After 7 months, it was co-sponsored by Qualcomm, China Zhigu·Chongqing Economic Development Zone, CSDN, Testin Cloud Test, OPPO, Extreme View, Zhongke Chuangda, and Chuangyebang, Chongqing Economic Development Zone·Qualcomm China· Co-organized by Zhongke Chuangda Joint Innovation Center, the industry's professional "Qualcomm Artificial Intelligence Application Innovation Competition" with TensorFlow Lite as an open source technology partner ended successfully!
After two rounds of PK in the rigorous preliminary and final competitions, the judges finally selected the winning works from the 110 outstanding works submitted by the contestants. 9 participating teams/individuals won the Platinum Award, Business Model Innovation Award, and Chongqing Innovation Center Special Award, AI Visual Computing Innovation Application Award, Algorithm Innovation Award, AI Application Scenario Innovation Award, Terminal Side Artificial Intelligence Innovation Award, 5G Edge Application Technology Award and AIOT DSP Application Innovation Award!
Who are the best people who won this year's awards? What are the unique features of the winning works? Not much to say, let’s announce them one by one below.
Who are the major awards?
Platinum Award
Winner: Zhao Tianqi
Entry name: Artificial Intelligence Animation System: Cyber incarnation Cybatar
Team members and division of labor: Zhao Tianqi, founder CTO of JuliDimension, project direction and scientific research; Wang Yuan, software development team leader of JuliDimension, project implementation.
Introduction of works:
Cyber avatar Cybatar is an artificial intelligence animation production application, including real-time facial capture, intelligent motion capture, intelligent face capture, intelligent generation, intelligent sound capture and other digital human technologies, so that everyone can make a few minutes of professional animation every day works. The current version has several built-in CG characters, scene props, and synthesis templates of its own or cooperative IP. The production content can be used according to different settings and permissions, and the material library is diverse. Only ordinary cameras can accurately capture the user's facial expressions. The changes of emotions, anger, sorrow, and even various micro-expressions can be accurately displayed. It can also make precise facial adjustments for different controllers and usage scenarios, such as eye size, neck height, etc.
Future achievable functions of this application include 3D scene conversion, multi-person and multi-source control system, game time-sharing combination control mechanism, network synchronization update mechanism, network direct recording and broadcasting cooperation mechanism, artificial intelligence face-squeezing system, dress-up system and action Capture, gesture capture, voice capture and other AI algorithms.
Show results:
Juli Dimension has developed a new surface capture network structure model, which overcomes the shortcomings of traditional surface capture algorithms that are limited by modeling accuracy, shooting angles and ambient lighting. Models can be divided into general ID extraction model, expression extraction model, expression optimization model, pose extraction model and spatio-temporal fusion model.
Through this sub-task network and multi-level structure, accurate extraction and conversion of expressions and poses are achieved, and at the same time, the consistency and real-time performance of the model during face-catching and head-catching are guaranteed.
Then, this model was successfully transplanted to an Android phone equipped with Qualcomm Snapdragon 865, and combined with the UE4 three-dimensional engine to realize the real-time facial expression reproduction of the virtual digital person. Finally, the scene is optimized for the mobile platform, including lighting, materials, models, etc., so that the operating efficiency exceeds 24fps.
Business Model Innovation Award
Winner: Yuan Bo and Youxuemao R&D team
Entry name: Youxuemao AI system
Team members and division of labor: Yuan Bo Youxuemao Technical Partner; Sun Bo, Youxuemao CEO; Pan Zhigang Product Operation.
Introduction of works and development process:
Youxuemao children’s AI interactive learning platform contains multi-disciplinary enlightenment AI smart hardware and AI courses for preschool children, connects various smart screens, embeds AI chips in traditional play aids and upgrades to smart teaching aids, and uses a variety of educational scenarios and AI interactive technology to stimulate children Interest, using visual, auditory, tactile and other multi-sensory stimuli to perceive, help children explore and learn, "teach them to fish instead of fish."
The AI technology that Youxuemao mainly involves:
The first is voice recognition, which is mainly used in the Youxuemao English APP. Users will follow the audio of the game and the live recording teacher, and the system will judge the follow-up results and give feedback. Specifically, it performs preprocessing of the collected sounds, and uses the language model and probability statistics of CMUSphinx for speech recognition.
