Demosaicing Method Based on Backpropagation Neural Networks

High-Quality Color Demosaicing Method Using Backpropagation Neural Network Weights and Residual Refinement

This technology is a demosaicing method that restores full-color images from mosaic images captured by a single image sensor's color filter array (CFA). It uses weights learned via a backpropagation neural network to first interpolate the green channel, reconstructs the red and blue channels based on color differences, and then refines each channel using residual images relative to the green channel.

To reduce cost and size, digital cameras use a single sensor covered by a color filter array, which records only one color per pixel, necessitating interpolation for missing colors. Existing edge-detection and color-difference-based interpolation techniques, such as ACPI, ECI, and AHD, rely on fixed rules, which often leave aliasing artifacts along the edges of the restored image.

This technology applies neural network weights trained on an image dataset to both the green interpolation and residual image refinement stages, incorporating inter-channel correlations based on data. This effectively reduces edge aliasing and improves image quality. It can be integrated into the image signal processing (ISP) pipelines of smartphones, digital cameras, and automotive cameras, enabling sharp, full-color images even with low-cost single sensors.

‍

‍

Key Features:

• Reconstructing the green channel by acquiring missing green pixels from a mosaic image through interpolation based on weights obtained by training a backpropagation neural network on an image dataset
• Reconstructing the red and blue channels by interpolating missing pixels based on the color difference between red and green, and blue and green
• Obtaining a residual image by subtracting the intermediate red channel, formed via linear interpolation, from the reconstructed green channel, and refining it using neural network weights
• Obtaining the refined pixel values for the red and blue channels by combining the refined residual images with the reconstructed green channel

‍