ULBP Feature Quantity Generation for Orientation-Invariant Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing object detection methods face challenges in accurately detecting physical objects of various orientations due to the need for multiple discriminators and large discrimination parameters, which increases processing time and costs, especially when dealing with rotations and reversals of feature quantities like local binary patterns.

Innovation Solution

An object detecting apparatus that extracts directional characteristic patterns, generates feature quantities considering rotations and reversals, and converts these quantities to simplify the detection process using a pattern discriminating apparatus with a ULBP feature quantity and a boosting discriminator, reducing the need for extensive conversion tables and parameter sets.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If multiple discriminators are provided for different orientations, then detection accuracy for various orientations is improved, but the number of discriminators and parameter size increase

Engineering Contradiction:

Improvedetection accuracyVSAvoidnumber of discriminators

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple discriminators for different orientations into a single discriminator by generating feature quantities that encode orientation information. Instead of maintaining separate discriminators for each orientation, the invention combines their functionality into one discriminator that receives orientation-aware feature quantities as input, thereby reducing the number of discriminators while preserving detection accuracy across various orientations.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent changes the parameters of the feature quantities to incorporate orientation information. By generating feature quantities that are transformed according to the object's orientation (rotation and reversal operations), the single discriminator can adapt to different orientations through parameter transformation rather than requiring multiple specialized discriminators.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If feature quantities are rotated and reversed to detect various orientations, then detection coverage is improved, but processing time increases

Engineering Contradiction:

Improvedetection coverageVSAvoidprocessing time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-generating and storing transformation tables that contain the results of rotation and reversal operations for feature quantities. Instead of performing computationally intensive rotation and reversal operations in real-time during detection, the invention pre-computes these transformations and stores them in lookup tables, allowing the system to quickly retrieve pre-computed feature quantities during the detection process.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If conversion tables for feature quantity transformation are created, then orientation invariance is improved, but memory requirements increase

Engineering Contradiction:

Improveorientation invarianceVSAvoidmemory requirements

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent extracts and stores only the essential transformation information in conversion tables. Instead of storing complete feature quantity datasets for all possible orientations, the invention extracts and stores only the transformation relationships (rotation and reversal mappings) in compact conversion tables. This allows the system to generate orientation-adapted feature quantities on-demand by applying these extracted transformation rules, reducing memory requirements while maintaining orientation invariance.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS9659227B2Detecting object from image data using feature quantities

Publication Date: 2017.05.23 CANON KK

AI summary

Object detecting apparatus and method enable to easily and accurately generate a feature quantity for detecting an object at low cost. To achieve this, in the object detecting apparatus and method, a value is assigned in consideration of relations of rotation and reversal in uniform patterns by using a small-sized conversion table, and, an ULBP feature quantity after rotation and reversal is generated from a previous ULBP feature quantity by simple calculation without using a conversion table indicating a correspondence relation between previous values and values after rotation and reversal.