Image Recognition Apparatus Using Dynamic Dimension Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition techniques face challenges in accurately discriminating objects due to variations in illumination, pose, expression, and occlusion, requiring a large number of registration patterns that increase storage and memory demands, leading to inefficiencies in parameter management and power consumption.
Innovation Solution
An image recognizing apparatus that extracts partial feature quantities from images, reduces their dimension using a compressing unit with dynamically generated parameters, and calculates similarity between input and registration images, thereby reducing the number of parameters and improving accuracy while minimizing storage and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of registration patterns are used to cope with variations in illumination, pose, expression, and occlusion, then recognition accuracy is improved, but storage requirements and memory bandwidth increase
Solution Approach 1:
The patent divides the face image into multiple local areas (e.g., eyes, nose, mouth regions) and extracts features independently from each area. This segmentation allows the system to handle variations more efficiently by focusing on discriminative local features rather than requiring numerous full-face registration patterns, thereby reducing storage requirements while maintaining recognition accuracy.
Solution Approach 2:
The patent extracts only the essential feature quantities from local areas that are most discriminative for identification, rather than storing complete registration patterns. By taking out and storing only these critical feature elements, the system achieves high recognition accuracy with significantly reduced storage requirements.
2Measurement precision
If a large number of registration patterns are used to cope with variations, then recognition accuracy is improved, but memory bandwidth and power consumption increase
Solution Approach 1:
By segmenting the face into local areas and processing features from each area independently, the system reduces the total data volume that needs to be read from memory during authentication. This decreases memory bandwidth requirements and consequently reduces power consumption while maintaining recognition accuracy through the use of discriminative local features.
Solution Approach 2:
The system extracts and stores only the essential feature quantities from local areas rather than complete registration patterns. This extraction approach minimizes the amount of data that must be transferred through memory during operation, thereby reducing power consumption while preserving recognition accuracy.
3Reliability
If conversion is performed using linear operation with high dimensionality, then feature robustness to variation is improved, but parameter quantity and storage area increase
Solution Approach 1:
The patent applies conversion operations separately to features extracted from each local area rather than to the entire face feature vector. This segmentation of the conversion process reduces the dimensionality of parameters needed for each local area, thereby reducing total parameter quantity while maintaining robustness to variations through the combined information from multiple segmented areas.
4Measurement precision
If local portions are selected for feature extraction to cope with variations, then recognition accuracy is improved, but the number of parameters for conversion increases
Solution Approach 1:
The patent applies different conversion parameters to different local areas based on their specific characteristics and discriminative value. By tailoring the conversion process to each local area's quality and importance, the system achieves high recognition accuracy while optimizing the number of parameters needed for each area, thereby managing overall parameter quantity more efficiently.
Data Source
AI summary
An image recognizing apparatus comprises: an extracting unit extracting a partial feature quantity from an object in each of a registration image and an input image; a compressing unit reducing a dimension of the extracted partial feature quantity; a storing unit storing the partial feature quantity of the object in the registration image of which the dimension has been reduced; and a calculating unit calculating similarity between the object in the input image and the object in the registration image, using the partial feature quantity of the object in the input image of which the dimension has been reduced and the partial feature quantity of the object in the stored registration image. The compressing unit switches the reduction of the dimension of the partial feature quantity using a preset dimension compression parameter and the reduction of the dimension of the partial feature quantity by dynamically generating a dimension compression parameter.