Online Distillation Using Frame Cache for Image Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing image analysis apparatuses using lightweight models for moving image data analysis face limitations in improving analysis accuracy compared to offline learning, as they rely solely on online learning with newly inputted frames without leveraging previously analyzed data effectively.
Innovation Solution
An image analysis apparatus that utilizes a combination of a high-accuracy and lightweight image analysis model, where the lightweight model is updated using both newly obtained frames and cached analyzed frames, with evaluation-based extraction and weighting of update data to enhance learning efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If online learning is performed using only newly inputted frames, then the model can be updated in real-time, but the analysis accuracy improvement is limited compared to offline learning
Solution Approach 1:
The system pre-processes and stores analysis results of previously inputted frames in a storage unit before they are needed for model updates. This preliminary action allows the model to access high-quality training data from past frames without re-processing them, thereby improving analysis accuracy while maintaining real-time update capability.
Solution Approach 2:
A storage unit acts as an intermediary between the inputted frames and the image analysis model. It stores analysis results of past frames and selectively provides them to the model during online learning, enabling the model to learn from both new and historical data without direct re-processing of original frames.
2Quantity of substance
If all analyzed frames are used for model updates, then more training data is available, but the update process becomes less efficient
Solution Approach 1:
The system extracts only the necessary and useful analysis results from the storage unit based on specific conditions (e.g., frames where the lightweight model's analysis differed from the expensive model's analysis). This selective extraction provides sufficient training data for effective model updates while avoiding the computational overhead of processing all available frames.
Solution Approach 2:
Instead of using all available training data, the system employs partial action by selectively choosing a subset of frames that are most beneficial for model updates. This approach achieves effective learning with reduced computational resources and improved update efficiency.
3Speed
If a lightweight image analysis model is used, then processing speed is improved, but analysis accuracy decreases
Solution Approach 1:
The system implements feedback by comparing the analysis results of the lightweight model with those of a more expensive high-accuracy model. The differences between these results are used to generate update data that refines the lightweight model, enabling it to progressively improve accuracy while maintaining its speed advantage.
Solution Approach 2:
The system changes the parameters of the lightweight model through iterative updates based on feedback from both new frames and cached analyzed frames. By adjusting model parameters using carefully selected training data, the lightweight model achieves improved accuracy without sacrificing its processing speed.
Data Source
AI summary
The image analysis apparatus includes: a first analysis unit that analyzes image frames by using a first image analysis model; a second analysis unit that can analyze the image frames by using a second image analysis model an analysis accuracy of which is lower; a storage unit that stores therein analyzed frames, which are already analyzed by using the first and second image analysis models, in association with an evaluation value for evaluating a result of an analysis performed with the second image analysis model; an extraction unit that extracts an analyzed frame that satisfies an extraction condition based on the evaluation value; and an update unit that updates the second image analysis model by using a result of an analysis performed with the first image analysis model on the extracted analyzed frame, and a result of an analysis performed with the first image analysis model on a new frame.