Low-Resolution Face Recognition via Feature Adaptation Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based face recognition technologies face challenges in accurately recognizing low-resolution face images due to a mismatch in information amounts between high-resolution and low-resolution images, leading to deteriorated recognition performance, especially in real-world environments like CCTV systems.
Innovation Solution
A low-resolution face recognition device and method that generates actual low-resolution face images from high-resolution images using a combination of high-resolution and low-resolution face images, and employs a feature adaptation network to reduce the feature gap between domains, allowing for improved recognition performance by learning common features in a middle domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning-based face recognition model is learned using high-resolution face images, then the model can extract distinctive features based on texture information, but the recognition performance deteriorates when applied to low-resolution images due to information mismatch
Solution Approach 1:
The patent introduces a low-resolution image generator as an intermediary component that synthesizes low-resolution face images from high-resolution images. This generator creates artificial low-resolution images that mimic real low-resolution image characteristics (such as compression artifacts and noise), enabling the model to learn how to effectively process low-resolution inputs without actually training on real low-resolution data, thus resolving the information mismatch problem
2Area of stationary object
If image interpolation method is used to increase low-resolution image size, then the image dimensions are enlarged, but the amount of information cannot be increased and recognition performance cannot be improved
Solution Approach 1:
The patent uses a low-resolution image generator to create synthetic low-resolution images that copy the essential characteristics of real low-resolution images (including degradation patterns, noise, and compression artifacts) from high-resolution source images. This copying approach allows the model to learn from synthetic images that preserve the information content of real low-resolution images, overcoming the limitation of traditional interpolation methods that merely enlarge pixels without adding information
3Measurement precision
If super-resolution imaging technique is used to generate high-resolution images from low-resolution images, then high-resolution images can be obtained, but a paired high-resolution and low-resolution image does not exist for learning, so learning is performed on interpolated low-resolution images which cannot model actual environment
Solution Approach 1:
Instead of using super-resolution techniques to go from low-resolution to high-resolution (the conventional approach), the patent inverts the approach by using a low-resolution image generator to synthesize low-resolution images from high-resolution images. This inversion allows the model to learn the transformation process and adapt to low-resolution inputs by understanding how high-resolution images degrade to low-resolution ones, enabling better modeling of actual environmental conditions
Data Source
AI summary
The present disclosure relates to a low-resolution face recognition device, which includes a high-resolution face image inputter; a low-resolution face image inputter; a high-resolution face feature extractor configured to extract a high-resolution face feature by using high-resolution and low-resolution face images; a face quality feature extractor configured to extract face quality features by using the high-resolution and low-resolution face images; a feature combiner configured to detect the high-resolution and low-resolution face features by concatenating the high-resolution face feature and the face quality feature; a feature adaptation network configured to extract a high-resolution face feature map and a low-resolution face feature map by using the detected high-resolution and low-resolution face features, respectively; and a consistency meter configured to determine a face ID by measuring consistency of a face feature map by using the extracted high-resolution and low-resolution face feature maps.