Image Classification via Segmentation and Confidence Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition technologies face challenges in accurately identifying objects in images due to issues like changes in character angle, aspect ratio, scale, lighting, and overlapping images, which affect the determination of a class associated with an image.
Innovation Solution
A method that determines a segmentation score and confidence score for image segments by comparing them to regions within an image, using these scores to identify classes without requiring known character fonts or orientations, enabling the recognition of uneven, skewed, or deformed strings and characters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional object recognition methods are used, then the system can identify objects in standard conditions, but it fails to accurately recognize objects with deformities such as skew, scale changes, lighting variations, and overlapping images
Solution Approach 1:
The patent segments the image into multiple candidate regions and generates multiple segmentations of the same image with different parameters. Each segmentation produces a set of candidate objects, allowing the system to evaluate objects under different segmentation assumptions and select the most reliable classification.
Solution Approach 2:
The system varies segmentation parameters to generate multiple segmentations of the image. By changing parameters such as threshold values, region boundaries, and candidate object definitions, the system creates multiple interpretations of the same image data, enabling robust classification that accounts for various deformities and conditions.
2Ease of manufacture
If the system requires known character fonts or orientations for recognition, then classification can be performed with predefined models, but it cannot recognize uneven, skewed, or deformed strings and characters
Solution Approach 1:
The patent creates a classification system that is universal and does not depend on specific fonts, orientations, or predefined models. The system can classify objects based on their visual characteristics alone, making it applicable to a wide variety of object types including deformed, skewed, or uneven characters without requiring specialized models for each case.
Solution Approach 2:
The system performs self-validation by generating multiple segmentations and candidate objects, then using confidence scores to determine the most reliable classification. The system does not rely on external predefined models but instead uses its own internal consistency checks across multiple segmentations to validate classifications.
3Reliability
If the system uses multiple segmentations and candidate objects, then it can improve classification reliability, but the computational complexity increases
Solution Approach 1:
The system generates multiple segmentations and candidate objects, which may seem excessive, but this redundancy enables reliable classification by allowing the system to compare results across different segmentations. The additional computational effort is justified by the significant improvement in classification confidence and accuracy, especially for difficult-to-classify objects.
Solution Approach 2:
The system uses confidence scores as feedback to evaluate the quality of classifications. By computing confidence scores based on the agreement across multiple segmentations and candidate objects, the system can identify high-confidence classifications and potentially focus subsequent processing on lower-confidence cases, managing computational complexity through feedback-driven prioritization.
Data Source
AI summary
The technology is directed to determining a class associated with an image. In some examples, a method determines the class associated with an image. The method can include determining a segmentation score for an image segment based on a comparison of the image segment and a region of an image. The region of the image can be associated with the image segment. The method further includes determining a confidence score for the image segment based on the segmentation score and a classification score. The classification score can be indicative of a similarity between the image segment and at least one class. The method further includes determining a class associated with the image based on the confidence score. The method further includes outputting the class associated with the image.