Autofocus Image Processing With Attention-Map Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In ultra-low-light environments, existing autofocus (AF) methods struggle with focus detection due to fluctuating evaluation values caused by high noise levels, leading to blurring or failure of AF functions.
Innovation Solution
An image processing apparatus uses a neural network to generate an attention map indicating regions with specific spatial frequencies, reducing noise in images and providing focus adjustment information through a trained machine learning model, enabling accurate autofocus even in high-noise conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If high gain is employed to improve subject visibility in ultra-low-light environments, then subject visibility is improved, but image quality deteriorates due to increased noise
Solution Approach 1:
The patent segments the image processing into multiple stages: initial noise reduction processing, attention map generation, and selective noise reduction. By dividing the processing into stages and applying different processing strengths to different regions (subject vs. background), the system can reduce noise while preserving subject visibility and detail.
Solution Approach 2:
The patent applies different noise reduction strengths to different regions of the image by generating an attention map that identifies subject regions. The noise reduction processing is applied more strongly to background regions while preserving detail in subject regions, achieving local optimization of both noise reduction and subject visibility.
2Object-affected harmful factors
If noise reduction processing is applied to reduce noise levels, then noise is reduced, but focus detection accuracy deteriorates due to loss of high-frequency components
Solution Approach 1:
The patent segments the image into subject regions and background regions using attention map generation. By applying noise reduction selectively to background regions while preserving high-frequency components in subject regions, the system maintains focus detection accuracy while reducing noise in non-critical areas.
Solution Approach 2:
The patent applies noise reduction with different strengths to different regions: strong noise reduction to background regions where high-frequency components are less critical, and gentle or no noise reduction to subject regions where high-frequency components are essential for focus detection. This local differentiation resolves the contradiction between noise reduction and focus accuracy.
3Object-affected harmful factors
If strong noise reduction is applied to improve image quality, then noise is reduced, but autofocus stability deteriorates due to fluctuating evaluation values
Solution Approach 1:
The patent segments processing into stages with an attention map that identifies subject regions. By applying noise reduction selectively rather than uniformly, the system preserves the evaluation values needed for autofocus in subject regions while reducing noise in background regions, maintaining autofocus stability.
Solution Approach 2:
The patent applies noise reduction with spatially varying strength based on the attention map. Regions identified as subjects receive gentler processing to preserve evaluation value stability for autofocus, while background regions receive stronger noise reduction. This local quality differentiation maintains autofocus reliability while improving overall image quality.
Data Source
AI summary
There is provided with an image processing apparatus. A first generating unit generates an attention map indicating a region having a specific spatial frequency from a first image that is captured by an image capturing apparatus. An outputting unit outputs, based on the first image and the attention map, a second image in which noise has been reduced from the first image by using a trained machine learning model. A second generating unit generates, based on the second image and the attention map, information for adjusting focus of the image capturing apparatus.