The second is to use human body key point recognition and human body semantic segmentation to complete the character matting of complex scenes, and to improve the part of the CocosNet network to complete the human body posture redirection, placing children in the APP scene to achieve interaction and animation; also used The traditional CV method has completed the edge detection and filtering functions of some objects, so that children can make DIY creations by themselves.
Refer to CocosNet to realize human posture migration across domains:
Based on the key points of the human body and the Spine skeleton animation to achieve the interactive fusion of task scenes, and at the same time realize the binding of Mesh weights through human body semantic segmentation, and finally act in the APP scene, allowing children to see themselves entering the classroom virtual.
The third Youxuemao course uses video automatic slicing technology according to different error feedback operations to calculate seamlessly connected videos of different feedbacks, simulating the live broadcast process of a real teacher.
The Fourth Youxuemao implements AI teaching aids for intelligent training, recording children's learning data intelligence, analyzing and evaluating, intelligently commenting and other different functions.
Chongqing Innovation Center Special Award
Winner: Jiang Qi and the rhythm team
Entry name: Rhythm (cardiopulmonary disease auxiliary diagnosis terminal and service)
Team members and division of labor: Jiang Qi, Davidson Capeland consultant, Panda car CIO, responsible for overall product and back-end implementation; Wang Shoukun, former Douban chief scientist, who provided algorithm tuning guidance as a consultant; Yu Zihao, responsible for front-end interaction.
Introduction of works and development process:
Rhythm provides assisted diagnosis of cardiopulmonary diseases, terminals and services positioned in the home pre-diagnosis market. In the past two decades, although there have been achievements in the field of electronic auscultation and assisted cardiopulmonary diagnosis, there are many dilemmas to be overcome, such as lack of databases, expensive sampling terminals, slow iteration of feedback systems, and small scale of mass production. Rhythm collects heart and lung sound data through a cost-effective electronic stethoscope, and then uses artificial intelligence algorithms to compare the types of potential diseases to solve the above problems. Rhythm achieves extreme cost control, ease of use, and accuracy. The hardware solution is controlled within 200 yuan, the sampling scale of a single case is 10,000 cases, and the commercialization achieves 100 types of cardiopulmonary related cases coverage, and a rapid feedback system is established to optimize diagnosis model.
This project fills the gap in the domestic database of heart and lung sound diseases. It only takes a few simple steps to insert the electronic stethoscope into the earphone jack of the mobile phone in a quiet environment to collect heart sound or lung sound data, and use Tensorflow Lite to speculate the type of disease. At the current stage, the recognition accuracy rate has reached 82%. With the addition of samples, the width and accuracy of the identified diseases will continue to improve.
AI Visual Computing Innovation Application Award
Winner: Wang Shiguang
Team members and division of labor: Individual
Entry name: mobile-based crop pest identification system
Introduction of works and development process:
"Agriculture, rural areas and farmers" is a social issue that our country pays close attention to. The identification of crop diseases and insect pests based on image processing technology has the characteristics of fast, accurate, and real-time, which can assist farmers to take effective prevention and control measures in time. The system uses large-scale professional image data to train the deep learning model, and now supports the recognition of nearly 15 crops and 80 common pests and diseases. The recognition accuracy is greater than 90%, and the recognition time for a single image is less than 200ms.
Using this system, farmers only need to download an App, take out their mobile phones and take pictures to delineate the area to be identified, and the system can automatically identify the disease status and types of crops.
The picture above is an overview of the crop pest identification project system based on the mobile terminal. First, the server uses large-scale professional image data to train the deep learning model, and uses AutoML and knowledge transfer technologies to search for higher-precision models under the limitation of limited time delay, and then deploy the trained models to the device On the OPPO Ace2 mobile phone of Qualcomm Snapdragon 865 5G mobile platform, Qualcomm's fifth-generation AIEngine was used to accelerate the performance of the entire model deployment, and the prototype of the entire project was obtained. Finally, use cloud big data to provide accurate and personalized comprehensive solutions for each farmer.
