3D Position Inference Across Image Domains for Abnormality Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality detection methods for autonomous moving devices, such as drones, struggle to accurately detect abnormalities due to varying natural environments without requiring extensive training data, leading to incomplete detection capabilities.
Innovation Solution
An inference device that utilizes a first learning model to analyze temporal changes in the three-dimensional position of detection targets by combining feature amounts from images with different domains and recognition tasks, using a second learning model to regress temporal changes without explicitly considering environmental conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a learning model is learned in consideration of all states of the natural environment, then detection accuracy across different environments is improved, but the amount of training data required increases excessively
Solution Approach 1:
The patent segments the detection task into two separate components: (1) a domain conversion model that transforms images from different environmental domains to a reference domain, and (2) a detection model trained only on the reference domain. This segmentation allows the detection model to focus on learning object detection without being burdened by environmental variations, thereby reducing training data requirements while maintaining detection accuracy across different environments.
Solution Approach 2:
The patent introduces an intermediary domain conversion model that acts as a mediator between images from different environmental domains and the detection model. This intermediary transforms input images from various environments into a standardized reference domain, allowing the detection model to process all inputs uniformly without requiring separate training for each environment, thus solving the contradiction between detection reliability and training data quantity.
2Adaptability or versatility
If the domain of the inference target image is different from the domain of the training image, then the model can handle diverse environmental conditions, but the detection accuracy decreases without domain adaptation
Solution Approach 1:
The patent applies preliminary action by performing domain conversion on inference target images before they are input to the detection model. The domain conversion model pre-processes images from different environmental domains, transforming them into the reference domain in advance. This preliminary transformation ensures that the detection model receives standardized inputs, maintaining high detection accuracy while enabling the system to handle diverse environmental conditions.
3Measurement precision
If time-series data is divided into multiple intervals for analysis, then temporal changes can be detected more precisely, but the complexity of processing increases
Solution Approach 1:
The patent segments the time-series data into multiple divided data sets at predetermined intervals, with each segment processed by the domain conversion model and detection model. This segmentation enables precise temporal change detection by analyzing each interval separately, while the modular architecture of the domain conversion model and detection model keeps processing complexity manageable through reusable components.
Data Source
AI summary
An image signal indicating a target image is acquired when a domain of the target image is different from that of a training image or a recognition task of the target image is different from a pre-learned task. The image signal is provided to a trained learning model. An inference time feature amount obtained by combining feature amounts of the detection target after the feature amounts are blurred is acquired from the learning model. A three-dimensional position of the detection target is estimated on the basis of a representative feature amount being a registered feature amount of the detection target appearing in an image for conversion whose domain and recognition task of the image are the same as those of the target image, and the inference time feature amount. A temporal positional change of the detection target in the target image is analyzed on the basis of the estimated position.