Image Feature Extraction Using Overlap-Aware DNN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image feature extraction methods using deep neural networks (DNNs) face challenges in achieving high accuracy due to the need for human-labeled true value labels and confusion during learning, particularly in determining overlap regions between images.
Innovation Solution
A method that involves determining overlap points between images to divide them into overlap and non-overlap regions, using three-dimensional information for training a DNN to automatically label overlap regions and separate feature extraction processes, thereby improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for image feature extraction, then the capability to extract features is improved, but the need for human-labeled true value labels increases and confusion during learning occurs
Solution Approach 1:
The patent segments the image into multiple regions (first region, second region, third region, fourth region) based on overlap relationships between images. This segmentation allows the DNN to process different regions separately with different training strategies, reducing confusion during learning while maintaining high feature extraction accuracy. The overlap region is specifically identified and handled differently from non-overlap regions.
Solution Approach 2:
The patent performs preliminary actions by automatically determining overlap points and dividing images into overlap and non-overlap regions before training the DNN. This preliminary region division eliminates the need for manual true value label creation, as the overlap regions are automatically identified based on geometric relationships. The preparation work is done in advance, simplifying the overall learning process.
2Reliability
If manual labeling of overlap regions is performed, then training data quality is improved, but the time and effort required increases
Solution Approach 1:
The system performs self-service by automatically determining overlap points between images and dividing them into overlap and non-overlap regions without human intervention. The overlap regions are automatically identified based on the geometric relationships between images, and the DNN is trained using these automatically generated labels. This eliminates the need for manual labeling while maintaining high training data quality.
Solution Approach 2:
The patent performs preliminary action by automatically creating the training data structure before model training. Overlap points are determined, regions are divided, and training datasets are prepared in advance through automated processes. This preliminary automated preparation eliminates the time-consuming manual labeling step while ensuring reliable training data quality.
3Ease of operation
If overlap regions are not distinguished during training, then the training process is simplified, but feature extraction accuracy decreases
Solution Approach 1:
The patent applies local quality by treating different regions of the image differently during training. The overlap region (where images intersect) is distinguished from non-overlap regions, and the DNN is trained to process these regions with different considerations. This local differentiation improves feature extraction accuracy for overlap regions while maintaining overall training feasibility.
Solution Approach 2:
The patent segments the image into distinct regions (overlap and non-overlap) based on geometric relationships. This segmentation allows the training process to specifically address the unique characteristics of overlap regions, improving feature extraction accuracy for these areas without overly complicating the overall training procedure.
Data Source
AI summary
An information processing apparatus including circuitry configured to receive a first image, receive a model from a learning device, and output a first position and a first posture based on a first image feature amount extracted from the first image and the model, wherein the model is obtained by determining at least one overlap point between a second image and a third image, dividing the second image into an overlap region including the at least one overlap point and a non-overlap region in response to determining the at least one overlap point, and performing training based on a second image feature amount corresponding to the overlap region and a third image feature amount corresponding to the non-overlap region.