Neural Network Focus Detection for Phase Difference Pixels

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The existing pupil splitting phase difference detection method faces challenges in accurately calculating the image shift amount due to mismatched phase difference waveforms and sensitivity distributions of phase difference pixels, leading to inappropriate conversion coefficients and inaccurate defocus values.

Innovation Solution

A focus detection device that utilizes a neural network for mechanical learning, processing received light amount distributions from phase difference pixels with different properties to output accurate defocus-related information, including image shift amounts and reliability parameters, thereby improving focus control accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If center-of-gravity calculation is used to detect image shift amount, then calculation speed is improved, but measurement precision deteriorates due to inappropriate conversion coefficients

Engineering Contradiction:

Improvecalculation speedVSAvoidimage shift amount detection accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent replaces the conventional mechanical/mathematical center-of-gravity calculation method with a neural network-based inference system. The neural network learns the complex relationship between phase difference pixel outputs and image shift amounts through training data, substituting traditional mathematical operations with a trained computational model that can accurately determine image shift amounts without relying on potentially inappropriate conversion coefficients.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the approach from using fixed conversion coefficients to using learned parameters from neural network training. By training the neural network on diverse image data with known ground truth labels, the system adapts its internal parameters to capture the actual relationship between phase difference waveforms and image shifts, improving measurement precision while maintaining calculation efficiency.

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If correlation calculation method is used to detect image shift amount, then measurement precision is improved, but device complexity increases and calculation time increases

Engineering Contradiction:

Improveimage shift amount detection accuracyVSAvoidcalculation process complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network on a comprehensive dataset of images with known focus states before actual use. This training phase captures the complex relationships between phase difference waveforms and image shift amounts in advance, allowing the inference process during actual operation to be simplified while maintaining high measurement precision without requiring complex real-time calculations.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent substitutes the computationally intensive correlation calculation method with a neural network inference system. The neural network, having learned the complex patterns during training, can quickly determine image shift amounts through forward propagation, replacing the need for time-consuming correlation computations while maintaining or improving measurement precision.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Adaptability or versatility

If phase difference pixels with different properties are used, then adaptability is improved, but difficulty of detecting and measuring increases due to waveform mismatch

Engineering Contradiction:

Improveimaging target adaptabilityVSAvoidimage shift amount measurement difficulty

Core Design Contradiction:

Adaptability or versatilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent replaces traditional waveform-based measurement methods with a neural network inference system. The neural network is trained to handle phase difference waveforms from pixels with different properties and sensitivity distributions, learning to correctly interpret and measure image shift amounts despite variations in waveform characteristics, thereby reducing measurement difficulty while maintaining adaptability.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the measurement approach from direct waveform correlation to neural network-based inference. By training the network on diverse phase difference data with known ground truth, the system learns to compensate for waveform mismatches and property variations, making measurement easier and more reliable across different imaging targets and conditions.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11503201B2Focus detection device and method

Publication Date: 2022.11.15 SONY GROUP CORP

AI summary

The present technology relates to a focus detection device and method and a program by which an accurate image shift amount can be detected. A calculation unit performs calculation based on learning on the basis of a received light amount distribution of an A pixel group having a first property for phase difference detection and a received light amount distribution of a B pixel group having a second property different from the first property and outputs defocus amount related information relating to a defocus amount. The present technology can be applied to an imaging apparatus that performs focus detection by a phase difference detection method.