To get a better browsing experience, please use Google Chrome.Download Chrome
Free TrialAsk for Price
  • HOME
  • Products
  • Solutions
  • Customers
  • Knowledge Hub
  • About Us
  • Demo


How to identify images during the content review process?

Aug 2, 2023

In the process of content review and recommendation, the identification of target objects is a key link, which can be roughly divided into four categories:portrait recognition, object recognition, image/logo recognition and scene recognition.

1. Portrait Moderation

Portraits are the most intuitive visual subject when discussing visual recognition. Therefore, the ability to recognize portraits in pictures is the basis for solving compliance issues and realizing more advanced functions. Portraits include various factors such as posture, skin color, and attire, requiring enterprises to have a more comprehensive recognition capability.
Gender moderation: When performing gender identification on portraits, in addition to males and females, attention should also be paid to some minority groups that exist. In order to avoid the bias caused by the direct recognition of external male/female characteristics, causing unnecessary reputational risks.
Body posture recogni: recognize portrait posture and body shape, such as bust/full body, standing/sitting/prone, tall, short, fat, thin, etc. Because poses imply basic tendency information, for example: busts are more likely to be selfies, prone poses are more likely to contain pornographic risks, and so on. Therefore, the ability of gesture recognition can be used as one of the reference factors for comprehensively determining the risk of pictures. The height, shortness, fatness and thinness of the body shape can not only assist the accuracy of other recognition, but also serve as a reference factor for intelligent recommendation.
Glothing moderation:The recognition of portrait clothing is a necessary recognition ability for risk judgment and advanced functions. In terms of compliance, the area covered by clothing can be used as a judgment factor for pornographic risk; whether the clothing is police uniform, military uniform, royal noble clothing or militant clothing can be used as a reference factor for violations. When it has the ability to identify clothing styles, such as lolita clothing, gentleman suits, etc. it can provide advanced functions of interest recommendation.
Skin color moderatio:In the context of globalization, it is necessary to consider the generalization changes brought about by differences in skin color, especially some extreme scenarios that did not consider the effect of the model in the past, such as the high proportion of Asians in Latin America and the change in skin color caused by the history of intermarriage. Therefore, skin color recognition is not to be able to determine the specific skin color, but to ensure the model's good adaptability to portrait recognition.

2. Item Moderation

Items are an important visual element contained in the picture. The "items" here actually include inanimate items and living animals and plants. The ability to identify items is directly related to the effect of risk judgment and intelligent recommendation. Here we can divide "items" into three categories: sensitive items, ordinary items, and animals and plants.
Sensitive items:The identification of sensitive items is directly related to risk judgment, nd certain objects appearing in the screen directly indicate risks, such as guns and ammunition, drugs and related plants, gambling tables and slot machines, etc. But at the same time, regional differences should also be considered in policy settings.
Fox example: Due to the relatively common European colonial history and extensive religious beliefs in Latin America, special attention should be paid to religious items such as crosses and Bibles.
Ordinary Items:The ordinary items defined here are all kinds of common items, and the recognition of ordinary items serves more for the intelligent recommendation function. For example, if multiple smart phones, tablets, smart watches, and computer devices appear in a user's dynamic picture, then he is likely to be a digital enthusiast, or engaged in the production, research and sales of related products.
Animals and plants: The identification of animals and plants is also related to the requirements of content compliance and intelligent recommendation. In terms of content compliance, some animals and plants need to be identified because of their preciousness, such as tigers and other protected animals. There are also some animals and plants that need to be identified because of regional cultural taboos, such as toad patterns and animal totems in Latin America, which require targeted training of corresponding animal recognition models. In comparison, the requirements for intelligent recommendation are easier to understand. For example, if trees frequently appear in dynamic pictures, then he may like to hike outdoors, or work as a forest ranger. As for the cats and dogs appearing in the pictures, they can also be used as a basis for guessing pet preferences.

3. Symbol/Logo Moderation

Among the components of a picture, symbols and logos are often fixed graphics that occupy a small image area, but have very obvious and strong symbolic meanings. Symbols and logos include various national emblems, military logos, logos, station logos, religious symbols, etc. Even maps, which have a complete shape and great significance, can also be regarded as a kind of symbolic logo.
Here we can be divided into high-sensitivity signs and low-sensitivity signs.
High Sensitivity Logo:This kind of logo often has a very high sensitivity due to its meaning, and a sensitive strategy of "rather mistake than miss" should be adopted when identifying it. For example: Nazi-related logos, Buddhist "Swastika" logos, cross logos, maps, etc.
Low Sensitivity Logo:This type of logo mainly includes some other common logos, such as business logos, QR codes, TV station logos, watermarks, association logos, etc. The identification of such logos is mainly to meet the individual needs of the platform. For example, if you want to judge whether a video posted by a user involves a competing platform or whether it is suspected of copyright infringement, you need to identify whether there are key logos such as logos, watermarks, and logos in the screen.

4. Scene Moderation

In a general screen, the remaining part after excluding characters, objects, and logos can be regarded as a scene. In the design of some technical routes, the idea of ‚Äč‚Äčtarget detection and judging the risk of the subject is adopted, but the background of the screen actually contains quite a lot of reference information, which can also be used as a reference factor for the visual recognition results.
For example, scenes such as outdoors, street scenes, bedrooms, and bathrooms already contain the tendency of the subject's behavior, such as the combination of "street view" (scene) + "full body portrait" (body posture) + "trend" (dress) + "trend brand" (logo) It shows that the picture is likely to be about the content of fashion street photography, and the combination of the subject's behavior and background information can often form a more accurate comprehensive judgment result.

Live Chat