Night Image Enhancement Technology For Automatic Driving

Enhancing low-light images is a complex but rewarding task with essential appli-cations in many fields.For autonomous driving,capturing images in low-light environ-ments severely affects visual performance and can create a host of problems,such as high noise and loss of image detail.For simultaneous localization and mapping(SLAM),low-light images will severely affect the localization and map building process.Be-sides,many advanced tasks,such as image instance segmentation and depth estimation,will be severely affected.These problems will seriously affect the driving safety of self-driving vehicles.Over the past few decades,researchers have developed several theories to achieve low-light image enhancement.With the use of neural network techniques,image low-light enhancement models have also achieved good results,but training mostly requires a large number of low-light paired images,making it difficult to use them widely.In this paper,a multi-discriminator generative adversarial network is proposed to recover the color and texture features of images.The discriminator consists of three branches.The first branch is a multi-scale discriminator,which correctly checks the image features at different scales.The second branch is the color discriminator,which serves to determine whether the colors in the generated image are realistic or not.The third branch is a texture discriminator used to determine the sharpness of the generated image edges.In addition,we adopt the structure that the generator and discriminator share the encoder,and this approach greatly reduces the number of model parameters and speeds up the training.In addition,we designed an attention module to focus on information-rich regions.The results of a series of experimental studies show that our method outperforms the state-of-the-art methods in terms of the quality of the generated images and all metrics.The experimental results show that our model can effectively improve the autopilot effect in low-light environments,such as SLAM repositioning and detecting drivable regions under severe light changes.