Adaptive Image Recognition Model Switching for Embedded Vision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Embedded devices with limited computational capacity struggle to execute convolutional neural networks effectively, particularly when recognizing small targets in high-resolution images, often setting suboptimal image capture conditions.
Innovation Solution
A recognition process execution program that employs a first trained model for initial image recognition, with a determination function to assess reliability, and switches to a second trained model with higher processing load if initial reliability is below a threshold, or adjusts image capture settings to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a first trained model is used for initial recognition to reduce computational load, then processing speed is improved, but recognition reliability deteriorates when the target is small or the image resolution is high
Solution Approach 1:
The system dynamically switches between a first trained model (lower processing load) and a second trained model (higher processing load) based on real-time determination of reliability. When reliability falls below a threshold, the system transitions to the second model to improve recognition accuracy, thereby adapting the processing approach to current conditions.
Solution Approach 2:
The determination function continuously monitors the reliability output from the first trained model and provides feedback to the control unit. This feedback mechanism triggers a switch to the second trained model when reliability is insufficient, creating a closed-loop system that adjusts processing based on actual performance.
2Reliability
If image capture conditions are set for high-resolution images to improve recognition accuracy, then recognition reliability is improved, but computational capacity requirements increase
Solution Approach 1:
The system uses a two-model approach where the first trained model handles routine recognition tasks with standard image quality, while the second trained model is activated only when necessary (e.g., when reliability is low or image resolution is high). This partial application of high-computational models only when needed resolves the contradiction between reliability and power consumption.
3Reliability
If a second trained model with higher processing load is used to improve recognition accuracy, then recognition reliability is improved, but device complexity increases
Solution Approach 1:
The recognition system is segmented into two distinct trained models with different processing loads, each handling specific scenarios. The control unit manages switching between these segmented models based on reliability determination, dividing the complex task into manageable parts that can be selected based on current needs.
Data Source
AI summary
A recognition process execution program that inputs first image data indicating a first image to a first trained model, makes the first trained model execute a first recognition process for recognizing a target appearing in the first image, makes the first trained model output first reliability data indicating a reliability of a result of the first recognition process, determines a reliability indicated by the first reliability data, and based on a determination result by the determination function, inputs second image data indicating a second image different from the first image data to a second trained model, and makes the second trained model execute a second recognition process, having a processing load greater than or equal to a processing load of the first recognition process, for recognizing the target appearing in the first image and the second image.