Extreme Low-light Imaging Based On Improved U-net

Low-light imaging is an important aspect of image processing.Due to the low number of photon count and low SNR,extreme low-light imaging is challenging.So far,there have been many image enhancement algorithms that can be used for low-light imaging,such as histogram equalization,gray-scale transformation,Retinex algorithm,and compensation methods based on the perceptual characteristics of the human visual system.However,the effectiveness of these methods in extremely low light conditions is limited.The popular convolutional neural networks now play a role.Compared with traditional low-light imaging methods,convolutional neural networks have a strong learning ability by simulating biological brain structure,can handle complex mapping processes between input and output,and have been widely used in image processing.The algorithm in this paper is used to process sensor data under low light conditions and train deep neural networks to learn the image processing flow of low light raw data,including color transformation,demosaicing,noise reduction and image enhancement.Train the neural network end-to-end to avoid noise amplification and error accumulation.How to improve the U-net neural network to enhance its effect in the field of imaging in extremely dark conditions is the research focus of the paper.The main research content of the paper is as follows:1)This article first summarizes traditional low-light imaging methods,such as histogram equalization and gray-scale transform methods,and analyzes the problems of tedious steps,low signal-to-noise ratio,and poor visual effects of these methods in imaging in extremely dark conditions.Then,the important role of convolutional neural network in imaging of extremely dark conditions is derived.2)Next,the overall process of imaging in extremely dark conditions based on a convolutional network is explained.Train a U-net network to directly process low-luminance images in fast imaging systems.This method can not only suppress blind spot noise and achieve color conversion,but also convert images from Bayer Raw format to RGB format.It then explains the input processing steps before entering the network,including converting the input to four channels,eliminating black pixels,and enlarging the input data by the desired multiple.3)Then the article introduces the U-net network structure,and proposes corresponding improvements and sources of inspiration.Then it describes the process and trade-offs of improving the U-net network.Finally,the experimental results are emphasized.The experiments show that the U-net can effectively improve the structural similarity and peak signal-to-noise ratio of imaging results in very dark conditions.