Gaussian-Masked Model Training for Fast, Precise Image Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks face a trade-off between detection speed and precision, with large networks being slow but highly accurate, and small networks being fast but less accurate.
Innovation Solution
A model training method using Gaussian masks to highlight local features in multiple networks, followed by training the second network based on feature and classification losses, and further refinement with additional networks to improve performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large neural network is used, then detection precision is improved, but detection speed deteriorates
Solution Approach 1:
The patent segments the training process into multiple stages with different network configurations. Initially, a large network (first network) is trained to achieve high precision. Then, a smaller network (second network) is trained to approximate the features of the large network, achieving a balance between precision and speed. This segmentation allows the system to benefit from both high-precision and fast networks at different training phases.
Solution Approach 2:
The patent creates a copy relationship where the second network (smaller, faster) learns to replicate the feature extraction capabilities of the first network (larger, more precise). Through feature loss calculation and knowledge distillation, the second network copies the essential detection features from the first network, enabling it to achieve comparable precision with faster inference speed.
2Speed
If a small neural network is used, then detection speed is improved, but detection precision deteriorates
Solution Approach 1:
The patent performs preliminary training of the large network first to establish high-precision feature extraction capabilities. This preliminary action creates a teacher model that guides the subsequent training of the small network. The small network benefits from this preliminary high-precision model through knowledge distillation, allowing it to achieve better precision than it would through independent training alone.
Solution Approach 2:
The first network serves as an intermediary or teacher model that transfers knowledge to the second network. Through the mechanism of feature loss calculation and gradient-based optimization, the first network mediates the learning process of the second network, enabling the smaller network to acquire detection precision that would otherwise require a larger architecture.
3Measurement precision
If network depth is increased, then feature extraction capability is improved, but computational complexity increases
Solution Approach 1:
The patent changes the parameter of network depth dynamically during different training phases. In the first phase, a deep network (first network) is used to extract rich features. In the second phase, a shallower network (second network) is configured to approximate these features with reduced computational complexity. This parameter change allows the system to optimize between feature extraction capability and computational efficiency at different stages.
Data Source
AI summary
A model training method and a related apparatus are provided and may be used in computer vision to perform image detection. The method includes: extracting feature information from a target image; further separately extracting features of a target object from the feature information by using a Gaussian mask to obtain a first local feature and a second local feature; determining a feature loss by using the first local feature and the second local feature; performing prediction by using the first network and the second network based on a same region proposal set to obtain a first classification predicted value and a second classification predicted value, and obtaining a classification loss based on the first classification predicted value and the second classification predicted value; and training the second network based on the classification loss and the feature loss to obtain a target network.