During the development process, the main problem the author encountered was how to train a model with higher accuracy and generalization with less data, and the need to prepare an appropriate environment. Because the author himself has no experience in App design and development, he is not proficient in Android development and needs to learn while developing. Although the process is a bit bumpy, the problems are finally solved one by one. Thanks to friends who gave help during the project.
Algorithm Innovation Award
Winner: Yan Qing
Entry name: Image enhancement based on color compensation
Team members and division of labor: Individual
Introduction of works and development process:
In recent years, although relatively good results have been achieved in the field of underwater image processing, there are still many pain points. For this reason, this project proposes a low-cost and good restoration method for underwater image color restoration. It is expected to correct the color cast of underwater images while removing the interference of underwater imaging backscattering.
The main steps to achieve color restoration and definition restoration of underwater images are as follows:
The original image uses WCID to estimate the global attenuation rate and local attenuation rate of light in different bands of underwater images
Perform color compensation for the more attenuated color channels according to the attenuation rate calculated in the first step
White balance the compensated image to remove the negative effects of over-compensation
According to the blue and green channels of the image, the transmittance of the water image is estimated by defogging
The atmospheric scattering model is used to remove the backscattering of the water image to obtain a clear image
Experimental results show:
AI Application Scenario Innovation Award
Winner: Xie Yongming (Hong Kong Optical Cloud Technology Team)
Entry name: TrueToF—Active AI ToF imaging and application
Team members and division of labor: Xie Yongming, founder of Hong Kong Optics Cloud Technology, project coordinator; Zhou Jingwei, 3D imaging algorithm research and development; Li Jianying, project management and implementation; Wang Xinghe, product design.
Introduction of works and development process:
With the development of ToF products, coupled with ToF output depth information, there will be no color image information output, ensuring personal privacy. Based on this consideration, Hong Kong Optical Cloud Technology has launched the TrueToF portrait application. The three-dimensional image information provided by TrueToF is very conducive to image segmentation and 3D recognition, and can quickly locate the area of interest, thereby reducing the data processing complexity of the later AI algorithm. Simplify the user's face ID registration process and improve the accuracy of the 3D face recognition system.
TrueToF ToF imaging AI optimization
Compared with traditional ToF imaging, TrueToF uses machine learning and machine learning of the environment to correctly identify and improve the ToF imaging accuracy and accuracy of black objects and bright light environments, and perform dynamic imaging to ensure the power saving requirements of mobile devices .
TrueToF ToF portrait cutout
Based on the fast face recognition machine learning model, it can achieve millisecond-level face detection and portrait extraction, while ensuring the accuracy of 3D portraits and objects on the edges of complex backgrounds.
TrueToF ToF face recognition
The system can extract more than 400 3D feature points at one time as feature comparison points, and is accelerated by Qualcomm's fifth-generation AI Engine, with millisecond-level acquisition, which effectively simplifies the user's face ID registration process, and through millisecond-level face recognition The module can quickly locate human faces based on 3D features and perform real-time face tracking and recognition.
TrueToF ToF emoticon mapping
Through the real-time matching of more than 400 feature points, real-time facial information can be obtained, which ensures the "any angle" tracking of the face. By using Qualcomm's fifth-generation AI Engine acceleration, real-time facial expressions can be obtained, which can be used for 3D expressions Real-time mapping, etc.
Terminal-side artificial intelligence innovation award
Winner: Cong Xiaofeng
Entry name: Android-based photo2cartoon application
Team members and division of labor: Team leader: Cong Xiaofeng (second graduate student), responsible for Android studio development of APP, using deep learning framework for model training and end-to-side deployment.
Introduction of works and development process:
Currently, mobile apps have become an important entrance for children to learn. This project aims to improve the fun of children's drawing learning, using deep learning image-to-image translation technology to convert the "sticky strokes" drawn by users into "real images".
The App provides four panels, which are used to record the images generated at different stages during the user's sketching process. The user can assign the generated image to any drawing board by selecting the display label (1,2,3,4), as follows:
The system uses the image2image conversion method generally recognized by the current academic circles, and performs model design according to the real-time requirements of the end-side. Use Conv/DeConv + InstanceNorm + ReLU as the basic module to design the encoding and decoding network (Conv means convolution, Deconv means deconvolution). The model adopts the idea of generative adversarial network in the field of deep learning, and conducts joint training on the generator and discriminator to ensure the synchronization and optimization of the generative network and the discriminant network.
