Simultaneous Feature Extractor and Generator Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training feature extractors and generators to restore ideal images from non-ideal images are unfeasible as they require a pre-trained feature extractor, preventing simultaneous optimization of both components.
Innovation Solution
An image processing apparatus that acquires both ideal and non-ideal images, extracts feature amounts using a feature extractor, classifies objects, generates restored images, and trains both the extractor and generator simultaneously to minimize differences in classification results and restored image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a pre-trained feature extractor is used to train the generator, then the generator can be trained, but the feature extractor and generator cannot be simultaneously optimized
Solution Approach 1:
The patent merges the training processes of the feature extractor and generator into a single unified training framework. The loss function combines classification loss (from the feature extractor) and restoration loss (from the generator), allowing both models to be trained simultaneously and jointly optimized rather than sequentially
Solution Approach 2:
The patent creates a universal training system where the same infrastructure (processor, memory, training algorithm) can simultaneously optimize both the feature extractor and generator. The system universally handles both model training tasks through a single integrated approach, eliminating the need for separate training pipelines
2Measurement precision
If only ideal images are used for training, then face recognition accuracy is high, but the system cannot handle non-ideal images with occlusion or poor imaging conditions
Solution Approach 1:
The patent converts the harmful effect of non-ideal images (occlusion, poor lighting, profile views) into a beneficial training opportunity. By using non-ideal images as input to the generator and comparing the restored images with ground truth ideal images, the system learns to handle challenging imaging conditions while maintaining high recognition accuracy
Solution Approach 2:
The patent performs preliminary restoration of non-ideal images to ideal images before using them for training. This preliminary action by the generator ensures that the feature extractor receives cleaned, idealized images during training, allowing the system to pre-compensate for imaging conditions and maintain high accuracy across diverse scenarios
Data Source
AI summary
An image processing method includes acquiring a first image that represents an object, a second image that represents the object and is different from the first image in imaging condition, and identification information for identifying the object; extracting a feature amount from the second image by using an extractor; classifying the object by using the feature amount and a classifier; restoring by generating a restored image close to the first image from the second image, using the feature amount and a generator; and training the extractor and the generator such that values corresponding to a first difference between a result of classification of the object and the identification information and a second difference between the restored image and the first image become small.