2018 flying technology transfer, alibaba machine intelligence technology laboratory Liu Lei bring speech titled global precision image search introduction, mainly expounded from four aspects, the first part introduces the basic concepts of image search, the second part is mainly on the technology architecture of image search and its advantages, the third part analyses the application scenario and case, Finally, the use of goods and pricing to do a simple introduction. Here are the highlights:

Alibaba MIT

MIT (Machine, Intelligence,Technologies Machine Intelligence Technology Laboratory) was founded in 2018, composed of a group of outstanding scientists and engineers, mainly distributed in Hangzhou, Beijing, Seattle, Silicon Valley, Singapore and other places. Alibaba is responsible for ai technology research and development of the core team. MIT based on alibaba precious huge amounts of data, depending on the technology of machine learning/deep learning, build the cover image video, voice interaction, natural language understanding, intelligent decision, such as the core of artificial intelligence technology, fully energized electrical business, finance, logistics, social networking, entertainment and other important business of alibaba group, moreover also on ecological partners output, Join hands with enterprises to create an intelligent future.

Image search

Image Search service (Image Search) is a deep learning and large-scale machine learning technology as the core, through Image recognition and Search functions, Image Search by Image intelligent Image Search products. Its main service content includes the same image search and similar image search two parts, based on the image recognition technology, image search service combined with the application of different industries and business scenes, to help users to achieve the same or similar image search. Image search business is also more extensive, for all users with image database can enjoy image search services. The development process of image search is mainly divided into three stages. In August 2014, It was successfully launched in Bealitao, taobao, and users could shop by taking photos for the first time. In February 2017, the product began to be exported commercially and image search technology was shared with major business partners. This year began to attempt to extensively empower image search technology on Ali Cloud, and completed the simultaneous release of image search on Ali Cloud and the world in July.

Technical architecture and its advantages

Image search technology is mainly divided into five algorithm modules:

• Category prediction: it is limited to distinguish images only by image features, so target judgment can be made by category prediction. In this way, there is no need to search the whole image library globally, but only need to search the corresponding image of a certain category. Category search can not only improve the efficiency of search, but also improve the accuracy of search.
• Subject detection: In general, the subject target of the image obtained is small and the background is complex. In order to reduce the interference of background and other subjects, subject detection is needed.
• Feature extraction: Feature extraction is an important module of image search, which mainly applies two dimensions of feature depth and local feature. At present, feature extraction of deep learning has been far superior to traditional feature extraction. Local feature extraction is the expression of local heterogeneity of image features. At present, some improvements have been made in local features, including dimension compression and optimization of extraction speed.
• Retrieval and sort: retrieval index is mainly divided into offline build indexes and quick query two modules online, offline build index is through offline process to extract image features, the characteristics of the online process mainly through user input images to extract features, and then to extract the features into the distributed engine, a quick query retrieval. The sorting module can effectively combine the depth feature and the local feature. The depth feature mainly extracts the information from the high level, while the local feature focuses on the local information according to the image.

The technical process of image search

The offline process selects images from the image library, detects the main body of the selected images, and then extracts the features of the detected images, thus establishing an image retrieval engine. The online process is to predict and judge the class object when the user enters Query, and then perform subject detection and feature extraction. After retrieval and sorting, the result is finally output.

Advantages of image retrieval

The industry’s leading deep learning the deep learning algorithms and vast amounts of data, has the advantages of effect precision, strong robustness, and independent research and development support tens order of magnitude of other fast indexing technology, can achieve the response of the millisecond time, while products are all from alibaba electric business platform, experience of actual combat, so has the advantages of reliable, stable and effective. In addition, we can also do some customized development for different business scenarios, and integrate the computing capabilities of Ali Cloud and the visual AI capabilities of machine intelligence technology laboratory to create image search services that are fully suitable for users’ whole business scenarios.

Application scenarios and cases

Electricity market scene

Users simply take or upload photos and search based on them. This saves the tedious text description, simplifies the product search process, greatly improves the user’s shopping experience, and at the same time, merchants can also recommend products to users more quickly.

Common gallery scenario

Image sharing and social networking websites usually have a large number of images for users to search. By using image search service, an image search engine of 10 billion level can be quickly built on the cloud to provide the function of searching images by images and improve user experience.

case

This case is to make an image search technology in the system of Samsung Galaxy S8. Taobao related commodities can be searched in the album by searching pictures, which provides customers with a better shopping search experience.

Commercial and Pricing

Image search has been commercially launched on July 11 at both domestic and international sites, and has been widely praised by customers. At present, there are two billing methods, namely prepayment resource pack and postpayment. Prepaid resource packs are billed by purchasing a monthly/yearly (natural month/natural year) prepaid resource pack in image search services. After the purchase, the number of calls will be deducted in the resource bundle. When the resource bundle is used up, the default mode is postpaid. At the same time, the pre-paid resource package adopts the tiered payment mode based on the adjustment amount of resource package and the purchase period of resource package. The larger the adjustment amount of resource package is, the longer the purchase period of resource package is (for example, the selection of package year instead of month), the lower the unit price will be. Post-payment is the default mode of post-payment when the pre-paid resource package of image search service is used up. Post-payment mode is pay-per-view.

Author: Cloud Trace Kyushu