Neural Network Training with Dual Loss Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training neural network models for person recognition in video understanding are inefficient in optimizing parameters, particularly when dealing with image data that requires both position and serial number information for effective training.
Innovation Solution
A training method for neural network models that involves receiving image data with detection and person search data, using a neural network model to perform person recognition, and employing multiple loss functions to adjust the model parameters based on recognition results, detection data, and person search data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional single loss function training is used for person recognition, then the training process is simple, but the parameter optimization efficiency is low and recognition accuracy is insufficient
Solution Approach 1:
The training method segments the person recognition task into two distinct components: person detection (locating persons in images) and person reidentification (matching persons across different images using serial numbers). Correspondingly, two separate loss functions are designed - detection loss for position accuracy and reidentification loss for serial number accuracy - allowing each component to be optimized independently while contributing to overall recognition accuracy.
2Productivity
If multiple loss functions are used for comprehensive parameter optimization, then the training effectiveness improves, but the computational complexity increases
Solution Approach 1:
The patent merges the detection loss and reidentification loss into a unified total loss function through weighted summation. This allows both person detection accuracy and person reidentification accuracy to be optimized simultaneously within a single training framework, improving overall training efficiency while maintaining manageable computational complexity through the use of standard backpropagation algorithms.
Data Source
AI summary
The present invention provides a training method of a neural network model, wherein the training method includes the steps of: receiving image data including a plurality of frames, and for first frames in the frames, the image data further includes detection data, and the detection data includes position of at least one person within the corresponding first frame; and for second frames in the frames, the image data further includes person search data, and the person search data includes position and serial number of at least one person within the corresponding second frame; using the neural network model to perform a person recognition operation on the frames to generate a recognition result; and using loss functions to process the recognition result of each frame, the detection result of each first frame and the person search data of each second frame, for adjusting parameters of the neural network model.