请选择 进入手机版 | 继续访问电脑版
客户端

铂链研究成果于本月在国际科技期刊IJCIT发表!

所在版块: 区块链技术 2017-11-27 16:59   [复制链接] 查看: 1721|回复: 233
日前,由铂链首席科学家高振教授领衔的铂链研究团队,经过长期、深入的研究,成功取得了AI人脸识别领域研究成果。研究的主体——AI人脸识别模型,由铂链系统采集数据,并根据采集到的精确小数据训练AI,最终成功得到模型,相关研究论文将在本月于国际科技期刊IJCIT发表。


21.png

IJCIT是首个开放同行评议的国际学术双月期刊,旨在提供一个平台,鼓励全球新兴行业学者和学者分享他们在计算机科学、工程、技术和相关学科领域的专业和学术知识。 IJCIT 期刊在全球人工智能及新兴科技领域享有盛誉。
此次基于小数据的人脸识别论文"Artificial Intelligence based Fast Facial Recognition with High-Quality Small-Scale Data" 被《计算机与信息技术国际期刊》(《IJCIT》)经评议后被正式录用发表,是国际学术界对于铂链研究团队研究方向及研究能力的认可。

本次由高教授带领的铂链研究团队,旨在在研究中对人工智能更深层次应用高质量的小数据的瓶颈进行破冰。 铂链则在此次研究中作为数据提供商、图像处理和人工智能技术的集成。这种方法被设计并应用于较大规模的环境中,同时将在便携式摄像机和微控制器单元中进行研究,该单元能对未知面部进行检测和分类。几种不同的人工智能学习技术和神经网络也在低成本微控制器中进行了研究和测试。这项研究的预期是建立一个基于独特的识别方法的安全系统。此外,还有一个基于共识的一站式应用平台,可帮助人工智能快速、轻松地开始工作。

铂链CEO宋欣先生表示,AI项目及学术研究的成果,离不开数据的真实可靠。如何获取数据定将成为急需解决的问题与痛点,铂链愿与世界各地的优质AI项目及学术研究领域学者紧密合作,通过区块链技术及铂链独创的数据入股合约,为更多AI产业提供优质数据,共同助力人工智能时代的快速到来。

那么是怎样的模型与研究成果让这样一个具有国际重量级学术地位的科技期刊破例加快流程评议并刊登发布呢?让我们一起来了解一下吧!
以下为期刊正文内容:注意:本学术论文内容及版权属于铂链所有,未经允许请勿转载刊登。
Artificial Intelligence based Fast Facial Recognition with High-Quality Small-Scale Data

Zhen Gao*, Tingting Wang, Xin Song, Chao Wang
Tuzi Network Technology Shanghai
Shanghai, China
*Corresponding Author's E-mail: bottos2050[AT]gmail.com

AbstractThe bottleneck of the deeper applications of artificial intelligence is the lack of high-quality small-scale data. Using the integration of image processing and artificial intelligence technologies a system is developed to provide full automation based on the data supported by BOTTOS data provider. In order to apply this method to a larger scale, a proof of concept will be investigated that utilizes a portable camera and microcontroller unit that detect and classify an unknown face. Several different machine learning techniques and neural networks were also investigated and tested in order to determine the performance on a low cost microcontroller and the effectiveness in successfully identifying a random face. The plan of this research is to build a safety system based on unique identification methods. Furthermore, a consensus-based one-stop application platform to help AI projects get training data fast and easily through smart contract and to generate wealth value for personal data through data mining is under construction.
Keywords- Facial recognition; Image processing; Artificial intelligence methods; Automated system   
Introduction
Vision is the most essential function for an intelligent machine to explore and interact with the surroundings especially when an autonomous system wants to be aware of the real-life environment and the existence of agents, obstacles, terrain, and targeted objectives. That explains why image processing and machine vision were extensively investigated by researchers from various fields including but not limited to computer science, automation engineering, biomedical engineering, mechanical engineering and electrical engineering [1-5]. With the development of artificial intelligence methods and their real-world applications in science and technology, most of the decision making process will be deployed with least human involvement. Besides, the continuously evolutionary artificial intelligence methods have the ability or potential to produce anthropomorphic results [6-8].
Face recognition is very natural methods of identification for users as they do not have to perform any strange tasks to interact with the technology. This technology is non-invasive which makes easy to prepare and begin with the recognition processes. A user with very limited technical ability would still be able to interact with the face recognition.
As a primary modular of artificial intelligence methods, artificial neural network (ANN) played an important role for non-linear modelling, process prediction, pattern classification and matching based on abstraction and simulation of human nervous system with weighted directed connection which represents the bio-physical connection of human’s neurons [9-11]. The major role of neural network is to train and derive the complicated, nonlinear and dynamic mapping between input singles and output singles without an explicit mathematical discerption. This process could be treated as modelling a black box with collected sample data for training. Due to the powerful learning ability and rapid development of hardware, presently ANN as a method has been widely accepted for various real-life applications [12-14]. With the seamless integration between image processing and artificial intelligence methods, plus the high-quality small-scale data supported by BOTTOS data provider the process of real-time classification and recognition will be more efficient and the results are expected to be more reliable and robust.  
Objective
The scope of this project is to build a small, portable image processing and classification subsystem for a home security system utilizing open source software and easily sourced hardware components based on the integration of vision and artificial intelligence methods for facial recognition. Initial research was completed to examine current trends in implementing artificial intelligence into a process, and specifically an artificial neural network with high efficiency, requiring low processing power with a high success rate.

