Image Recognition Using Dual-Model Feature Vector Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in accurately identifying objects from low-quality images, such as those with motion blur, out-of-focus blur, or low resolution, due to the inability to correctly extract features, leading to erroneous authentication.
Innovation Solution
An image recognition apparatus that calculates a first feature vector from a high-quality image and a second feature vector from multiple low-quality images using different models, allowing for comparison to determine if the objects in both images are the same, thereby improving authentication accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If low-quality images are used for object recognition, then the system can operate with existing monitoring camera data without requiring additional high-quality image capture infrastructure, but the feature extraction accuracy deteriorates leading to erroneous authentication
Solution Approach 1:
The patent applies parameter changes by transforming low-quality images into high-quality representations through super-resolution reconstruction and synthetic data generation. The system changes the quality parameter of images from low-resolution/blurry to high-resolution/sharp through computational methods, enabling accurate feature extraction while maintaining adaptability to existing monitoring camera data
Solution Approach 2:
The patent introduces an intermediary approach by using synthetic images generated from low-quality monitoring camera data as a bridge. These synthetic high-quality images serve as intermediaries that enable accurate feature extraction and model training without requiring direct capture of high-quality images, thus resolving the contradiction between using existing data and achieving high accuracy
2Reliability
If high-quality images are used for feature extraction, then authentication accuracy is improved, but the system requires additional infrastructure for capturing high-quality images which increases system complexity and cost
Solution Approach 1:
The patent replaces the mechanical system of high-quality image capture infrastructure with a computational approach. Instead of using better cameras or controlled imaging conditions, the system uses algorithms to synthetically generate high-quality images from low-quality monitoring camera data, thereby achieving high authentication accuracy without increasing physical hardware complexity
Solution Approach 2:
The patent creates copies of low-quality monitoring camera images that are synthetically enhanced to high quality. These synthetic copies retain the essential information from the original low-quality images while adding the quality attributes of high-quality images, enabling accurate authentication without requiring the original high-quality capture infrastructure
3Measurement precision
If image resolution is increased through super-resolution, then the detail information is enhanced for better recognition, but information not found in the original image may be added causing erroneous authentication
Solution Approach 1:
The patent implements feedback mechanisms in the form of loss functions during the synthetic image generation process. The system uses ground truth labels and authentication results as feedback to guide the synthetic image generation, ensuring that only accurate and reliable information is synthesized while preventing the addition of spurious details that could lead to erroneous authentication
Solution Approach 2:
The patent performs preliminary actions by pre-processing and validating the low-quality monitoring camera data before generating synthetic high-quality images. This includes quality assessment, selective processing, and verification steps that ensure the synthesized information is reliable before being used for authentication, thus preventing erroneous results
Data Source
AI summary
An image recognition apparatus that identifies an object in an image includes at least one memory storing instructions, and at least one processor that, upon execution of the instructions, operates as a first calculation unit configured to calculate a first feature vector from a first image including the object by using a first model, a second calculation unit configured to calculate a second feature vector from second images by using a second model, wherein a number of the second images is greater than a number of the first image used by the first calculation unit, and wherein a quality of at least one of the second images is lower than the first image, and an identification unit configured to compare the first feature vector and the second feature vector to determine if an object in the second images is a same object as an object in the first image.