Image Recognition Apparatus Using Gamma-Value Transformation for Robust Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies face challenges in robustly recognizing objects across varying image capturing conditions, as they often learn irrelevant features such as brightness and lens characteristics, leading to over-learning and misclassification.
Innovation Solution
An image recognition apparatus that acquires images, changes their parameters through γ-value transformation, extracts features using local region extraction and classification, and integrates recognition results using a classifier to output a target value, thereby improving accuracy and reducing nonsystematic misdiscrimination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If image processing is performed to increase variations of learning examples by adding noise, changing brightness values, and applying affine deformation, then the number of learning examples is increased, but the processed images partially differ from actual distribution of variation in the image causing over-learning
Solution Approach 1:
The patent applies parameter changes by performing γ-value transformation on learning images to adjust brightness and contrast parameters. This creates varied learning examples that reflect actual distribution of variation in images, preventing over-learning while increasing the effective number of learning examples. The γ-value transformation specifically addresses the contradiction by modifying image parameters in a way that maintains realism.
2Productivity
If learning is performed based on a small number of learning examples, then the learning process is simpler and faster, but irrelevant features such as brightness and lens characteristics are learned incorrectly as part of object features
Solution Approach 1:
The patent performs parameter changes on learning images using γ-value transformation to create multiple variations from a small number of original images. This allows the system to learn robust features without requiring a large dataset, maintaining learning efficiency while preventing the model from incorrectly learning irrelevant features like brightness and lens characteristics as object features.
Solution Approach 2:
The patent applies preliminary action by pre-processing learning images with γ-value transformation before the actual learning process. This preliminary parameter adjustment ensures that the learning model is exposed to varied conditions from the start, preventing it from learning incorrect associations between irrelevant features and object identities, thereby improving feature extraction accuracy without increasing the number of original images needed.
Data Source
AI summary
In the present disclosure, an image parameter of an input image is changed, features is extracted from each of a plurality of generated images, a category of each region is determined based on the features in each image, and the results are integrated.