Target Detection Model Training for Self-Similar Object Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target detection models struggle with inaccurate detection and difficult annotation when dealing with targets that exhibit self-similarity, such as flames, due to ambiguous annotation standards and manual annotation processes, leading to less precise model identification.
Innovation Solution
A method and apparatus for training a target detection model that involves determining a relationship between predicted and annotated regions, setting a low loss function value when the predicted region is sufficiently within the annotated region, and performing weighted summation of loss function values to enhance model precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If manual annotation is used for self-similar targets, then annotation can be performed, but annotation precision deteriorates due to ambiguous standards
Solution Approach 1:
The system uses the target detection model itself to generate annotation suggestions for self-similar targets. The model predicts bounding boxes and the system automatically determines loss function values based on intersection ratios, allowing the system to annotate itself without external manual intervention, thereby resolving the ambiguity of manual annotation standards
Solution Approach 2:
The patent changes the parameter of loss function value assignment based on the intersection ratio between predicted and ground truth regions. When the ratio exceeds a threshold, a predetermined low loss value is assigned; otherwise, the calculated loss value is used. This parameter change enables precise handling of self-similar targets by adapting the loss value to the detection confidence
2Productivity
If conventional loss function calculation is used, then model training can proceed, but detection precision deteriorates for self-similar targets
Solution Approach 1:
The patent applies different loss function values to different predicted regions based on their intersection ratio with ground truth. Regions with high intersection ratios (above threshold) receive predetermined low loss values, while other regions receive calculated loss values. This local differentiation improves detection precision for self-similar targets without compromising overall training efficiency
Solution Approach 2:
The training process is segmented into two pathways: one for regions satisfying the target relationship (high intersection ratio) with predetermined loss values, and another for regions not satisfying the relationship with calculated loss values. This segmentation allows the model to learn differently from correct and incorrect predictions, improving overall detection precision
3Measurement precision
If strict annotation standards are enforced, then annotation precision improves, but annotation workload increases
Solution Approach 1:
The system performs self-annotation using the target detection model's predictions. The automated determination of loss function values based on intersection ratios eliminates the need for complex manual annotation standards, reducing annotation complexity while maintaining precision through the model's own learning capability
Data Source
AI summary
Provided is a method for training a target detection model. The method includes: determining a first region in a sample image; counting training periods of the target detection model, and determining a relationship between an intersection region and the first region in response to a number of training periods of the target detection model reaching a target number; determining, in response to the relationship between the intersection region and the first region satisfying a target relationship, a predetermined low loss function value as a loss function value of the first region; and training the target detection model with reference to the loss function value, wherein the target detection model is used to detect one or more of fireworks, dust, clouds, shorelines, flame, smoke, steam, or night light in an image.