Feature Extraction Apparatus for Object Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing object recognition technologies face accuracy issues due to shifting or turning of objects, which affects the positional relationship between pixels, and require extensive processing time for multidimensional parameter calculations.

Innovation Solution

A feature extraction apparatus that divides an object region into smaller regions, calculates small-region features, and computes the product of features from these regions to obtain image features, reducing the impact of object movement and processing load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If feature calculation is performed for each pixel using positional relationships with mapped pixels, then object recognition can be performed, but accuracy deteriorates when objects shift or turn due to changes in positional relationships

Engineering Contradiction:

Improvefeature calculation accuracyVSAvoidrecognition accuracy under object movement

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent divides the image into multiple small regions and calculates features for each small region independently. By segmenting the image at a coarser level than individual pixels, the method reduces sensitivity to positional relationship changes caused by object shifting or turning, thereby maintaining recognition accuracy while avoiding the accuracy deterioration that occurs with pixel-level feature calculation.

Inventive Principle:

Principle #1Segmentation

2Adaptability or versatility

If multidimensional parameters are calculated for each pixel to improve recognition capability, then more comprehensive features are obtained, but processing time increases significantly

Engineering Contradiction:

Improverecognition capabilityVSAvoidprocessing time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs multidimensional parameter calculations on small regions rather than on individual pixels. This segmentation approach maintains the comprehensive feature extraction capability needed for versatile object recognition while dramatically reducing the total number of calculations required, thus decreasing processing time significantly.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent calculates features for a subset of small regions rather than for every pixel in the image. By performing partial action on representative small regions, the method achieves sufficient recognition capability without the excessive processing time required for complete pixel-level analysis.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS11281926B2Feature extraction method and apparatus

Publication Date: 2022.03.22 DENSO CORP

AI summary

In a feature extraction apparatus, a small region setter divides an object region set in a captured image into small regions smaller in size than the object region. A small region extractor extracts, from each small region, a small-region feature that has two or more feature parameters associated with the corresponding small region. A feature calculator calculates a product of a first feature representing the small-region feature of a first small region selected as a target region from the small regions, and a second feature representing the small-region feature of a second small region in the small regions, the second small region having a predetermined positional relationship with the first small region. The feature calculator obtains, based on the product of the first feature and the second feature, an image feature for the first small region as the target region.