Hierarchical Super-Class Grouping for Road Sign Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition systems, particularly in advanced driver assistance systems (ADAS), face challenges in accurately recognizing multiple road signs in real-time due to intra-class similarity, occlusions, blur, inappropriate lighting, and geometrical variations, leading to decreased accuracy and throughput, especially when biased towards classes with abundant training data.
Innovation Solution
A method and system that group similar image classes based on feature correlation analysis, forming super-classes and subsequently performing sub-class classification, which reduces the number of classes to be recognized and mitigates bias towards classes with extensive training data, using deep learning techniques and correlation matrix analysis to determine and apply a threshold for grouping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If single shot image recognition techniques are used for multi-class recognition, then real-time processing is achieved, but accuracy and throughput decrease with increase in number of road sign classes
Solution Approach 1:
The patent segments the multi-class recognition problem into two stages: first recognizing super-classes (coarse-grained categories) and then recognizing sub-classes (fine-grained categories). This hierarchical segmentation reduces the complexity of single-shot multi-class recognition by breaking it down into manageable steps, thereby maintaining real-time processing while improving accuracy.
Solution Approach 2:
The patent introduces a hierarchical dimension to the recognition process by organizing classes into super-classes and sub-classes. This dimensional transformation allows the system to process images through multiple levels of abstraction, improving recognition accuracy without sacrificing throughput by performing computations in a structured, multi-stage manner.
2Measurement precision
If image recognition systems rely on abundant training data for particular classes, then recognition performance improves for those classes, but the system becomes biased towards those classes
Solution Approach 1:
The patent segments the class hierarchy into super-classes that aggregate multiple sub-classes. By training recognition models at the super-class level where data distribution is more balanced, the system reduces bias towards classes with abundant training data. The hierarchical structure allows under-represented sub-classes to benefit from the balanced super-class recognition.
Solution Approach 2:
The patent transforms the flat class structure into a hierarchical dimension with super-classes and sub-classes. This dimensional change enables the system to operate at multiple levels of granularity, allowing balanced recognition at the super-class level while maintaining the ability to distinguish between sub-classes, thereby reducing bias from imbalanced training data distribution.
3Adaptability or versatility
If the number of image classes to be recognized is increased, then comprehensive road sign recognition is achieved, but processing time and computational complexity increase
Solution Approach 1:
The patent segments the large number of classes into hierarchical groups (super-classes containing multiple sub-classes). This segmentation reduces the effective number of classes that need to be processed in each recognition stage, thereby reducing processing time while maintaining comprehensive recognition capability across all classes.
Solution Approach 2:
The patent introduces a hierarchical dimension to organize the large number of classes into super-classes and sub-classes. This dimensional organization allows the system to process images through multiple levels, reducing the computational burden at each stage while maintaining the ability to recognize all classes, thereby reducing overall processing time.
Data Source
AI summary
A system and method for for grouping of similar image classes for image recognition is provided. The invention comprises extracting one or more features from multiple classes of images for determining a correlation value between each of the multiple classes of images based on assessment of the extracted features of each one of the classes of images with respect to other classes of images in the multiple classes of images. Further, the one class of image is grouped with the other class of image in the multiple classes of images to form one or more groups of super-classes of similar class of images based on analysis of the determined correlation values with respect to a pre-determined threshold value. An input image is recognized based on the formed groups of super-classes followed by sub-class classification of the images.