The major components of the home security system to be built are the used is as follows. Voice control is being performed using a licensed protocol / server called BitVoicer server. This server is being connected to an Arduino microcontroller and mic for audio input and control to unlatch the safe as an output. A Raspberry Pi will be programmed to use a Pi Camera in conjunction with the OpenCV library to perform facial recognition. Once the face is recognized, the Pi will turn the servo and open the lock. The muscle control is designed to unlock the lock after an arm movement pattern has been inputted and identified. As shown in Figure 1, EMG sensors and an Arduino microcontroller are being used to analyze the muscle data.

112721.png
Figure 1. Workflow
Experitmental Setup
The topology of the facial classification system is relatively simple due to the complexity of the machine learning and artificial neural network architectures. All inputs and outputs of the system will be directly connected to the ports available on the Raspberry Pi. The camera module connects to the Raspberry Pi using the ribbon that directly connects to the board. The camera module will send color images directly back to the Raspberry Pi. The HMI will be connected to the HDMI or USB port depending on the device that will be chosen. The audio system connects to the audio port of the Raspberry Pi and will be configured receive input queues from the Raspberry Pi. Power is provided at 5 Volts at 2A that can be attained using a micro USB to standard outlet connection. An industrial environment device remote access is set up to ensure that multiple users can operate or monitor the operation of the device. Remote device access is completed by configuring the Ethernet port and connecting a network switch. A remote access point will only be set up if the initial design criteria are met.
The neural network will work in four sequential layers: input, multiple feature analysis, classification and output. Each layer will be designed to capture and compute relative information and pass it on to the next layer. The conceptual design focuses on the characteristics of shape & size, colour and texture to train and test the program for successful determination at a high confidence level. The HSV or L*a*b* colour space models will be used to define the colour of the object. These models will allow the network to be trained on which colours to recognize. The algorithm will be adjusted to recognize the gradient of the colour to identify ripeness. The shape of the object by first defining the area and perimeter of the object that are found by considering the pixels they inhibit. Using the relationship between these two features the shape of the object can then be defined. The texture of the fruit is found by passing the image through a Log-Gabor filter that analyzes the frequency domain of the defined region of the object. The processed image is then analyzed to find the repeating feature, defining that as the texture. The artificial neural network now has to be built using a pre-set database and the m
ultilayer perceptron neural network libraries in OpenCV. To attain the desired accuracy a database of 50 images for each object will be used. The artificial neural network will be trained on the computer first and then transferred to the microcontroller once refinement of the system is complete. Following figure shows the logic flow of this method.
The motion detection program was tested using the original setup. The size of the search window that defined the object in motion had to be adjusted to account for any issues. Furthermore the timing of image capture and the amount of pixels that defined movement had to be adjusted to best simulate a real world application.
After the enclosure was built all modules of project were put together and transferred to the Raspberry Pi. A main program was built that integrated and utilized all modules of the program. The Raspberry Pi had minor issues running the more intensive neural networks, running out of memory if a classification could not be easily resolved. This problem occurred infrequently and did not have a large impact on the overall operation.
Test
The current setup of the device is meant purely for capturing data samples and performing analysis on them, as well as testing and comparing the effectiveness of several machine learning techniques. When the user starts up the device they are presented with four options. The first option is to allow the user to capture samples to expand the current database.
If all required samples have been acquired the user can then choose the second option, which is to process these new images. The image processing will begin on new samples that have not been analyzed which are determined by referring to a document that saves the progress of the image processing event at each stage. A reference image will be shown to the user and then the user will provide the system with an id of the object. The program will then complete the extraction of 14 features of the object and save this data to a file that will be referenced during the training phase.
The image processing and classification device will have the following functions:
1. Central programming structure that allows easy navigation to different modes of the device.
2. Three modes that the user may choose from: image capture, image processing, or testing of new samples.
3. Detect and track an object that comes into view of the camera.
4. Perform image analysis.
5. Classify an object.
6. Report to the user the ID of the object.
The system will use both audio and visual system to report the id of the object to the user. The visual system will notify the user of the progress of the system, the name of the object as well as the confidence level of the identification. During the image capturing, processing and identification processes the user will be shown the image of the object at each major step to ensure the image analysis is being completed without error.
