Autofocus Neural Network Lens Position Prediction
Solution Overview
Problem
Current autofocus methods, such as CDAF and PDAF, face challenges in accuracy and efficiency, with CDAF being time-consuming and prone to inaccuracy, and PDAF having lower performance and phase errors in low light conditions.
Innovation Solution
A processor-implemented method using a neural network model to determine lens position by generating input data with encoded phase information, lens position information, and region of interest (ROI) data, allowing for precise autofocus prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If contrast detection autofocus (CDAF) method is used to obtain image frames while moving a lens, then the definition of the obtained image frames can be maximized, but the method becomes time-consuming and prone to inaccuracy
Solution Approach 1:
The patent replaces the mechanical trial-and-error lens movement method of CDAF with a neural network-based computational system. The neural network directly predicts the optimal lens position from phase information and image data, eliminating the need for physical lens movement and multiple image frame captures, thus resolving the contradiction between accuracy and time consumption.
Solution Approach 2:
The patent performs preliminary encoding of phase information and lens position information into structured input data before neural network processing. This pre-processing step prepares the data in an optimal format for rapid prediction, enabling the system to achieve high accuracy without time-consuming iterative adjustments.
2Speed
If phase detection autofocus (PDAF) method is used to separate light incident on a camera, then the autofocus speed is improved, but the performance and accuracy deteriorate in low light conditions
Solution Approach 1:
The patent merges phase information from PDAF with intensity information from the image sensor into a unified neural network input. This combination allows the system to leverage the speed advantage of PDAF while compensating for its low-light performance deficiencies through the additional intensity data that the neural network can utilize for more accurate focus determination.
Solution Approach 2:
The patent transforms the raw phase information and image data into encoded neural network input data with specific parameter structures. This parameter transformation enables the neural network to process the information in a way that maintains speed while improving accuracy, particularly in challenging lighting conditions where traditional PDAF struggles.
3Measurement precision
If neural network model is used to determine lens position by processing encoded phase information and lens position information, then the autofocus performance and accuracy are improved, but the device complexity increases
Solution Approach 1:
The patent introduces encoded neural network input data as an intermediary representation between the raw sensor data and the final lens position determination. This intermediate encoding layer simplifies the neural network's processing task while maintaining high accuracy, effectively managing the complexity by structuring the data flow in a manageable way.
4Measurement precision
If region of interest (ROI) position information is encoded and input to the neural network model, then the autofocus precision for specific objects is improved, but the data processing complexity increases
Solution Approach 1:
The patent applies local quality encoding by specifically processing and encoding the position information of the region of interest separately from the rest of the image data. This allows the neural network to focus computational resources on the important ROI areas, improving focus precision for specific objects while managing overall data processing complexity through selective processing.
Data Source
AI summary
A processor-implemented method includes: generating first input data comprising phase information of an input image; generating second input data in which lens position information is encoded; and determining position information of a lens corresponding to autofocus by inputting the first input data and the second input data to a neural network model.