Image Recognition System Using Entropy-Based Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition technologies face challenges in achieving accurate recognition results across different fields due to varying image characteristics, such as low brightness in 'Under an overpass' or large light changes in 'Intersection of roads', as they rely on general pre-trained models without field-specific training.
Innovation Solution
An image recognition system comprising a cloud server and a local server, where the local server recognizes images using a local model, calculates a local entropy indicator based on field and confidence scores, and transmits images to the cloud server for updating the local model, incorporating field information and confidence scores to enhance recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general pre-trained image recognition model is used for all fields, then the device complexity is reduced and ease of operation is improved, but the recognition accuracy deteriorates in field-specific conditions
Solution Approach 1:
The patent divides the image recognition system into two segments: a general pre-trained model for initial recognition and a field-specific updated model for improved accuracy. The system automatically segments images into those requiring general model processing and those requiring updated model processing based on entropy calculation, thereby resolving the contradiction between using a simple general model and achieving high field-specific accuracy.
Solution Approach 2:
The patent changes the model parameters dynamically by calculating entropy indicators for different images and selectively applying updated models only when needed. This parameter-based selection mechanism allows the system to maintain low complexity for most images while achieving high accuracy for field-specific images, resolving the contradiction between model complexity and recognition accuracy.
2Measurement precision
If field-specific updated models are trained for all images, then the recognition accuracy is improved, but the loss of time and computational resources increases
Solution Approach 1:
The patent applies partial action by using the field-specific updated model only for a portion of images that have high entropy indicators and require improved accuracy, rather than applying it to all images. This selective application reduces processing time while maintaining high accuracy where needed, resolving the contradiction between comprehensive model application and processing efficiency.
Solution Approach 2:
The system implements feedback through entropy calculation, where the general model's recognition results are evaluated and fed back to determine whether field-specific updated models should be applied. This feedback mechanism optimizes processing time by avoiding unnecessary updated model application while ensuring high accuracy for images that benefit from it.
3Adaptability or versatility
If field-specific updated models are deployed, then the adaptability to different fields is improved, but the device complexity and storage requirements increase
Solution Approach 1:
The patent creates a universal system that can handle both general and field-specific recognition tasks through a single automated framework. The entropy-based selection mechanism provides multi-functionality, allowing the system to adapt to different fields without requiring separate manual configuration for each field, thereby resolving the contradiction between field adaptability and system complexity.
Solution Approach 2:
The system implements self-service through automated entropy calculation and model selection, where the system automatically determines which images require field-specific updated models without human intervention. This self-service capability improves field adaptability while minimizing the complexity burden on users, as the system autonomously manages the complexity of deploying multiple models.
Data Source
AI summary
An image recognition system includes a cloud server that stores a cloud recognition model and a local server that stores a local recognition model. The local server recognizes an image through the local recognition model to generate a local recognition result. The image corresponds to a field, the image includes at least one object, and the local recognition result includes local category confidence scores of each object corresponding to each of a plurality of categories. For each object, the local server calculates a local entropy indicator according to the field and the corresponding local category confidence scores. The local server determines to transmit the image to the cloud server according to the local entropy indicators of the image. The cloud server recognizes the image through the cloud recognition model to generate a cloud recognition result, and updates the local recognition model according to the cloud recognition result.