Logo Detection via Agnostic Region Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing logo detection systems face challenges in accurately and efficiently detecting multiple logos in images and videos, especially with limited annotated training samples, and struggle to adapt to new logo classes without retraining models.
Innovation Solution
The use of an agnostic logo detection model in combination with a logo detection model, trained using both real-world and synthetic training samples, to identify candidate regions and detect logos, allowing for the detection of multiple logos and new classes without retraining, by extracting feature vectors and comparing them with a database of known logos.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional logo detection models are used, then detection accuracy for known logos is achieved, but the system cannot detect new logo classes without retraining
Solution Approach 1:
The system segments the logo detection task into two independent components: (1) an agnostic logo detector that identifies candidate regions regardless of logo class, and (2) a separate logo recognition module that matches extracted features against a database. This segmentation allows new logo classes to be added to the database without retraining the detection model, resolving the contradiction between adaptability and retraining time.
Solution Approach 2:
The patent introduces an intermediary feature extraction and matching layer between the logo detector and the logo database. Instead of directly training the detector on specific logo classes, the system extracts features from candidate regions and compares them against logo templates in the database. This intermediary mechanism enables flexible addition of new logo classes without affecting the detection model.
2Adaptability or versatility
If multiple logo detection models are used to improve detection coverage, then more logo classes can be detected, but system complexity increases
Solution Approach 1:
The agnostic logo detector is designed to be universal and class-agnostic, capable of detecting candidate regions for any logo class without being specialized for specific brands. This single universal detector replaces the need for multiple class-specific detectors, reducing system complexity while maintaining broad detection coverage across diverse logo classes.
3Measurement precision
If more annotated training samples are used to improve model accuracy, then detection precision increases, but data acquisition and annotation time increases
Solution Approach 1:
The system performs preliminary action by pre-processing and storing logo templates in the database before detection. Instead of requiring extensive annotated training data for each new logo class, the system pre-extracts and stores reference logo features that can be quickly compared during detection. This preliminary preparation reduces the need for time-consuming annotation of new logo classes.
Data Source
AI summary
Disclosed herein are techniques for detecting logos in images or video. In one embodiment, one or more candidate regions are detected for determining logos in an image. A logo is determined to be the logo in the candidate region based on matching a feature vector of a candidate region to a feature vector of the logo.