An ISP Realization Method Based on Weakly Supervised Learning

Abstract

An ISP implementation method based on weakly supervised deep learning, which includes the following steps: S100: collecting and demosaicing the original RAW image to obtain a demosaiced image; S200: performing brightness adaptation on the demosaiced image Adjust to obtain the brightness-adjusted image; S300: parameter setting and initialization of the network model; S400: generator G generates a predicted high-definition RGB image according to the brightness-adjusted image; S500: reverse generator F according to the prediction Generate a predicted input image from the high-definition RGB image, and then adjust the parameters of the generator G; S600: Input the predicted high-definition RGB image and the target high-definition R6B image into the discriminator at the same time, and use the feedback from the discriminator to adjust the parameters of the generator G.

Description

technical field [0001] The disclosure belongs to the fields of image processing technology and computer vision technology, and in particular relates to an ISP implementation method based on weakly supervised deep learning. Background technique [0002] ISP (Image Signal Processing) is image signal processing, which specifically refers to the post-processing of digital image data output from the CMOS sensor. After correction and enhancement, it is closer to what the human eye sees in reality in terms of color, brightness and clarity. to high-definition real images. The ISP in traditional shooting equipment uses a dedicated digital integrated circuit, and as the core part of the shooting equipment, the ISP undertakes various image processing functions during the shooting process. Therefore, the quality of the ISP determines the imaging quality of the camera to a large extent. [0003] Since 2010, powerful ISPs that can reconstruct high-quality images have appeared for mobile...

AI Technical Summary

Problems solved by technology

Strongly supervised learning requires pixel-by-pixel aligned image pairs for network training, and since the processing effect of ISP has a great relationship with image sensor parameters, the strong supervised learning method needs to collect a large amount of control data sets for each shooting device, which is time-consuming. It is labor-intensive, and the quality of the generated image is limited by the RGB image quality after the camera ISP processing

Images