AI Object Recognition Using Dynamic Model Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition technologies can only use one recognition model and struggle to accurately identify objects in image data, especially in varying environments, with no mechanism to improve recognition accuracy when the confidence level is low.
Innovation Solution
An artificial intelligence apparatus and method that uses a default recognition model to generate identification information, measures its confidence level, and constructs a compound recognition model when necessary to improve accuracy, combining multiple recognition models learned from different image data sets based on environment and image type information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single recognition model is used, then the device complexity is low, but the recognition accuracy and confidence level are insufficient
Solution Approach 1:
The patent divides the recognition system into multiple specialized recognition models, each trained on specific image data sets categorized by environment and image type. Instead of using one monolithic model, the system segments the recognition function across multiple models (e.g., outdoor models, indoor models, night models, day models) to handle different conditions optimally.
Solution Approach 2:
The patent implements a dynamic model selection mechanism that adjusts which recognition model(s) to use based on the input image characteristics. The confidence level measurement allows the system to dynamically switch from a default model to compound models when low confidence is detected, making the system adaptable to varying conditions without requiring all models to be active simultaneously.
2Reliability
If multiple recognition models are combined, then the recognition accuracy improves, but the device complexity increases
Solution Approach 1:
The patent merges multiple recognition models into a unified compound recognition model that integrates predictions from individual models. When low confidence is detected, the system combines the results from multiple models (e.g., outdoor and indoor models) to produce a more reliable recognition result, effectively merging their strengths to overcome individual model limitations.
Solution Approach 2:
The patent changes the operational parameters of the recognition system by adjusting which models are activated based on confidence levels. When confidence is high, only the default model processes the image (simpler path). When confidence is low, the system switches to compound models that combine multiple model predictions, dynamically changing the processing parameters to optimize reliability.
3Productivity
If a default recognition model is used for all cases, then the processing speed is fast, but the recognition accuracy decreases in various environments
Solution Approach 1:
The patent performs preliminary classification of image data into different categories (outdoor, indoor, night, day) and pre-trains separate recognition models for each category. This preliminary organization allows the system to quickly select the appropriate pre-trained model based on image characteristics, avoiding the need to evaluate all models from scratch and maintaining fast processing speeds while improving accuracy.
Data Source
AI summary
Disclosed herein is an artificial intelligence apparatus for recognizing at least one object, comprising: a memory configured to store a plurality of recognition models for generating identification information corresponding to the object from image data; and a processor configured to: obtain image data for the object, generate first identification information corresponding to the object from the image data using a default recognition model composed of at least one or more of the plurality of recognition models, measure a confidence level for the first identification information, obtain the first identification information as a recognition result of the object if the confidence level is equal to or greater than a first reference value, and obtain second identification information corresponding to the object from the image data as a recognition result of the object using a compound recognition model composed of at least one or more of the plurality of recognition models if the measured confidence level is less than the first reference value, wherein the default recognition model is a model defined by first weights for the plurality of recognition models, and wherein the compound recognition model is a model defined by second weights for the plurality of recognition models.