Research On Automatic Grayscale Rendering Based On Deep Learning

As a very important part of the automotive production cycle,automotive styling also needs to improve efficiency.After finalizing the proportion of the vehicle,the styling designer needs to draw sketches and renderings.The renderings are used as an intuitive way to show the shape of the car,and as a basis for the creation of the digital model of vehicles,they are used as a guide for the digital model.However,it takes a skilled designer a long time to draw the rendering,and in the output process,it is affected by the designer’s personal design experience and aesthetic cognition,thus presenting different effects.With the popularity of digital rendering,traditional hand-drawn renderings are more often used in the concept car development stage,and how to make the renderings more efficient and allow the stylist’s design style to be inherited becomes a difficult point to shorten the pre-development cycle of the car.Therefore,the research topic of this paper is the style migration from design sketches to hand-drawn renderings in grayscale.The main research work of this paper is as follows.1.This paper filters the rendering sample pictures to improve the data quality by determining the dataset specifications.However,limited by the small dimensionality of the automotive hand-drawn sketch dataset,the insufficient number of samples and the large style differences.This paper proposes a method to apply an improved Sobel edge detection method for edge extraction of the rendering dataset samples to obtain the same number of hand-drawn style-like samples as the rendering dataset。2.Since the grayscale rendering generated using traditional Pix2 Pix neural networks and Cycle GAN(Cycle Generative Adversarial Networks)are poorly implemented,this paper proposes a synthesis method N-Cycle GAN for grayscale rendering based on Cycle GAN networks.3.Finally,PSNR and SSIM picture quality evaluation methods are used to compare the picture quality among the three networks.The styling designers subjectively evaluate the specific elements of the car body,and review the grayscale renderings generated from six angles.The main innovations of this paper are as follows: the quantitative specification of the dataset construction and the quantitative evaluation of the generated rendering are established,the Sobel edge detection method using linear light is proposed,and the two data augmentation methods are applied to the Cycle GAN network,and the N-Cycle GAN network is proposed.The research results will effectively increase the rendering speed of designers and shorten the design cycle.