HDR Image Generation via Deep Learning Alignment and Attention Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current high dynamic range (HDR) image generation techniques often suffer from ghosting due to hand tremor, object movement, and other jitter issues when capturing images with both high-brightness and low-brightness areas, leading to suboptimal results.
Innovation Solution
A device and method utilizing a deep learning architecture, specifically a homography network trained through a convolutional neural network (CNN), to align and merge multiple low dynamic range (LDR) images captured at different aperture values or shutter speeds, employing an attention mechanism to weight and combine image regions based on their brightness and association with objects, thereby generating a high-quality HDR image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If multiple images are taken and superimposed to form HDR image, then dynamic range is improved, but ghosting occurs due to hand tremor and object movement
Solution Approach 1:
The system performs preliminary alignment of multiple captured images before merging them to create the HDR image. By pre-aligning the images to compensate for hand tremor and object movement, the system eliminates ghosting artifacts while maintaining the dynamic range benefits of multi-image superposition.
Solution Approach 2:
The system uses a deep learning architecture that processes alignment information as feedback to adjust the merging process. The attention mechanism selectively weights different image regions based on their quality and relevance, allowing the system to adapt to movement and tremor conditions while producing high-quality HDR output.
2Measurement precision
If deep learning architecture with attention mechanism is used, then alignment accuracy is improved, but device complexity increases
Solution Approach 1:
The deep learning architecture is segmented into functional modules: a homography network for alignment estimation, an attention mechanism for selective weighting, and a merging network for HDR synthesis. This modular segmentation allows the complex processing to be broken down into manageable stages, improving alignment accuracy while enabling more efficient implementation.
Solution Approach 2:
The system changes parameters dynamically during processing, including adjusting the degree of attention weighting and merging coefficients based on input image characteristics. This adaptive parameter adjustment allows the system to achieve high alignment accuracy without requiring excessive computational complexity for all possible scenarios.
Data Source
AI summary
A method for generating images with high dynamic range (HDR) based on multiple images captured at different aperture values, under different conditions, or at different shutter speeds is applied in a device. The method inputs the original multiple images into a predetermined model and aligns the multiple images. The method further confirms object images that need to be attended among multiple aligned images and obtains a merge weighting for each of the object images, and merges the images for a generated HDR according to the merge weighting of each image. The device utilizing the method is also disclosed.