On April 22nd, World Earth Day, Baidu and the International Fund for Animal Welfare (IFAW) jointly released “AI Guardian for Endangered Species”, the first-ever platform that uses AI technologies to combat the illegal trade of wildlife products. Based on Baidu's deep learning platform PaddlePaddle, the two sides leverage image recognition models to identify pictures of endangered wildlife products, thus blocking relevant illegal trade online.

With the basic data of amost 100,000 images of animal products provided by IFAW, the team from PaddlePaddle spent 5 months in repeated testing, labeling, and optimization. As the result of the first stage, images of products concerning elephants, tigers, and pangolins have been brought into the scope of the platform with a recognition accuracy rate reaching 75%, providing practical help to IFAW. Up to now, it has helped screen nearly 250,000 pictures, among which 3,348 images of target products were accurately identified, and 807 illegal trade advertisements were traced.

Identification and labeling of tiger products (right)

Identification and labeling of ivory products (right)

Identification and labeling of pangolin products (right)

Despite former practices of AI technology in wildlife identification, the development and accuracy of the system are still troubled by some unfavorable factors, namely, the varied morphology, color, and luster of processed wildlife products as well as the limited amount of available raw data for machine learning.

To make it worse, an increasing number of sellers tend to publish illegal information in implicit ways that traditional methods fail to monitor, such as pictures, short videos, live streaming platforms, etc. A 2019 study of an online platform’s information issuing situation tells that 49% of the 6,396 illegal advertisements have their key products information in pictures instead of texts.

IFAW's past experiences of online research on wildlife crimes show that although a large number of illegal advertisements can be collected, it is time-consuming to recognize the target products, let alone the undesirable accuracy.

To this end, our "AI Guardian" has tailored an image recognition model with the help of PaddleClas, an industrial-level image classification kit, and open-sourced the model through PaddleHub, a pre-trained model application tool. The PaddlePaddle team, together with IFAW, will continue to dedicate to the iterations of the model for better results and broader recognition scope. The image classification kit boasts a rich model library, including 23 series of classification networks and 117 pre-trained models and performance evaluations. The ImageNet 1k pre-trained model, based on PaddleClas and its leading SSLD knowledge distillation scheme, is trained on an IFAW dataset of 100,000 images of wildlife products, and can be easily deployed through Paddle Serving to deliver online prediction services. At the same time, wildlife products recognition will be included in Baidu’s AI Open Platform, providing free services.

Panorama of PaddlePaddle’s vision technology

"Upholding the guideline of 'Technology for better', it is our responsibility to leverage the technology to protect wildlife and the earth. And our PaddlePaddle will keep working on that. The newly developed recognition model will be open-sourced in the PaddlePaddle platform, which serves as an AI-based software platform to support the innovative activities of developers and researchers and help to shape a more intelligent ' AI guardian'," said Wu Tian, the vice president of Baidu.

This year's World Earth Day, themed on "cherish the earth and live in harmony with nature", makes combating the illegal wildlife trade a natural proposal involved. The International Criminal Police Organization (Interpol) once pointed out that such trade, both in scale and profit, should be considered the third-largest illegal trade in the world, second only to drugs and arms.

The birth of the "AI Guardian" comes at the right time, through which IFAW can efficiently probe into the new features of online wildlife crimes while the relevant information could be promptly returned to the relevant platforms for subsequent works, or reported to law enforcement departments for investigation.

IFAW General Representative of Asia Ge Rui said: “We are now in the middle of the pandemic, and on the World Earth Day, the application of AI and other online high technologies, represented by the 'AI Guardian', has empowered us to end the illegal wildlife trade, building up our confidence to eliminate the threats of illegal trade to wildlife. We hope the whole society will take action and make each and every one the guardian of wild animals."

In addition to pictures of elephants, tigers, and pangolins, more capabilities to recognize wild animals that are seriously threatened by illegal trade will be added with usage scenarios covering more forms such as video and live broadcast.

PaddlePaddle also joined hands with the People's Daily client to call on people to "paint for the departed wildlife with AI", allowing the public to create the image of the departed animal in their mind based on news of wildlife poaching with our model. Each painting contributed 10 minutes of Baidu Intelligent Cloud computing power, which is equivalent to 12,000 clues screening.

In addition to the cooperation with IFAW, PaddlePaddle has publicized its technical capability of "AI Guardian", hoping to unite more public welfare organizations to crack down on illegal wildlife trade with technology and maximize its role in protecting wildlife.

Baidu’s Chairman and CEO Robin Li once said, "Do Better, Technology for Better. This is what technology for, and what every member of Baidu pursues with great innovation. " Looking forward, Baidu will continue to utilize AI capabilities to integrate and analyze multiple forms of communication, including texts, images, and videos, as well as the associated information across different network platforms, so as to offer more efficient assistance to wildlife protection and fulfill "technology for better" with practical deeds.

Please refer to the follow documents for more details of PaddlePaddle.

·       Official Website：

https://www.paddlepaddle.org.cn

·       Image Classification Kit PaddleClas：

GitHub:

https://github.com/PaddlePaddle/PaddleClas

·       PaddleHub Pre-trained Models：

https://www.paddlepaddle.org.cn/hubdetail?name=resnet50_vd_wildanimals&en_category=ImageClassification

·       PaddlePaddle Server-side Deployment Library：

https://github.com/PaddlePaddle/Serving

·       PaddlePaddle Open-source Frameworks：

GitHub: https://github.com/PaddlePaddle/Paddle

Gitee:  https://gitee.com/paddlepaddle/Paddle

Official QQ group:  Join the group chat to exchange with like-minded friends in deep learning.