Convolutional Neural Network for High Dynamic Range Image Processing

The patent divides the image processing task into multiple specialized branches: a first branch processes over-exposed regions to recover details, a second branch processes under-exposed regions to enhance visibility, and a third branch handles normally exposed regions. This segmentation allows each branch to be optimized for its specific exposure condition, improving overall dynamic range enhancement capability while managing complexity through modular design.

The patent applies different processing strategies to different regions of the image based on their exposure characteristics. Over-exposed regions receive one type of processing (detail recovery), under-exposed regions receive another (enhancement), and normally exposed regions receive a third type of processing. This local quality approach ensures that each region is processed appropriately for its specific conditions, maximizing detail recovery across the entire image.

The patent applies preliminary processing steps to identify and isolate over-exposed and under-exposed regions before applying the main enhancement algorithms. By pre-segmenting the image into different exposure zones and preparing appropriate processing paths for each, the system recovers details more effectively while minimizing information loss in extreme exposure regions.

The patent introduces intermediate processing stages that act as mediators between the input image and final output. These intermediate steps include generating confidence maps, creating processed versions of over/under-exposed regions, and blending results from multiple branches. These intermediaries enable gradual refinement and preserve details that would otherwise be lost in extreme exposure conditions.