Neural Network Estimation Confidence via Wavefront Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks may produce low estimation accuracy for three-dimensional optical characteristics of objects, making it difficult to determine the reliability of estimated results, especially when objects are small or have optical characteristics not present in training data.
Innovation Solution
A data processing method involving a neural network that inputs measurement data, generates estimation data, performs forward propagation to obtain wavefronts passing through the object, and calculates a confidence level based on measurement and restoration data, using trained models and multiple training data sets to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a deep neural network is used to estimate three-dimensional optical characteristics from measurement data, then the processing speed is improved, but the estimation accuracy deteriorates making it difficult to determine reliability
Solution Approach 1:
The patent introduces restoration data as an intermediary element that mediates between the measurement data and the estimation data. The restoration data is generated from the estimation data through forward propagation operations, and serves as a reference to evaluate the reliability of the estimation. This intermediary allows the system to maintain fast processing through the neural network while providing a mechanism to assess estimation accuracy by comparing restoration data with original measurement data.
Solution Approach 2:
The patent implements a feedback mechanism where the confidence level calculation uses both measurement data and restoration data to evaluate estimation reliability. The system calculates a confidence level based on the comparison between measurement data and restoration data, and this confidence information can be used to adjust the neural network's learning process. This feedback loop enables the system to maintain high processing speed while continuously improving and verifying estimation accuracy.
2Speed
If a trained neural network model is used for fast image quality improvement, then processing speed is improved, but reliability of results deteriorates when objects are small or have optical characteristics not present in training data
Solution Approach 1:
The patent performs preliminary action by generating restoration data from the estimation data before final output. This restoration data serves as a preliminary verification step that checks whether the estimation is reliable. By performing this restoration and comparison operation beforehand, the system can identify unreliable estimations (particularly for small objects or objects with optical characteristics not present in training data) before they are accepted as final results, thus maintaining reliability without sacrificing the speed advantage of using a trained neural network.
3Reliability
If confidence level calculation is added to verify estimation reliability, then reliability of results is improved, but device complexity increases
Solution Approach 1:
The patent applies universality by using the existing neural network infrastructure for multiple functions. The same neural network that performs the primary estimation task is also utilized to generate restoration data through forward propagation. This means the existing computational resources are leveraged for the additional reliability verification function without requiring completely separate hardware or computational systems, thus improving reliability while minimizing the increase in device complexity.
Data Source
AI summary
A data processing method includes an input step S1 of inputting measurement data into a neural network, an estimation step S2 of generating estimation data from the measurement data, a restoration step S3 of generating restoration data from the estimation data, and a calculation step S4 of calculating a confidence level of the estimation data, based on the measurement data and the restoration data. The neural network is a trained model, the measurement data is data obtained by measuring light transmitted through an object, the estimation data is data of a three-dimensional optical characteristic of the object estimated from the measurement data, and the three-dimensional optical characteristic is a refractive index distribution or an absorptance distribution. In the estimation, the neural network is used, in the restoration, forward propagation operations are performed on the estimation data, and in the forward propagation operations, wavefronts passing through the interior of the object estimated from the measurement data are sequentially obtained in a direction in which light travels.