News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Computer vision and deep learning
Computer vision and deep learning
Artificial intelligence is a wonderful dream for human beings, like interstellar roaming and immortality. We want to create a machine that has the same ability to perceive things as people, such as seeing the world.
In the 50s of last century, mathematician Turing put forward the standard of judging whether the machine had artificial intelligence: Turing test. The machine is put in a room, the human tester is in another room, people talk to the machine, the tester doesn't know whether the other room is human or machine. After chatting, if the tester can't be sure whether he or she is chatting with a person or machine, then Turing test is passed, that is to say, this machine has the same perception ability as human.
But out of the Turing test to start at the beginning of this century, 50 years have countless scientists put forward many machine learning algorithms, and tries to let the computer have same level of intelligence, but until 2006 the deep learning algorithm, to bring a glimmer of hope for a solution.
The deep learning.
Deep learning in many academic fields, compared with non - depth learning algorithms, often has the improvement of 20-30%. Many large companies are also starting to invest in this algorithm, and set up their own deep learning team. The largest investment is Google, which disclosed Google brain project in June 2008. Google bought DeepMind in January 2014, and in March 2016 the development of the Alphago algorithm in the Chess Challenge, defeated nine players Li Shishi, that deep learning algorithms can overcome the world's strongest players.
In terms of hardware, Nvidia began to make display chips. But from 2006 and 2007, GPU chip was used to perform general computation. It is especially suitable for a lot of simple and repeated computation in deep learning. At present, many people choose the CUDA toolkit of Nvidia for the development of deep learning software.
Since 2012, Microsoft has used deep learning to do Machine Translation and Chinese speech synthesis. Behind AI Na, there is a set of data algorithms for Natural Language Processing and speech recognition.
In 2013, Baidu announced the establishment of a Baidu Research Institute, the most important of which was the Baidu deep study institute, when Dr. Yu Kai, a famous scientist, was recruited. But then he left Baidu and created another company horizon for the development of deep learning algorithms.
Facebook and Twitter have carried out deep learning research respectively, in which the former joined professor Yann Lecun of New York University, and established their own deep learning algorithm laboratory. In October 2015, Facebook announced its open source deep learning algorithm framework, namely Torch framework. Twitter acquired Madbits in July 2014 to provide high precision image retrieval services for users.
Computer vision in the era of advanced learning
The Internet giant, which emphasizes deep learning, is certainly not for the sake of learning, but mainly that it can bring a huge market. Why did the traditional algorithms fail to reach the precision of deep learning before the depth of learning?
Before deep learning algorithm comes out, for visual algorithm, it can be roughly divided into 5 steps: feature perception, image preprocessing, feature extraction, feature selection, inference prediction and recognition. In the early machine learning, the dominant statistical machine learning group was not concerned about the features.
In my view, computer vision is the application of machine learning in the field of vision, so computer vision has to design the first 4 parts of its own when adopting these machine learning methods.
But it's a difficult task for anyone. The computer to the traditional recognition methods of feature extraction and classifier design separately, then when applied together again, for example if the input is an image of the motorcycle, the process must first have a characteristic expression or feature extraction, and then the characteristics expressed in classification learning algorithm.
In the past 20 years, there have been a number of excellent feature operators, such as the most famous SIFT operator, the so-called operator that keeps the scale rotation constant. It is widely used in image comparison, especially in the so-called structure from motion applications, with some successful application examples. The other is the HoG operator, which can extract objects, compare the robust edge of the object, and play an important role in the detection of objects.
These operators also include Textons, Spin image, RIFT and GLOH, all of which occupy the mainstream of visual algorithm before deep learning is born, or deep learning is really popular.
Several (half) examples of success
These characteristics and some specific classifier combinations have made some successful or semi successful examples, which basically meet the requirements of commercialization, but have not yet been fully commercialized.
One is the fingerprint recognition algorithm in 80s and 90s. It has been very mature. It usually looks for some key points on the fingerprint pattern, looks for points with special geometric characteristics, then compares the two fingerprint key points to determine whether they match.
Then, the face detection algorithm based on Haar in 2001, which has been able to achieve real-time face detection at the time of hardware condition. Now we all face detection in all cell phone cameras is based on it or its variants.
The third is the object detection based on the HoG feature, which is combined with the corresponding SVM classifier which is the famous DPM algorithm. The DPM algorithm exceeds all algorithms in object detection and has achieved good results.