As the core network deployed on the mobile terminal, the generator network is responsible for converting user sketches into real images. The generator is composed of multiple convolution modules and deconvolution modules. The overall structure is an encoding-decoding idea, as shown in the figure below:
5G Edge Application Technology Award
Winner: Chen Peng
Entry name: AI Smart Pictorial
Team members and division of labor: Chen Pengbot Intelligent VP World Top 100 IT Enterprise Senior R&D Engineer + background algorithm and business logic development; Zhang Jing has management experience in the design department, good at supporting enterprise CIS system design service VIS visual design, corporate brand visual design, enterprise Publicity related design, etc.; Xu Yuan senior marketing expert, creative and communication, product and interaction design.
Introduction of works:
The AI Smart Illustrated Book is based on the DiTing AI Smart Platform, which uses a fully-pipelined deep learning algorithm, a GPU/CPU-core server cluster parallel distributed architecture, through a large number of model training, the use of algorithms to achieve automatic filtering of illegal image content, including Multi-dimensional and multi-categorized content review and filtering mechanisms such as pornography, vulgarity, politics, violence and terrorism, and prohibited advertisements, as well as the identification of scenes, characters, actions, objects and other scene tags.
This system is divided into two main parts: the front-end business system and the back-end algorithm service. It is designed by using the microservice architecture design idea and combining the open source spring framework.
The image content review in this system is implemented based on a deep neural network model, using pixel set classification and counting and network regeneration technology for image content review. In addition, the system adopts a distributed architecture deployment. The latest version of the image model training set sample library has reached 250,000. The current average recognition accuracy of the image algorithm is 92%, especially the recognition rate of pornographic images. The recognition accuracy of the model is 95.6%.
Product workflow display: initial interface, upload review, and result display from left to right
AIOT DSP Application Innovation Award
Winners: Bai Wenbo, etc. (Jiangxi Wo Visual Development Co., Ltd.)
Entry name: AI smart rearview mirror
Team members and division of labor: team
Introduction of works and development process:
The industry's first smart rearview mirror product using Qualcomm Snapdragon 8 cores. The overall solution is based on Qualcomm MSM8953 chip, which integrates software and hardware resources such as positioning, cameras, Android systems, algorithms, and vehicle applications. Local support for multi-channel video display and recording, front and rear ADAS early warning, DMS early warning, intelligent navigation, voice interaction, driving track playback, Bluetooth phone, rear view video, multimedia audio and video entertainment (QQ music and listening partner) and other functions; online support for video Live broadcast, driving track, file retrieval, electronic fence, navigation and search, WeChat interconnection, message reminder, voice intercom, online upgrade and other functions.
System frame diagram
It is worth noting that the above winners are the best among the many entries. A group of outstanding talents and AI solutions emerged during the competition, involving industry, agriculture, transportation, e-commerce, games, charity, and smart home. , Health, civil service and other fields, adding new blood to AI application innovation!
Meet at the award ceremony of "Qualcomm Artificial Intelligence Application Innovation Competition"
The final winners will be awarded at the "Qualcomm Artificial Intelligence Application Innovation Contest" awards ceremony held online at the end of September. There will be awards from Qualcomm, Chongqing Economic Development Zone, Extreme View, Zhongke Chuangda, Tsinghua University, OPPO Senior executives and technical directors of companies such as Google TensorFlow and Google will also attend the event to share their unique insights into the AI field and their prospects for the future.
According to the official website of the conference, this competition prepared a Lynk & Co 05 SUV equipped with Qualcomm Snapdragon 820A automotive platform (only the price of the naked car) for the only platinum award winner, and was named the winner of the track gold award Each will also receive prizes worth 30,000 yuan.
Innovation never stops. The Qualcomm Artificial Intelligence Application Innovation Competition continues to bring more talents and practical solutions to the industry applications. I look forward to the awards ceremony to bring you more surprises!
Click "Read the original text" and follow the live broadcast of the award ceremony!