Due to the design of the system and the environment it will be used in the following constraints are present:
1. System to be designed tested and built within a low cost budget.
2. Utilize open source software to allow accessibility to a larger user base.
3. Hardware used must be able to easily integrate with all systems of the device.
4. System must be used in an environment with a static background to prevent object detection confusion.
The constraints of this research are meant as guidelines to follow during the design and implementation phases. If the project is completed by the allotted deadline, the algorithms utilized will be refined to adjust for environments with noise due to light conditions and background variation.
The applied co-evolutionary ANN is regarded as a sole population. In this figure, the connection condition between different layers and the related weights are treated as a sub-population. By splitting it into several sub-structures, each sub-structure will be evolved in a more efficient way by communicating and cooperating with other sub-structures, as shown in the following figure.
112722.png
Figure 2. Algorithm step and flow chart
Totally, 100 images are used for training, and 20 images are collected for testing. Figures 3 and 4 show a training sample and a passing test example, respectively. The results of several training methods are compared as follows:
EigenFace: 75%
FisherFace: 75%
LBPH: 80%
CNN: 90%
Co-evolutionary ANN: 95%
112723.png
Figure 3. Traning sample (note: the portrait is only used for research purpose)
112724.png
Figure 4. Testing Sample (note: the portrait is only used for research purpose)
Conclusions
Using the training data provided by BOTTOS user, we were able to successfully recognize the test subject in with good to great accuracy. The testing uncovered some performance issues with the Raspberry Pi that we were unaware of previously. Further testing will be done to optimize the number of photos required to train the model. If time permits, we would also like to look into using machine learning for recognition. In terms of additional work for the PI, we are looking into interacting with the servo motor to unlock the deadbolt once the face has been recognized.
References
[1] Teng Wang, Juequan Chen, Xiangdong Gao and Wei Li. Quality Monitoring for Laser Welding Based on High-Speed Photography and Support Vector Machine. Appl. Sci. 2017, 7(3), 299; doi:10.3390/app7030299
[2] Xiaochun Zheng, Yankun Peng and Wenxiu Wang, A Nondestructive Real-Time Detection Method of Total Viable Count in Pork by Hyperspectral Imaging Technique, Appl. Sci. 2017, 7(3), 213; doi:10.3390/app7030213
[3] M.V. Ananyev, D.I. Bronin, D.A. Osinkin, V.A. Eremin, R. Steinberger-Wilckens, L.G.J. de Haart, J. Mertens, Characterization of Ni-cermet degradation phenomena I. Long term resistivity monitoring, image processing and X-ray fluorescence analysis, Journal of Power Sources, vol 286, 2015, pp. 414-426
[4] Jih-Gau Juang, Yi-Ju Tsai and Yang-Wu Fan, Visual Recognition and Its Application to Robot Arm Control, Appl. Sci. 2015, 5(4), 851-880; doi:10.3390/app5040851
[5] Bingfei Nan, Zhichun Mu, Long Chen and Jian Cheng, A Local Texture-Based Superpixel Feature Coding for Saliency Detection Combined with Global Saliency, Appl. Sci. 2015, 5(4), 1528-1546; doi:10.3390/app5041528
[6] Xibin Jia, Shuangqiao Liu, David Powers and Barry Cardiff, A Multi-Layer Fusion-Based Facial Expression Recognition Approach with Optimal Weighted AUs, Appl. Sci. 2017, 7(2), 112; doi:10.3390/app7020112
[7] Robail Yasrab, Naijie Gu and Xiaoci Zhang, An Encoder-Decoder Based Convolution Neural Network (CNN) for Future Advanced Driver Assistance System (ADAS), Appl. Sci. 2017, 7(4), 312; doi:10.3390/app7040312
[8] Chi-Ying Lin and Hong-Wu Jheng, Active Vibration Suppression of a Motor-Driven Piezoelectric Smart Structure Using Adaptive Fuzzy Sliding Mode Control and Repetitive Control, Appl. Sci. 2017, 7(3), 240; doi:10.3390/app7030240
[9] Junya Lv, Huiyu Ren and Kai Gao, Artificial Neural Network-Based Constitutive Relationship of Inconel 718 Superalloy Construction and Its Application in Accuracy Improvement of Numerical Simulation, Appl. Sci. 2017, 7(2), 124; doi:10.3390/app7020124
[10] Dajie Song, Lijun Song, Ye Sun, Pengcheng Hu, Kang Tu, Leiqing Pan, Hongwei Yang and Min Huang, Black Heart Detection in White Radish by Hyperspectral Transmittance Imaging Combined with Chemometric Analysis and a Successive Projections Algorithm, Appl. Sci. 2016, 6(9), 249; doi:10.3390/app6090249 Rong Shan, Zeng-Shun Zhao, Pan-Fei Chen, Wei-Jian Liu, Shu-Yi Xiao, Yu-Han Hou, Mao-Yong Cao, Fa-Liang Chang and Zhigang Wang
[11] Network Modeling and Assessment of Ecosystem Health by a Multi-Population Swarm Optimized Neural Network Ensemble, Appl. Sci. 2016, 6(6), 175; doi:10.3390/app6060175
[12] Lei Si, Zhongbin Wang, Ze Liu, Xinhua Liu, Chao Tan and Rongxin Xu, Health Condition Evaluation for a Shearer through the Integration of a Fuzzy Neural Network and Improved Particle Swarm Optimization Algorithm, Appl. Sci. 2016, 6(6), 171; doi:10.3390/app6060171
[13] Ashfaq Ahmad, Nadeem Javaid, Nabil Alrajeh, Zahoor Ali Khan, Umar Qasim and Abid Khan, A Modified Feature Selection and Artificial Neural Network-Based Day-Ahead Load Forecasting Model for a Smart Grid, Appl. Sci. 2015, 5(4), 1756-1772; doi:10.3390/app5041756
[11]Atsushi Yona, Tomonobu Senjyu, Toshihisa Funabashi, Paras Mandal and Chul-Hwan Kim, Optimizing Re-planning Operation for Smart House Applying Solar Radiation Forecasting, Appl. Sci. 2014, 4(3), 366-379; doi:10.3390/app4030366