But there are too few examples of success, because the handmade design features require a lot of experience.
In the 50s of last century, mathematician Turing put forward the standard of judging whether the machine had artificial intelligence: Turing test. The machine is put in a room, the human tester is in another room, people talk to the machine, the tester doesn't know whether the other room is human or machine. After chatting, if the tester can't be sure whether he or she is chatting with a person or machine, then Turing test is passed, that is to say, this machine has the same perception ability as human.
But out of the Turing test to start at the beginning of this century, 50 years have countless scientists put forward many machine learning algorithms, and tries to let the computer have same level of intelligence, but until 2006 the deep learning algorithm, to bring a glimmer of hope for a solution.
The deep learning.
Deep learning in many academic fields, compared with non - depth learning algorithms, often has the improvement of 20-30%. Many large companies are also starting to invest in this algorithm, and set up their own deep learning team. The largest investment is Google, which disclosed Google brain project in June 2008. Google bought DeepMind in January 2014, and in March 2016 the development of the Alphago algorithm in the Chess Challenge, defeated nine players Li Shishi, that deep learning algorithms can overcome the world's strongest players.
In terms of hardware, Nvidia began to make display chips. But from 2006 and 2007, GPU chip was used to perform general computation. It is especially suitable for a lot of simple and repeated computation in deep learning. At present, many people choose the CUDA toolkit of Nvidia for the development of deep learning software.
Since 2012, Microsoft has used deep learning to do Machine Translation and Chinese speech synthesis. Behind AI Na, there is a set of data algorithms for Natural Language Processing and speech recognition.
In 2013, Baidu announced the establishment of a Baidu Research Institute, the most important of which was the Baidu deep study institute, when Dr. Yu Kai, a famous scientist, was recruited. But then he left Baidu and created another company horizon for the development of deep learning algorithms.
Facebook and Twitter have carried out deep learning research respectively, in which the former joined professor Yann Lecun of New York University, and established their own deep learning algorithm laboratory. In October 2015, Facebook announced its open source deep learning algorithm framework, namely Torch framework. Twitter acquired Madbits in July 2014 to provide high precision image retrieval services for users.
Computer vision in the era of advanced learning
The Internet giant, which emphasizes deep learning, is certainly not for the sake of learning, but mainly that it can bring a huge market. Why did the traditional algorithms fail to reach the precision of deep learning before the depth of learning?
Before deep learning algorithm comes out, for visual algorithm, it can be roughly divided into 5 steps: feature perception, image preprocessing, feature extraction, feature selection, inference prediction and recognition. In the early machine learning, the dominant statistical machine learning group was not concerned about the features.
In my view, computer vision is the application of machine learning in the field of vision, so computer vision has to design the first 4 parts of its own when adopting these machine learning methods.
But it's a difficult task for anyone. The computer to the traditional recognition methods of feature extraction and classifier design separately, then when applied together again, for example if the input is an image of the motorcycle, the process must first have a characteristic expression or feature extraction, and then the characteristics expressed in classification learning algorithm.
In the past 20 years, there have been a number of excellent feature operators, such as the most famous SIFT operator, the so-called operator that keeps the scale rotation constant. It is widely used in image comparison, especially in the so-called structure from motion applications, with some successful application examples. The other is the HoG operator, which can extract objects, compare the robust edge of the object, and play an important role in the detection of objects.
These operators also include Textons, Spin image, RIFT and GLOH, all of which occupy the mainstream of visual algorithm before deep learning is born, or deep learning is really popular.
Several (half) examples of success
These characteristics and some specific classifier combinations have made some successful or semi successful examples, which basically meet the requirements of commercialization, but have not yet been fully commercialized.
One is the fingerprint recognition algorithm in 80s and 90s. It has been very mature. It usually looks for some key points on the fingerprint pattern, looks for points with special geometric characteristics, then compares the two fingerprint key points to determine whether they match.
Then, the face detection algorithm based on Haar in 2001, which has been able to achieve real-time face detection at the time of hardware condition. Now we all face detection in all cell phone cameras is based on it or its variants.
The third is the object detection based on the HoG feature, which is combined with the corresponding SVM classifier which is the famous DPM algorithm. The DPM algorithm exceeds all algorithms in object detection and has achieved good results.
But there are too few examples of success, because the handmade design features require a lot of experience.