Neural Network Sensor Parameter Optimization via Backpropagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional technologies face difficulties in optimizing the operation parameters of sensors used in neural network models for improved identification and recognition accuracy, particularly in deep learning applications for autonomous vehicles and robots, due to limitations in image formation models and approximation methods.
Innovation Solution
A method involving training a first neural network model with a first operation parameter and second sensing data to generate first sensing data, then connecting it with a second neural network model to optimize the operation parameter through backpropagation, allowing for the adjustment of sensor parameters without relying on specific input device models, thereby enhancing identification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional commercially available cameras are used for deep learning input, then the camera is optimized for human viewing, but the identification accuracy for deep learning tasks deteriorates
Solution Approach 1:
The patent changes the optical parameters of the camera system by introducing intentional optical aberrations (chromatic aberration, astigmatism, spherical aberration) and using coded aperture masks instead of conventional lens designs. These parameter changes transform the camera from being optimized for human viewing to being optimized for providing discriminative features needed by deep learning models for accurate object identification and depth estimation.
2Device complexity
If conventional image formation models are used, then the modeling is simpler, but the optimization of sensor operation parameters for neural network input deteriorates
Solution Approach 1:
The patent implements a feedback loop where the neural network model's identification results are used to evaluate and optimize the sensor operation parameters. The system trains the neural network with images captured under different sensor parameters, then uses the model's performance feedback to iteratively adjust and optimize parameters such as aperture patterns, focal length, and optical aberrations, achieving parameter optimization that conventional methods cannot accomplish.
3Speed
If approximation methods are used in image formation modeling, then the computation is faster, but the identification accuracy of the neural network model deteriorates
Solution Approach 1:
The patent performs preliminary optimization of sensor operation parameters during the training phase using comprehensive computational methods. By pre-training the neural network model with optimally configured sensor parameters and pre-processing images under these optimized conditions, the system achieves high identification accuracy while maintaining reasonable computation speed during actual deployment, avoiding the need for complex real-time calculations.
Data Source
AI summary
A third model training part trains a second neural network model by backpropagation using an error difference between: an identification result which a third neural network model including a trained first neural network model and the second neural network connected to each other outputs after receiving second sensing data and a first operation parameter, and correct identification information corresponding to the second sensing data. A second operation parameter acquisition part acquires a second operation parameter by updating the first operation parameter via the first neural network model by the backpropagation.