Image Recognition Discriminator Adaptation via Incremental Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face accuracy issues when the image capturing apparatus's environment changes, such as direction, position, or angle, leading to decreased recognition performance due to differences between training and captured image tendencies.
Innovation Solution
An image processing apparatus that generates additional training data from captured images and their recognition results, using incremental training to update a discriminator for improved recognition accuracy, allowing the system to adapt to changing environments without user-inputted correct recognition results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a recognition model is trained using supervisory data prepared in advance, then image recognition can be performed using the obtained recognition model, but recognition accuracy decreases when the image capturing apparatus's environment changes
Solution Approach 1:
The recognition model is updated dynamically through incremental training using newly captured images. The system continuously adapts the discriminator by incorporating new training data generated from captured images, allowing the model to evolve and maintain accuracy across changing environmental conditions rather than remaining static
Solution Approach 2:
Training data is generated in advance from captured images by automatically creating label information through the recognition result generation unit. This preliminary preparation of training data with automatic labeling enables the incremental training process to proceed efficiently without requiring manual annotation before deployment
2Measurement precision
If additional training data is generated from captured images, then recognition accuracy on specific scenes is improved, but system complexity increases
Solution Approach 1:
The system generates its own training data autonomously by using the recognition model to create label information for captured images. The recognition result generation unit automatically produces ground truth labels from the captured images themselves, enabling the system to self-improve without external intervention or manual annotation processes
Solution Approach 2:
The system combines the recognition model with the training data generation capability into an integrated framework. The discriminator and the training data generation process work together as a unified system where the recognition results directly feed into training data creation, eliminating the need for separate manual annotation systems
3Measurement precision
If manual annotation is required for training data, then training data quality is high, but user burden increases and productivity decreases
Solution Approach 1:
The system automatically generates label information for training data using its own recognition model. The recognition result generation unit processes captured images and produces ground truth labels without requiring manual user annotation, enabling autonomous training data preparation that scales efficiently
Solution Approach 2:
The system uses the recognition results from captured images as feedback to generate improved training data. The label information generated from recognition results feeds back into the incremental training process, creating a continuous improvement loop that enhances training data quality without manual intervention
Data Source
AI summary
There is provided with an image processing apparatus, for example for image recognition such as object counting with machine learning. A generation unit, based on a first captured image, generates a first training data that indicates a first training image and an image recognition result for the first training image. A training unit, by performing training using the first training data, generates a discriminator for image recognition based on both the first training data and second training data that is prepared in advance and that indicates a second training image and an image recognition result for the second training image.