Distributed visual image recognition method based on virtual environment

Abstract

The invention discloses a distributed visual image recognition method based on a virtualized environment, which comprises the following steps: S1, preprocessing visual image data; S2, recognizing thetarget visual image data, and performing image classification and target detection; S3, deploying multi-server nodes and multi-identification containers and managing GPU resource allocation through internal web; S4: Making mirror image corresponding to image classification function and target detection function classification, and customizing resource allocation, mirror image loading, container instantiation and container internal service self-startup function of the system. By integrating a plurality of depth learning frameworks, fast instantiation of the capabilities of multiple in-depth learning framework models in a platform environment is achieved, The GPU resources of multi-server nodes are managed and deployed by virtualization of platform mechanism, and each functional module in the project is instantiated as an independent mirror container through virtualization scheme, so as to rapidly deploy the project and establish the cluster function on single or multiple computers.

Description

technical field [0001] The present invention relates to the technical field of virtualization, in particular to a distributed visual image recognition method based on a virtualized environment. Background technique [0002] With the popularization of mobile Internet and 4G network, the media information that people come into contact with every day is becoming more and more complicated. Through the mobile Internet, even many people can become producers of media information. Therefore, the security supervision of new media content has become an urgent task. For very large-scale Internet video image data, manual review can no longer meet the demand. With the rapid development of GPU computing and deep neural network for image recognition technology, the solution of visual image recognition through machine learning training has shown more and more important practical value in many fields. [0003] Several current mainstream deep learning frameworks, such as ensorflow, mxnet, an...

AI Technical Summary

Problems solved by technology

[0004] In the absence of virtualization, if you need to instantiate the models of preprocessing, caffe, mxnet, tensorflow and other frameworks on the same server, you must install all the environment dependencies of these frameworks in sequence on this server Well, complex environment dependency conflicts need to be resolved. If you want to move the environment of the current server to another server, you must re-install all the environment dependencies. This will be a very heavy work in the case of large-scale cluster servers. ;

[0005] In the process of project engineering, in order to be compatible with multiple deep learning frameworks in one platform, fast model training and instantiation of model capabilities have become very difficult problems, and there is no mature solution for GPU resource virtualization, so There is an urgent need for a unified platform specification to resolve multi-framework environment conflicts, make model training and instantiation of model capabilities easier, and manage and schedule the allocation and use of GPU resources for each model on multiple server nodes

Images