Object Detection via Illumination-Invariant Parameter Adjustment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing systems face challenges in accurately detecting objects when the environment of the input image differs from the environment in which the learning images were obtained, particularly due to variations in illumination, leading to reduced detection accuracy.

Innovation Solution

An image processing apparatus and method that includes a recognition section to detect objects based on a learning result, and a setting section to adjust parameters for subsequent frames in response to differences in image information between the object image and the learning image, thereby enhancing detection accuracy across varying environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a recognizer is obtained by learning using learning images under specific illumination conditions, then the object detection accuracy is improved under those conditions, but the detection accuracy deteriorates when the illumination environment differs from the learning environment

Engineering Contradiction:
Improveobject detection accuracyVSAvoidenvironmental adaptability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent changes the parameter of image information by selectively extracting pixels with intensity values within a specific range (between first and second threshold values) from the learning images. This parameter change in pixel selection allows the system to focus on mid-tone regions that are more consistent across different illumination conditions, thereby improving environmental adaptability while maintaining detection accuracy

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent creates an average image that copies and combines pixel information from multiple learning images taken under different illumination conditions. This averaged representation serves as a generalized model that captures common features across environments, enabling the recognizer to maintain accuracy when deployed in varying illumination conditions without requiring re-learning

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If multiple learning images under various illumination conditions are used to improve environmental adaptability, then the detection accuracy under different conditions is improved, but the complexity of the learning process and data requirements increase

Engineering Contradiction:
Improveenvironmental adaptabilityVSAvoidlearning process complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent simplifies the learning process by changing the parameter of pixel selection to focus only on pixels with intensity values between specific thresholds. This parameter-based filtering approach is computationally simpler than full-image processing and reduces the complexity of creating and processing multiple learning images under various conditions

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent uses copying and averaging of pixel information from multiple learning images to create a single average image that encapsulates environmental variations. This approach consolidates multiple complex learning images into one simplified representation, reducing the complexity of storing and processing numerous condition-specific images

Inventive Principle:
Principle #26Copying

Data Source

PatentUS8873800B2Image processing apparatus and method, and program
Publication Date: 2014.10.28 SONY GROUP CORP
  • US8873800B2 patent drawing
  • US8873800B2 patent drawing
  • US8873800B2 patent drawing

AI summary

The present disclosure provides an image processing apparatus, including: a recognition section adapted to recognize, based on a learning result obtained by learning of a learning image regarding a predetermined object, the object in a predetermined frame of an input image formed from a plurality of frames which are continuous in time; and a setting section adapted to set a parameter to be used for a process to be carried out for a later frame which is later in time than the predetermined frame of the input image in response to a difference in image information between an object image, which is an image in a region of the object recognized in the predetermined frame, and the learning image; the recognition section recognizing the object in the later frame for which the process is carried out based on the parameter set by the setting section.