Feature Extraction Model for Moving Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for moving image recognition face challenges in achieving sufficient recognition accuracy due to insufficient training data, as they rely heavily on annotated moving image data and lack effective learning methods when the amount of annotated data is limited.
Innovation Solution
A learning device and method that extracts features from both moving image and still image data using a feature extraction unit, with separate inference units for each type of data, allowing for learning based on the results of both types of inference to improve feature extraction for moving image recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data is insufficient, then the burden on generation of training data is large, but the recognition accuracy will not become sufficient accuracy
Solution Approach 1:
The feature extraction unit is designed to process both moving image data and still image data using the same parameters. By making the feature extraction unit multi-functional, the system can utilize still image data (which is easier and cheaper to obtain) to train the same feature extraction parameters that will be used for moving image recognition, thereby reducing the burden of generating large amounts of annotated moving image training data while maintaining sufficient recognition accuracy.
2Measurement precision
If annotated moving image data is limited, then it is difficult to prepare sufficient training data, but recognition accuracy cannot be achieved
Solution Approach 1:
Still image data serves as an intermediary resource in the training process. Instead of directly relying on limited annotated moving image data, the system uses still image data as a mediator to train the feature extraction unit. This intermediary approach allows the system to leverage the abundance of easily obtainable still image data to achieve effective training, thereby improving the ease of preparing training data while maintaining recognition accuracy.
3Measurement precision
If only moving image data is used for learning, then the learning process is simple, but the recognition accuracy is insufficient due to data limitations
Solution Approach 1:
The learning process merges the training of still image data and moving image data into a unified process. The feature extraction unit learns parameters from both types of data simultaneously, combining the benefits of abundant still image data with the temporal information in moving image data. This merging approach enhances recognition accuracy by leveraging diverse data sources while maintaining a relatively simple learning framework that processes both data types through the same feature extraction unit.
Data Source
AI summary
The learning device 10D is learned to extract moving image feature amount Fm which is feature amount relating to the moving image data Dm when the moving image data Dm is inputted thereto, and is learned to extract still image feature amount Fs which is feature amount relating to the still image data Ds when the still image data Ds is inputted thereto. The first inference unit 32D performs a first inference regarding the moving image data Dm based on the moving image feature amount Fm. The second inference unit 34D performs a second inference regarding the still image data Ds based on the still image feature amount Fs. The learning unit 36D performs learning of the feature extraction unit 31D based on the results of the first inference and the second inference.