Object Detection Model Training for Multi-Scale Detection Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models struggle to maintain robustness against varying object sizes and distortions, leading to reduced detection precision in complex scenarios.
Innovation Solution
The training method involves inputting images into an object detection model twice, first without size distinction for a global view, and secondly by classifying objects into proposal regions based on size clusters, using duplicated classifiers to enhance sensitivity to different object sizes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple detection frameworks (RCNN, Fast RCNN, Faster RCNN, YOLO, SSD) are trained separately for different object scales, then detection precision for specific scales is improved, but device complexity and training time increase significantly
Solution Approach 1:
The patent merges multiple detection frameworks (RCNN, Fast RCNN, Faster RCNN, YOLO, SSD) into a single unified detection model. This unified model integrates the architectural advantages of different frameworks and can handle multiple object scales simultaneously, reducing the need to train and maintain separate models for each scale while improving overall detection precision.
Solution Approach 2:
The unified detection model is designed to be universal, capable of detecting objects at various scales (small, medium, large) within a single framework. This multi-functional approach allows the model to replace multiple specialized models, reducing device complexity and training time while maintaining high detection precision across different object sizes.
2Measurement precision
If multiple detection frameworks are trained separately for different object scales, then detection precision for specific scales is improved, but training time and computational resources increase
Solution Approach 1:
By combining multiple detection frameworks into one unified model, the patent eliminates the need to train separate models for different object scales. This single training process replaces multiple sequential training operations, significantly reducing total training time and computational resource consumption while achieving high detection precision across all scales.
Solution Approach 2:
The unified detection model is pre-designed with architectural components that can handle multiple object scales from the outset. This preliminary design incorporates multi-scale feature extraction and detection mechanisms, avoiding the need for subsequent separate training processes for different scales, thus saving time and computational resources.
3Device complexity
If conventional detection models are used, then device complexity is low, but they cannot accurately detect objects of different scales simultaneously
Solution Approach 1:
The patent combines the strengths of conventional detection models (RCNN, Fast RCNN, Faster RCNN, YOLO, SSD) into a unified framework that maintains reasonable device complexity while achieving high detection precision for objects of different scales. The integration leverages the architectural advantages of each conventional model without requiring separate deployments.
Solution Approach 2:
The unified detection model introduces a new dimension of scalability by incorporating multi-scale feature extraction mechanisms. This allows the model to detect objects at various scales simultaneously by processing features at different resolution levels, thereby improving detection precision without proportionally increasing device complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This application discloses an object detection model training method performed by a computing device. In the method, a classifier that has been trained in a first phase is duplicated to at least two copies; and in a training in a second phase, each classifier obtained through duplication is configured to detect to-be-detected objects with different sizes, and train an object detection model based on a detection result. In the method, the object detection model obtained in the two phases in a training state has higher accuracy of detecting a to-be-detected object.