Camera Exposure Control via Deep-Learning Scene Semantics

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing automated exposure control systems in cameras often fail to provide acceptable image capture results, especially in back-lit situations, as they tend to produce poor contrast on the subject while maintaining good contrast on the background, leading to loss of image data due to incorrect exposure settings.

Innovation Solution

The Exposure Controller employs deep-learning techniques to perform real-time automated camera exposure control by predicting and setting optimal exposure values based on scene semantics, using a combination of pre-trained networks, reward networks, and contextual-bandit networks to bypass hardware metering functions and adapt to changing lighting conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If typical automated exposure control systems use intensity distributions or histograms to determine exposure settings, then the system can process images efficiently, but the system fails to provide acceptable image capture results in back-lit situations by producing poor contrast on the subject

Engineering Contradiction:
Improveexposure processing efficiencyVSAvoidexposure accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent segments the image into different regions (foreground subject and background) and applies different exposure evaluation criteria to each region. The system identifies the foreground region separately from the background and evaluates exposure metrics specifically for the foreground subject, allowing independent optimization of subject exposure even when background lighting conditions would otherwise dictate the exposure settings.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality assessment by evaluating exposure metrics specifically for the foreground subject region rather than the entire image. The system calculates exposure statistics (such as mean intensity and standard deviation) separately for the identified foreground region, enabling localized exposure optimization that prioritizes subject quality over overall scene exposure.

Inventive Principle:
Principle #3Local quality

2Manufacturing precision

If automated exposure systems prioritize subject contrast in back-lit situations, then image quality of the subject improves, but the background becomes washed out with lost image data

Engineering Contradiction:
Improvesubject exposure qualityVSAvoidbackground image data
Core Design Contradiction:
Manufacturing precisionVSLoss of information

Solution Approach 1:

The patent applies partial action by selectively optimizing exposure settings for the foreground subject rather than attempting to optimize the entire scene. The system intentionally allows the background to be under-exposed or washed out as a trade-off, focusing exposure resources on achieving proper subject contrast and detail while accepting information loss in the background region.

Inventive Principle:
Principle #16Partial or excessive action

3Manufacturing precision

If cameras use deep-learning techniques with pre-trained networks and reward networks for exposure control, then exposure accuracy and scene semantics handling improve, but device complexity increases

Engineering Contradiction:
Improveexposure control accuracyVSAvoidsystem architecture complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by using pre-trained neural networks that have been trained offline on large datasets. The complex deep learning models are trained in advance to recognize scene semantics and predict optimal exposure settings. During actual camera operation, the pre-trained network performs rapid inference without requiring real-time training, thus achieving high exposure accuracy while keeping runtime computational requirements manageable.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces a reward network as an intermediary component that bridges the gap between the pre-trained network and the final exposure control decisions. The reward network evaluates the quality of exposure predictions made by the pre-trained network and provides feedback signals that guide the optimization process, enabling the system to learn optimal exposure policies without requiring direct access to ground truth exposure settings during operation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3574644B1Real-time semantic-aware camera exposure control
Publication Date: 2023.07.05 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3574644B1 patent drawingFigure 1
  • EP3574644B1 patent drawingFigure 2
  • EP3574644B1 patent drawingFigure 3~4

AI summary

An "Exposure Controller" provides various techniques for training and applying a deep convolution network to provide real-time automated camera exposure control, as a real-time function of scene semantic context, in a way that improves image quality for a wide range of image subject types in a wide range of real-world lighting conditions. The deep learning approach applied by the Exposure Controller to implement this functionality first uses supervised learning to achieve a good anchor point that mimics integral exposure control for a particular camera model or type, followed by refinement through reinforcement learning. The end-to-end system (e.g., exposure control and image capture) provided by the Exposure Controller provides real-time performance for predicting and setting camera exposure values to improve overall visual quality of the resulting image over a wide range of image capture scenarios (e.g., back-lit scenes, front lighting, rapid changes to lighting conditions, etc.).