs to Candidate box regression and adjustment of a feature [4].
be based on one­stage The algorithm of ：

• YOLO v1 And many subsequent improved algorithms ：YOLO Series algorithm is an advanced target detection algorithm Law . Because the whole detection framework is a whole , Therefore, the performance of the algorithm can be optimized end-to-end .

• SSD Series algorithm ： Multi scale feature map is used to detect ., Set a priori box , Use convolution to detect .

Impulse neural network （Spiking Neural Network, SNN）, Originated from brain science , Because of its rich space-time Neurodynamic properties of the field 、 Diverse coding mechanisms and ultra-low power consumption are known as the third generation Neural Networks . Here is front , Neural networks have gone through several stages of development ： The first stage is the perceptron stage , It can simulate human perception Force and by American neuroscientists Frank Rosenblatt stay BM704 The simulation is completed on the machine . The second stage is based on connection The multi-layer artificial neural network of knot theory (Artificial Neural Network, ANN), It originated in the 20th century 80 years Mid - .20 century 80 End of the decade , Distributed expression and back propagation algorithm are proposed . stay 2006 Years later , Deep volume Product network plays an important role , It has led the development of artificial intelligence in recent ten years [5].
ANN All areas of deep learning （ Such as computer vision and natural language processing ） Great success , but ANN Is biologically imprecise , It cannot accurately imitate the working mechanism of biological brain neurons , Lack of certain Biological interpretability of . In order to make the neural network closer to the human brain ,SNN And then came . But with ANN On all sides The wide application of face is different ,SNN There are still many problems to be solved in the field , Its research is still in the early stage of rapid development Stage .

1.2 Research significance

SNN As the third generation artificial neural network , Event driven mechanism based on neurodynamics , Make it good at efficiency Deal with complexity 、 Sparse spatiotemporal information . also SNN In the hardware circuit, it has the advantage of ultra-low energy consumption .2019 year Developed by Tsinghua University ANN/SNN Heterogeneous fusion tianjixin board Nature cover , Point out the depth of computer science orientation The cross fusion of degree learning and neuroscience oriented impulse neural network will be the development direction of artificial general intelligence [6].
The significance of this design paper is to explore the application of impulse neural network in target detection , The current mainstream pulse The training algorithm of impulse neural network has direct BP Training 、STDP Unsupervised training and well trained ANN The transformation of , although There are many training algorithms , however SNN There is still no mature training algorithm . For example, in larger and deeper network training , Facing the problem of pulse signal coding 、 Problems such as high training costs [5]. And in the realization of target detection , Need more Complex network structure , At present, only Kim et al. (2020) Waiting for someone in the classic YOLO Model On spiking­yolo. So , Based on different network structures SNN, To facilitate implementation in hardware Low power consumption of , And compared with the existing results , It has a certain significance .

1.3 Paper organization and structure

The research content of this paper is ： In the summary and analysis of domestic and foreign ANN To transform SNN On the theoretical basis of , benefit Use the existing ANN Target detection model , Analyze the losses in the transformation process , And the realization of various transformation means The way ; At the same time, the transformation model is compared with the work done by predecessors , Analyze different model pairs SNN The impact of transformation , And in pytorch Design and implement the model under the framework .

This paper extends the problem to theoretical support , And then to algorithm research and specific solution design organization , It is divided into the following six chapters ：
The first chapter is this chapter ： Firstly, the background and significance of pulse neural network and target detection are described , And then put forward The main content of this paper is to design a transformation model and compare it with previous work to analyze the effects of different models on SNN The conversion influence , Finally, the structure and framework of the paper are given .
In Chapter 2, pulse neural network is introduced systematically ： Development trend 、 Advantages and disadvantages 、 Learning methods, etc . Focus on ANN
To SNN The method of transformation is described
The third chapter gives the commonly used artificial neural network model in target detection (Artificial Neural Network,ANN), With one­stage Representative algorithm SSD For example , The structure of the model 、 Loss function and other aspects .
The fourth chapter introduces the design ideas and methods of the model
Chapter five combines experiments , separately ssd,spiking­ssd Comparison of ;yolo,spiking­yolo Comparison of . Give real Test results , Analyze .
Chapter six summarizes and analyzes , Analyze the existing results of the transformation model from multiple perspectives , And it is given that this kind of knot Possible causes of fruit .

Catalog
1 introduction 2
1.1 Research background 2
1.2 Research significance 3
1.3 Paper organization and structure 3
2 Impulse neural network 5
2.1 Learning algorithms 5
2.2 ANN Transformed SNN 6
2.3 ANN Pulse implementation of operation 8
Deviation normalization 8
Normalize except for outliers 8
3 ANN Model structure 10
3.1 SSD 10
3.2 YOLO 13
4 SNN Model structure 16
4.1 spiking­ssd 16
4.2 spiking­yolo 18
5 experimental result 22
5.1 Data set introduction 22
5.2 Introduction to experimental indicators 22
5.3 Comparison of experimental results 22
6 Summarize and discuss 25
6.1 The impact of data on results 25
6.2 Several factors that affect the results of the model 25
reference 26


















Resource download address ：https://download.csdn.net/download/sheziqiong/85909649
Resource download address ：https://download.csdn.net/download/sheziqiong/85909649