Homogeneous Branch Neural Network for Person Re-Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing person re-identification (Re-ID) networks face challenges in improving accuracy and performance while reducing storage space occupation and calculation complexity, which hinders their deployment in portable devices and real-time applications.
Innovation Solution
The method involves creating a homogeneous training network with multiple branches of the same structure, using a knowledge synergy loss function to enhance feature interaction between branches, and training with cross-entropy and triplet loss functions to determine final weight parameters, thereby improving network accuracy and efficiency without increasing parameter quantity or calculation amount.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the network structure is deepened or widened to improve learning capability, then the accuracy and performance of person Re-ID networks is improved, but the parameter quantity and calculation amount increase
Solution Approach 1:
The network is divided into multiple homogeneous branches with identical structures, where each branch independently processes features. This segmentation allows the network to achieve better performance through ensemble learning while maintaining manageable parameter quantities in each individual branch, avoiding the need for a single large complex network
Solution Approach 2:
The patent introduces a knowledge synergy loss function that dynamically adjusts parameter updates during training by incorporating classification probabilities from multiple branches. This parameter change strategy enables the network to learn more effectively from diverse branch perspectives without increasing the fundamental parameter quantity, improving accuracy through smarter parameter utilization rather than more parameters
2Measurement precision
If the network structure is deepened or widened to improve learning capability, then the performance of person Re-ID networks is improved, but the calculation amount increases
Solution Approach 1:
By segmenting the network into multiple homogeneous branches with identical structures, the computational workload is distributed across branches. Each branch performs calculations on the same input data independently, but the homogeneous structure allows for efficient parallel processing and reduces redundant computations compared to heterogeneous deep networks
Solution Approach 2:
The knowledge synergy loss function modifies the training process by incorporating classification probabilities from multiple branches into the loss calculation. This parameter change enables more efficient learning convergence, reducing the number of training iterations needed to achieve high performance, thereby lowering the total calculation amount required
3Measurement precision
If the parameter quantity is increased to improve network capability, then the accuracy of person Re-ID networks is improved, but the storage space occupation increases
Solution Approach 1:
The network parameters are segmented across multiple homogeneous branches, where each branch contains a complete set of parameters for processing. This segmentation strategy allows the system to achieve high accuracy through ensemble effects while each individual branch maintains a compact parameter size suitable for storage-constrained environments like portable devices
Solution Approach 2:
Instead of creating a single large network with unique parameters for each function, the patent creates multiple copies of the same homogeneous branch structure. Each copy (branch) has identical parameter dimensions, but they are trained to produce diverse features through different initialization and the knowledge synergy loss function. This copying approach reduces storage requirements compared to a single large heterogeneous network while maintaining high accuracy through the ensemble of copies
Data Source
AI summary
A person re-identification method and apparatus, a device, and a readable storage medium. A homogeneous training network of an initial person re-identification network is trained by means of an objective function such as a knowledge synergy for dynamic classification probability loss function to obtain a final person re-identification network carrying more accurate final weight parameters, and a person re-identification task is performed by means of the final person re-identification network. In this way, the accuracy and performance of the person re-identification network to process the person re-identification task may be improved, the storage space in a device may be reduced, more beneficial to the storage and deployment of the portable device, and the amount of calculation of performing the person re-identification task may be reduced, thereby increasing the processing rate of the person re-identification task.