At present, the landing of CV is mainly focused on security, medical, retail, and automatic driving; there are also a small number of solutions for specific scenes in several industries, such as Internet short video, logistics, and intelligent education.
In 17 years, the artificial intelligence market size of 23.7 billion, CV accounted for 34.9%.
Among them, 67.9% of the revenue is contributed by the security field, and 18.1% from advertising and marketing.
At present, it is still security that occupies a large part of the CV market structure, which may be related to the industry characteristics and hardware/algorithm development path.
Compared to other traditional industries, the security industry has a stronger "identification identity" needs, the scene from the initial 1.1v1,1vN, fingerprint comparison? 2. ID card face comparison to the latest 3. static face automatic detection 4. dynamic face detection & object behavior trajectory analysis? 5. Intelligent case analysis The direction of development from the simple identification of identity, gradually to the subsequent more business scenarios.
Organize the financing information of AI startups from January to August 18, find out the latest industry landing direction, and analyze the competitiveness of products.
On 18 years, there were 199 startups **** financed, 36 in the direction of computer vision general services, and 30 related to CV.
The landing industry, in order of the number of companies: medical, retail, industrial inspection, intelligent driving, intelligent education, security, which has focused on the underlying technology to provide multi-industry solutions, such as Shangtang, Kuangshi, but also focus on one industry, such as HaHa retail, a pulse of sunshine.
Medical: based on computer vision technology, intelligent CT, intelligent X-ray assisted screening, assisting radiologists to diagnose the condition, in addition to the initial assisted diagnosis of the condition, there are other AI programs for medical pain points, such as the medical deep learning platform provided by the presumption of science and technology, and the drug research and development of deep learning platforms. In terms of operation mode, basically it is to cooperate with hospitals and provide intelligent diagnosis and treatment systems. The difficulties of the industry are in the difficulty of obtaining high-quality data, the relative complexity of business scenarios, and the need for labeling personnel with a professional medical background.
Retail: provide computer vision-based unmanned shelves program, automatic identification of the goods taken by the user, in addition to the analysis of foot traffic, merchandise sales, the output of a complete set of Supply - & gt; Transportation - & gt; Sales Digitalization The program is a great way to get the most out of your business.
Industrial inspection:? Provide deep learning-based visual inspection of appearance defects, accurate measurement technology equipment, the target industry is cell phone processing, automotive, 3C, etc., the value in improving the automation rate of the enterprise's processing links, reduce labor costs, improve product quality.
Intelligent education: and CV-related, machine marking.
According to the more important factors affecting AI, algorithms, data, and landing scenarios, the relevant companies are divided into several categories
Companies such as Shunfeng and Headline, which have a wealth of data accumulation, through algorithms, and output of AI solutions to solve the pain points of their own business.
There is also strategic cooperation with third-party traditional industry leaders, holding the thigh to pick up the demand, specializing in solving each other's business pain points, such as the polar perspective and China Resources strategic cooperation.
Finally attached a head CV company in various industries product layout.
Background: the process of seeing a doctor can be simplified into three steps, pre-diagnosis: daily exercise, great health care, prevention of disease. Diagnosis: intelligent triage - > auxiliary diagnosis - > prescription recommendations. Post-diagnosis: maintenance, other not in the mainstream process, there are drug development, doctor learning, etc..
Observing the entire diagnosis and treatment process, there are images generated & high repetitiveness & artificial a large number of concentrated links, mainly in the diagnosis, Tencent foraging to provide auxiliary diagnostic capabilities, but also in the diagnosis of the link, for the patient's CT / X-Ray, to the doctor to provide the possibility of disease advice, the colleague of the doctor's efficiency, but also improve the judgment of the middle-level doctors.
From the official website of Tencent Foraging, it is currently able to provide diagnostic capabilities for six cancers, including lung cancer, esophageal cancer, etc., and the detection rate OR the accuracy rate can be more than 90%.
From the hospital announcement, it mainly plays the role of a reminder to assist the doctor's operation.
The scene is very clear, the design of e-commerce banner, the initial stage is to assist, later replacement.
Four core generation steps:
First, let the machine understand what the design is composed of: through manual data annotation, the layers in the original file of the design to do the classification, the elements to do the labeling. A team of design experts will also refine the design techniques and styles. By telling the machine why these elements can be put together by way of data, we feed the experience and knowledge of the experts through the data. The core of this part is the algorithmic model of deep sequence learning.
The second step is to build the element center: when the machine learns the design framework, it needs a lot of production materials. We will establish an element library, do image feature extraction by machine, and then classification, and then manually control the image quality as well as copyright issues, we bought a copyrighted gallery, but also hope to avoid disputes over copyright from the beginning.
The third step, the generated system: the principle is a bit like Alpha Go playing Go. We build a virtual canvas on the design framework, similar to a chessboard, the generated system puts the elements in the center of the elements to the chessboard, where we use "reinforcement learning", as if you put a sweeping robot at home, let it run by itself, run a few laps, it will know where there are obstacles to be avoided. In the process of reinforcement learning, the machine refers to the original samples, and through continuous experimentation, gets some feedback, and then learns what kind of design is right and good.
The fourth step is the evaluation system: we will capture a large number of finished designs and evaluate them from both the "aesthetic" and "commercial" aspects. The aesthetic evaluation is carried out by human beings, and there are professional crowdsourcing companies in this regard; the commercial evaluation is to look at the click-through rate and the number of views that have been put out.
From the distribution of the products of the startups, there are three modes
1. Head startups, competitiveness in maintaining technical advantages, the future of the new industry direction, the first time to cut up, because of the technical advantages, so even if you find out a little late, you can also quickly form their own competitive program.
2. Utilizing industry resources, vertically do a direction, the algorithm may not be dominant, but because the competitiveness lies in the resources, you can ensure that the product is quickly landed, the accumulation of resources, data advantage.
3. Do programmers, they do not have data, arithmetic, output algorithmic capabilities, hold one or two traditional industry thighs, pick up demand, no precipitation for themselves.
At present, CV value in two aspects
1. Replacement of auxiliary labor, business efficiency
People can do the machine can do, the main gain in efficiency, corresponding to the Luban, machine marking, face monitoring, unmanned shelves are liberated from manpower, improve efficiency.
Specifically replace or assist, depending on the business difficulty of the scene.
2. Create a new scene experience
Face remote comparison, video advertising, AI beauty, processing massive data + low latency requirements, people can not do, you have to rely on the machine, creating a new scene and user experience.
1. looking for repetitive manpower concentration? 2. look for places where pictures/videos and users intersect.