Generalized Invariant Shape Descriptor for Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine vision techniques lack effective methods for feature selection and normalization, especially in non-isotropic feature spaces with limited training samples, and fail to integrate scale and reflection features effectively for shape classification.
Innovation Solution
A system and method for analyzing images using a generalized invariant shape descriptor that extracts features invariant to scale, translation, rotation, and symmetry, allowing for characterization and discrimination of objects with any shape, and combines these features with scale and reflection-dependent descriptors for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional feature extraction methods are used, then feature extraction can be performed, but feature selection and normalization become complex and unreliable in non-isotropic feature spaces with limited training samples
Solution Approach 1:
The patent applies universality by creating a comprehensive feature descriptor that serves multiple functions simultaneously: it extracts shape features, provides normalization, and works across different classification scenarios. The generalized invariant shape descriptor integrates multiple feature types (scale-dependent, reflection-dependent, and invariant features) into a single unified representation that can be directly used for classification without additional processing steps.
Solution Approach 2:
The patent applies parameter changes by transforming the feature space through mathematical operations that make features invariant to scale, translation, and rotation. By applying specific transformations to the raw shape features, the system converts problematic non-isotropic feature spaces into normalized isotropic spaces where classification becomes more reliable and less complex.
2Productivity
If scale and reflection features are not integrated, then classification can be performed with simpler features, but classification accuracy for shapes with arbitrary orientations and scales deteriorates
Solution Approach 1:
The patent applies segmentation by dividing the feature extraction process into distinct components: scale-dependent features, reflection-dependent features, and invariant features. Each component is extracted and processed separately, then integrated into a comprehensive feature descriptor. This segmented approach allows each feature type to be optimized independently while maintaining overall classification accuracy.
Solution Approach 2:
The patent applies the composite materials principle by creating a composite feature descriptor that combines multiple types of features (scale, reflection, and invariant features) into a single integrated representation. This composite descriptor leverages the strengths of each individual feature type to achieve robust classification across diverse shape variations.
3Reliability
If a comprehensive feature set is used, then classification accuracy for arbitrary shapes improves, but computational complexity and processing time increase
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing the generalized invariant shape descriptors for training samples during an offline phase. This preprocessing step transforms raw shape data into normalized feature representations in advance, so that during online classification, the system only needs to compute distances between pre-processed features rather than performing full feature extraction and normalization for each comparison.
Data Source
AI summary
System and method for analyzing an image. A received image, comprising an object or objects, is optionally preprocessed. Invariant shape features of the object(s) are extracted using a generalized invariant feature descriptor. The generalized invariant feature descriptor may comprise a generalized invariant feature vector comprising components corresponding to attributes of each object, e.g., related to circularity, elongation, perimeter-ratio-based convexity, area-ratio-based convexity, hole-perimeter-ratio, hole-area-ratio, and/or functions of Hu Moment 1 and/or Hu Moment 2. Non-invariant features, e.g., scale and reflection, may be extracted to form corresponding feature vectors. The object is classified by computing differences between the generalized invariant feature vector (and optionally, non-invariant feature vectors) and respective generalized invariant feature vectors corresponding to reference objects, determining a minimum difference corresponding to a closest reference object or class of reference objects of the plurality of reference objects, and outputting an indication of the closest reference object or class as the classification.