Object Recognition via Image Spectrum Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current food recognition systems face challenges in accurately identifying food in real-world environments due to the variability in visual appearance, as they often rely on feature-based approaches that struggle with invariant features across different food placements.
Innovation Solution
A system that combines image and spectrum features using an image/spectrum sensing device, data banks for image and spectrum features, fetching and analyzing modules, and a fusion module to improve object recognition precision by leveraging sparse codes and support vector machines for candidate object identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature-based approach is used for food recognition, then recognition works well in constrained environment, but it struggles to find invariant features robust to distinct visual appearances from different food placement in real-world environments
Solution Approach 1:
The patent combines image features and spectrum features into a fused feature representation. The image features capture visual appearance while spectrum features capture material composition, and their fusion creates a more robust and versatile recognition system that works reliably across different food placements and environments.
Solution Approach 2:
The patent transforms the feature extraction approach by changing from traditional hand-crafted features to deep learning-based features, and further enhances it by adding spectral parameters. This parameter expansion from purely visual to visual-spectral space enables the system to find invariant features that are robust to variations in food placement, lighting, and appearance.
2Measurement precision
If patch-based visual appearance is used directly instead of feature-based approach, then visual-based food recognition can be improved, but the system still faces challenges with variety of food appearance even though images captured from the same food class
Solution Approach 1:
The patent merges image appearance information with spectrum information to create a comprehensive feature representation. This combination allows the system to distinguish between different food classes more precisely while remaining adaptable to appearance variations, as the spectral component provides material-level discrimination that is invariant to visual appearance changes.
3Measurement precision
If only image-based recognition is used, then the system is simpler, but the precision to recognize the real object can be improved by combining image and spectrum
Solution Approach 1:
The patent integrates image sensing and spectrum sensing into a unified recognition system. The image/spectrum sensing device captures both visual and spectral information simultaneously, and the fusion module combines these complementary data sources to achieve higher recognition precision while managing system complexity through integrated architecture.
Solution Approach 2:
The patent creates a multi-functional sensing system that performs both image capture and spectral analysis through an integrated device. This universal system handles multiple types of information (visual and material) with a single platform, improving precision without proportionally increasing complexity.
Data Source
AI summary
A system for object recognition includes an image/spectrum sensing device, to fetch an object image from a real object and sense spectra at sensing regions of the real object. A fetching module obtains a real-object image feature pattern for each ROI of the object image. An analyzing module for object image feature searches for a first candidate object from data bank and analyzes a correlation between the real object and candidate object. A fetching module for object spectrum feature obtains a real-object spectrum pattern for each ROI of the object image. An analyzing module for object spectrum feature is to search for a second candidate object from data bank and to further analyze a match level between the real-object and the second candidate object. A fusion module further analyzes information of image feature and spectrum feature find whether or not having a matched object as an identify object.