Re-Identification Neural Network Update via Batch Normalization Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing re-identification neural network models for object recognition in surveillance systems face challenges in adapting to varying image characteristics across different camera installations, requiring costly data collection and refinement, and lengthy training processes, especially when deployed in edge devices.
Innovation Solution
A method for updating a re-identification neural network model by configuring a batch normalization layer to normalize image characteristic parameters using feedforward training data, without the need for labeling or computing a loss function, allowing adaptation to new camera environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a pre-trained re-identification neural network model is deployed at a new camera location, then the system can immediately operate, but the model fails to adapt to different image characteristics (brightness, background, blur) exhibited by the new location
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing reference image characteristics (mean and variance) during the training phase. When deploying at a new location, these pre-computed reference characteristics are immediately available for comparison, eliminating the need for time-consuming real-time data collection and training while enabling rapid adaptation to new image characteristics.
Solution Approach 2:
The patent changes parameters by updating the neural network model's statistical parameters (mean and variance) based on the difference between reference image characteristics and actual image characteristics at the new location. This parameter update mechanism allows the model to adapt to different image characteristics without requiring complete retraining, thus resolving the contradiction between adaptability and time loss.
2Reliability
If training data is collected and refined for each new camera location, then the model can be updated to match local characteristics, but the process requires significant time and computational resources
Solution Approach 1:
The patent extracts only the essential statistical characteristics (mean and variance) from the training data at each location, rather than using the complete dataset for training. By extracting these key parameters and storing them as reference characteristics, the system can quickly compare and update without processing the entire dataset, thus improving training efficiency while maintaining reliability.
Solution Approach 2:
The patent applies partial action by updating only the necessary statistical parameters (mean and variance) of the neural network model based on the difference between reference and actual image characteristics, rather than performing complete retraining. This partial update approach maintains sufficient reliability for re-identification while dramatically improving productivity by reducing computational resources and time required.
3Adaptability or versatility
If the neural network model is updated using conventional training methods, then the model can adapt to new data, but the process requires labeling training data and computing loss functions which increases complexity
Solution Approach 1:
The patent applies self-service by enabling the neural network model to update itself automatically based on the difference between reference image characteristics and actual image characteristics. The model computes its own update parameters without requiring external labeling or loss function computation, thus achieving adaptability while reducing device complexity by eliminating the need for complex training procedures.
4Adaptability or versatility
If multiple cameras are installed at different locations, then comprehensive surveillance coverage is achieved, but the cost and complexity of collecting and processing data from all locations increases
Solution Approach 1:
The patent segments the data processing task by handling each camera location independently, extracting and storing reference characteristics for each location separately. This segmentation allows the system to process and adapt to each location's data locally without needing to process all locations' data simultaneously, thus achieving comprehensive surveillance coverage while reducing the overall data processing volume and complexity.
Data Source
AI summary
A method for updating a neural network model for object re-identification includes: storing a neural network model pre-trained for object re-identification; acquiring images from a surveillance camera device; detecting objects from the images and, obtaining training data from among the objects to update the neural network model according to a predetermined criterion; and inputting the training data to the neural network model in a feedforward manner to obtain image characteristic parameters corresponding to the training data, and updating the neural network model by reflecting the image characteristic parameters in the neural network model.