回复

使用道具 举报

wubo

发表于 2017-11-27 17:06:30 来自手机 | 显示全部楼层
要搞事了么666
回复

使用道具 举报

bto147

发表于 2017-11-27 17:53:55 | 显示全部楼层
要搞事儿嘛,哈哈
回复

使用道具 举报

zw0832

发表于 2017-11-27 18:26:04 来自手机 | 显示全部楼层
VRC11月27号正式上线挖矿,免费注册领矿机  币圈资讯 2017-11-27 16:05 发帖在【莱特挖矿 比特币论坛  以太坊社区】     VRC矿机正式上线挖矿  11月27号免费注册送矿机       这是APP下载链接https://fir.im/vrc 下载完点击注册,填你的手机号,验证码,密码。最后一栏邀请码填2729 注册成功后去个人资料里实名认证和银行卡认证。认证完即可获得迷你矿机                                         VRC. qq交流群 658870905
回复

使用道具 举报

lixiwu

发表于 2017-11-27 19:06:34 | 显示全部楼层
铂链太牛逼了。。好样的
回复

使用道具 举报

bigbot

发表于 2017-11-27 23:05:08 来自手机 | 显示全部楼层
铂链好样的,加油
回复

使用道具 举报

HHJ1404593975

发表于 2017-11-27 23:28:06 | 显示全部楼层
66666666666
回复

使用道具 举报

caomin

发表于 2017-11-27 23:52:14 | 显示全部楼层
铂链强到一笔
忘记时间
回复

使用道具 举报

ghq518

发表于 2017-11-28 00:17:10 | 显示全部楼层
铂链太牛逼了..好样的
回复

使用道具 举报

18823369165

发表于 2017-11-28 01:09:23 | 显示全部楼层
铂链雄起
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则

扫描关注微信公众号及微博

Archiver|手机版|小黑屋|铂链BOTTOS社区 | 国内领先的区块链公益社区

GMT+8, 2018-5-23 13:04 , Processed in 0.105381 second(s), 34 queries .

铂链BOTTOS社区 | 国内领先的区块链公益社区

© 2017 bottos

快速回复 返回列表