Target Object Recognition with Logo-Assisted Region Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recognizing important information in game and competition videos suffer from low accuracy and adaptability across different scenarios due to frequent updates and complex backgrounds, leading to poor user experience.
Innovation Solution
A target object recognition method using a lightweight neural network model, combined with logo-assisted positioning, to accurately detect and recognize score areas in videos or images, employing MobileNet-based SSD for initial detection and multi-label classification for digit recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing methods are used for recognizing important information in game and competition videos, then the recognition process can be performed, but the recognition accuracy is low and adaptability across different scenarios is poor
Solution Approach 1:
The patent segments the recognition task into multiple stages: initial detection using a lightweight model, verification using a second model, and refinement through adjustment. This multi-stage segmentation allows each model to focus on specific aspects, improving both accuracy and adaptability without requiring a single complex model to handle all scenarios
Solution Approach 2:
The patent performs preliminary detection using a lightweight first detection model to identify candidate target objects before applying more complex verification. This preliminary action filters out obvious cases early, allowing the system to allocate computational resources efficiently and improve adaptability across different scenario complexities
2Adaptability or versatility
If frequent updates occur in game and competition scenarios, then new scenarios can be accommodated, but the adaptability of existing recognition methods deteriorates
Solution Approach 1:
The patent employs a universal detection framework that can handle multiple scenario types (games, competitions, videos, images) using the same lightweight first detection model. This multi-functional approach maintains reliability across diverse scenarios while adapting to frequent updates without requiring scenario-specific model retraining
3Loss of information
If complex backgrounds are present in videos or images, then more information is available for analysis, but the recognition accuracy deteriorates
Solution Approach 1:
The patent extracts and focuses on specific candidate regions identified by the first detection model, isolating them from complex backgrounds. By taking out only the relevant candidate areas for verification and refinement, the system maintains high recognition accuracy even when processing images or videos with complex backgrounds containing大量 irrelevant information
Data Source
AI summary
A target object recognition method is provided. The method includes: inputting a received initial picture into a first detection model, and obtaining an initial location of each target object in the initial picture; inputting a candidate picture corresponding to the initial location into a second detection model, and obtaining a verification object in the candidate picture and a verification location of the verification object in the candidate picture; adjusting the initial location of each target object based on the verification location to obtain a target location of each target object; and inputting a target picture corresponding to the target location into a recognition model to obtain each target object in the initial picture.