Object Tracking With Hybrid Neural Networks, Quantization, and Pruning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing number of parameters in neural networks for object detection leads to resource-intensive computations, necessitating mechanisms for reducing the network weight and accelerating computation while maintaining accuracy.
Innovation Solution
A hybrid neural network system comprising a first neural network module for lightweight quantization and a second neural network module for channel pruning, combined with a prediction and optimization module to generate and update parameters efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of parameters in neural network is increased to improve object detection accuracy, then detection accuracy is improved, but computation resource consumption increases
Solution Approach 1:
The patent divides the neural network into multiple modules with different functions (first neural network module for feature extraction, second neural network module for object detection, third neural network module for tracking). This segmentation allows each module to be optimized independently, reducing overall computational resource consumption while maintaining detection accuracy through specialized processing in each module
Solution Approach 2:
The patent dynamically adjusts the number of parameters and computational complexity of neural network modules based on the specific detection task requirements. By changing parameters such as the number of convolutional layers, filter sizes, and activation functions according to different scenarios, the system achieves high accuracy when needed while reducing resource consumption in less demanding situations
2Measurement precision
If the number of parameters in neural network is increased to improve object detection accuracy, then detection accuracy is improved, but computation speed decreases
Solution Approach 1:
The patent segments the detection process into multiple stages with different neural network modules, where simpler modules handle routine tasks quickly and more complex modules are only activated when needed. This segmentation enables the system to maintain high computation speed for common cases while achieving high accuracy when complex detection is required
Solution Approach 2:
The patent implements a hierarchical detection approach where a lightweight first neural network module performs initial feature extraction and filtering, and only potentially activates the more computationally intensive second neural network module for ambiguous or critical detections. This partial action approach maintains high speed for most cases while preserving accuracy for important detections
3Use of energy by moving object
If neural network weight is reduced to decrease resource consumption, then resource consumption is minimized, but detection accuracy deteriorates
Solution Approach 1:
The patent applies different levels of parameter reduction and optimization to different parts of the neural network based on their importance. Critical modules that directly impact detection accuracy are preserved with full precision, while less critical modules undergo more aggressive weight reduction and pruning, achieving resource minimization without significant accuracy loss
4Speed
If neural network weight is reduced to accelerate computation, then computation speed is improved, but detection accuracy deteriorates
Solution Approach 1:
The patent segments the computation into multiple passes with different levels of precision. The first neural network module performs rapid initial processing with reduced weights, and the second neural network module refines the results with higher precision weights for critical detections, achieving both speed and accuracy
Solution Approach 2:
The patent applies full-precision computation only when necessary for ambiguous or critical detections, while using reduced-precision computation for routine cases. This partial application of high-precision computation maintains accuracy for important cases while achieving speed improvements for the majority of detections
Data Source
AI summary
An object tracking learning system includes a first neural network module that expresses and learns a first parameter for an input image from a first type to a second type and outputs the learned result as a first learning result, a second neural network module that removes and learns a connection of a part of a second parameter for the input image and outputs the learned result as a second learning result, a prediction module that generates a prediction value for an object of the input image from a summation result obtained by summing the first learning result and the second learning result, and an optimization module that updates the first parameter and the second parameter based on the prediction value.