Target Detection Training Using Bounding Box and Category Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target detection models suffer from poor generalization and low accuracy in classifying and localizing objects in new image data due to overfitting on preset sample image sets, leading to reduced performance in real-world applications.
Innovation Solution
A model training method that incorporates a generation sub-model and a determination sub-model to iteratively update parameters based on bounding box distribution similarity and category similarity, using a determination sub-model to enhance the model's ability to learn and adapt to new data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If traditional object detection algorithms are used, then the algorithm is simple and easy to implement, but the detection speed is slow and cannot meet real-time requirements
Solution Approach 1:
The patent replaces traditional mechanical object detection algorithms with a neural network-based deep learning system. The neural network model is trained on pre-processed image data and deployed for real-time detection, substituting the mechanical algorithmic approach with a learned neural system that achieves both high speed and accuracy.
Solution Approach 2:
The patent implements pre-processing of training images before neural network training, including image enhancement, feature extraction, and data augmentation. This preliminary action prepares the data in advance to accelerate the detection process during actual use while maintaining high accuracy.
2Measurement precision
If deep learning methods are used to improve detection accuracy, then detection precision improves, but the training time and computational resources increase significantly
Solution Approach 1:
The patent performs extensive pre-processing of training images before neural network training, including image enhancement, feature extraction, and data augmentation. This preliminary action prepares the data in advance to accelerate the detection process during actual use while maintaining high accuracy.
Solution Approach 2:
The patent extracts key features from training images before feeding them to the neural network. By taking out and pre-processing essential features, the system reduces the computational burden during training and inference while preserving detection precision.
3Measurement precision
If more training data is used to improve model generalization, then detection accuracy improves, but data processing time and storage requirements increase
Solution Approach 1:
The patent extracts key features from training images before feeding them to the neural network. By taking out and pre-processing essential features, the system reduces the computational burden during training and inference while preserving detection precision.
Solution Approach 2:
The patent uses data augmentation techniques to create synthetic copies of training images through transformations such as rotation, flipping, and scaling. This increases the effective training data size without requiring additional physical images, improving model generalization while controlling data processing load.
Data Source
Figure 1~2
Figure 3
Figure 4a
AI summary
The present application provides a model training method, a target detection method and a device. In a model training phase, based on a first reference bounding box and its corresponding real bounding box and real category, a model to be trained is prompted to continuously learn a bounding box distribution and an identifying of a category of a target object; and determine a loss value based on a determination result set output by a determination sub-model, and then continuously update model parameters for multiple rounds of iterations based on the loss value, where the determination result set includes a first determination result that represents a bounding box distribution similarity, and a second determination result that represents a category similarity between the first prediction category and the real category.