Object Detection Using Salience and Information Amount
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods in images face accuracy degradation when the main object is not visually prominent and are susceptible to noise from environmental or observational factors.
Innovation Solution
An object detection apparatus that sets multiple partial regions within an input image, calculates salience degrees based on feature amount differences, and uses an information amount to determine the size of a third partial region for robust main object detection, employing various statistical divergence and entropy calculations to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the salience degree is calculated based on the difference in statistical feature amount distribution in the input image, then the main object in the image can be detected, but the accuracy in detecting the main object degrades if the main object in the image is not visually prominent
Solution Approach 1:
The image is divided into multiple partial regions (first partial region, second partial region, third partial region) to enable localized analysis. This segmentation allows the detection method to focus on specific regions with different characteristics, improving the ability to detect main objects that may not be visually prominent in the entire image.
Solution Approach 2:
Different partial regions are assigned different functions in the detection process. The first partial region is used for calculating salience degree, the second partial region for calculating information amount, and the third partial region (set based on information amount) for final detection. This local differentiation of quality and function improves detection accuracy for non-prominent objects.
2Measurement precision
If the size of the information amount (information entropy) contained in the main object in the input image is calculated, then the main object in the image can be detected, but this method is susceptible to noise caused by an environmental or observational factor
Solution Approach 1:
By segmenting the image into multiple partial regions and performing information amount calculation on a specific second partial region rather than the entire image, the method reduces the impact of noise from environmental or observational factors. The localized analysis in the second partial region makes the detection more robust against global noise.
Solution Approach 2:
The information amount calculation is performed locally in the second partial region, and the third partial region is set based on this local information amount. This localized approach quality-control measures reduce susceptibility to noise while maintaining detection accuracy.
Data Source
AI summary
An object detection apparatus first sets a first partial region having a preset size and a second partial region in a given point (pixel) in an input image. In addition, the object detection apparatus calculates a first information amount in the second partial region, and sets a third partial region based on the size of the first information amount. Furthermore, the object detection apparatus calculates a score based on a salience degree that is based on a difference in statistical feature amount distribution between the first partial region and the second partial region, and based on an information amount of feature amount in the third partial region. Lastly, the object detection apparatus detects a main object by calculating scores on the respective points in the image and applying a predetermined statistical process to the scores.