Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog High-Accuracy Population-based Image Search

High-Accuracy Population-based Image Search

In this article, Liu Lei talks about the applications of Image Search for e-commerce, developed by Alibaba's Machine Intelligence Technology Laboratory.

By Alibaba Cloud Data Intelligence Team

Liu Lei from Alibaba's Machine Intelligence Technology Laboratory gave a speech titled "High-Accuracy Population-based Image Search" on the 2018 Apsara technology platform. He introduced Image Search in four areas, including basic concepts, technical architecture and advantages, application scenarios and cases, and commercial use and price.

Alibaba Machine Intelligence Technology Laboratory

Established in 2018, the Machine Intelligence Technology Laboratory comprises of a group of outstanding scientists and engineers, with research centers located in Hangzhou, Beijing, Seattle, Silicon Valley, and Singapore. Machine Intelligence Technology Laboratory is Alibaba's core team responsible for the research and development of artificial intelligence technologies. Relying on Alibaba's valuable massive data and machine learning/deep learning technologies, the lab has developed image recognition, speech interaction, natural language understanding, intelligent decision-making, and other core artificial intelligence technologies. It fully empowers Alibaba Group's important businesses such as e-commerce, finance, logistics, social interaction, and entertainment, and also provides outputs to ecosystem partners to jointly build a smart future.

Image Search

1

Image Search is an intelligent image search product that enables search by image using image recognition and search functions, based on deep learning and large-scale machine learning technologies. It mainly consists of search by same image and similar image. Based on image recognition technologies, Image Search combines the application and business scenarios in different industries to help users research data by same image and similar image. Image Search can be used in a wide range of business areas because it is available for all users with image databases.

Image Search has experienced three main development phases. In August 2014, we launched the image search technology on Taobao Pailitao, allowing users to search commodities by photo for the first time. In February 2017, we started to share the image search technology with our business partners. In 2018, we began to empower the image search technology widely on Alibaba Cloud, and launched this technology on Alibaba Cloud and worldwide in July.

Technical Architecture and Advantages

2

The image search technology mainly involves five algorithm modules:

  1. Category prediction: Distinguishing images by features only is limited. Category prediction determines the category of a target, allowing users to search the images of a certain category instead of the entire image library. Category search improves both search efficiency and accuracy.
  2. Subject detection: An image subject target is generally small and involves complex background. Subject detection reduces the interference of background and other subjects.
  3. Feature extraction: As a key module of Image Search, feature extraction mainly uses two feature dimensions, deep feature and local feature. The deep learning feature extraction is far superior to the traditional feature extraction. Local feature extraction reflects the local uniqueness of image features. In this dimension, we have improved dimension compression, extraction speed, and other aspects.
  4. Retrieval and sorting: The offline indexing module and online quick query module are used for retrieval and indexing. The offline indexing module extracts image features in the offline process. The online quick query module extracts image features based on an input feature image, and then transfers the features to a distributed engine for quick query and retrieval. The sorting module can effectively combine deep features and local features. Deep features focus on deep information of images while local features focus on local information of images.

Technical Process of Image Search

3

In the offline process, the service selects an image from the image library, detects its subject, and then extracts its features to establish an image retrieval engine. In the online process, the service predicts and determines the category of the image you queried, detects its subject, and extracts features. It finally outputs the result after retrieval and sorting.

Image Retrieval Advantages

Deep learning is based on industry-leading deep learning algorithms and massive data, featuring good effect, high accuracy, and strong robustness. We have also independently developed the quick indexing technology that supports ten billions of images and instant response within milliseconds. All products are reliable, stable, and effective as they are verified on Alibaba's e-commerce platforms. Furthermore, we have customized some functions to make Image Search fully available for all business scenarios. For example, we have integrated Alibaba Cloud computing capabilities and MIT visual AI capabilities.

Application Scenario of Image Search

E-commerce Applications

4

Users can simply take or upload photos to search items by image. This saves tedious text descriptions and simplifies the commodity search process, greatly improving users' shopping experience. Sellers can also recommend their products to users more conveniently.

General Image Library

5

Image sharing and social interaction websites often provide plenty of images for users. Image Search quickly builds a cloud search engine with ten billions of images for search by image, improving user experience.

Taobao Product Search

6

An image search technology was deployed in the operating system of Samsung Galaxy S8 to search related commodities on Taobao by photo in the album. This improves user experience in shopping search.

About Image Search

7

Image Search was launched commercially at both Chinese sites and international sites on July 11 and is well received by customers. The two billing methods available at present are prepaid resource package and post-payment. You can purchase a prepaid resource package for Image Search by calendar month/year. After the package is purchased, the call count is deducted from the quota provided in the resource package. After the quota is used up, the billing method is switched to post-payment by default. A tiered discount policy is provided based on the call quota and subscription period in the prepaid resource package. The larger the call quota is or the longer the subscription period (such as yearly payment) is, the lower the unit price gets. After the quota in the existing prepaid resource package is used up, the billing method is switched to post-payment by default, and then the service is billed by the call count.

To learn more about Alibaba Cloud Image Search, visit www.alibabacloud.com/product/imagesearch

Trends Artificial Intelligence Deep Learning Machine Learning Image Search
1 1 1
Share on

Read previous post:

Alibaba Group Tackling Poverty through Technology

Read next post:

Link Vision Opens a New IoT Smart Video Service

Alibaba Clouder

2,599 posts | 764 followers

You may also like

Comments

Raja_KT February 19, 2019 at 6:05 am

In the "Technical Process of Image Search", the online, offline choice is interesting.

Alibaba Clouder

2,599 posts | 764 followers

Related Products

More Posts by Alibaba Clouder

See All