Investigating response time and accuracy in online classifier learning for multimedia publish-subscribe systems

Area of Science:

• Computer Science
• Artificial Intelligence
• Multimedia Systems

Background:

• The Internet of Multimedia Things (IoMT) generates vast multimedia data, posing real-time processing challenges for traditional event-based systems.
• Existing machine learning models require optimization for the low-latency demands of multimedia event processing.

Purpose of the Study:

• To develop an online classifier construction approach for efficient multimedia event processing.
• To enable systems to handle dynamic user subscriptions with low response times and acceptable accuracy.

Main Methods:

• Utilized current object detection methods, dynamically configuring hyperparameters for real-time classifier construction.
• Employed deep neural network-based object detection models for training online classifiers from scratch.

Main Results:

• Demonstrated that hyperparameter tuning of object detection models significantly improves performance.
• Achieved 79.00% accuracy with 15 minutes of training and 84.28% accuracy with 1 hour of training on a single GPU.
• Showcased the model's ability to answer previously unknown user subscriptions effectively.

Conclusions:

• The proposed online classifier training method is effective for real-time multimedia event processing in IoMT environments.
• Dynamic configuration and hyperparameter tuning of object detection models offer a viable solution for low-latency multimedia data analysis.
• This approach balances accuracy and training efficiency for handling evolving user demands in multimedia streams.