Research On Wood Texture Image Reconstruction Algorithm

Wood has been an important energy consumption in China since ancient times.At the same time,the simple and beautiful wood texture is also paid more attention in the decoration design.It can improve the utilization rate of the board by using the surface mount method for the plate reprocessing,but the printing wood patches needs to obtain enough wood grain patterns.Wood texture image reconstruction has important scientific research significance and practical application value in forestry engineering.This paper introduces image synthesis algorithm and generation model in deep learning to solve the problem of wood texture image reconstruction.Some achievements have been made in the following aspects:(1)In this paper,a texture data set called WSTS?4 including four types of wood(hickory,black walnut,oak and hackberry)is established.The data set covers three types of wood texture images(high structure,certain structure and randomness,high randomness)under the same light source,which has universality and representativeness.(2)This paper presents a new algorithm WTRM for wood texture from the perspective of traditional texture synthesis algorithm.The algorithm can analyze the different texture features of four kinds of wood in the data set WSTS?4,and automatically select the most appropriate mosaic size according to the input image,and then use the improved texture synthesis algorithm to generate a new image according to the input sample texture.(1)The selection of CS threshold for color similarity of different kinds of wood texture is analyzed and the corresponding values of all kinds of wood(hickory-0.88,black walnut-0.92,oak-0.90 and hackberry-0.94)are selected;(2)In the texture synthesis algorithm,the sampling space is expanded and the selection strategy of matching blocks is modified;(3)The weight ? is used to adjust the proportion of cumulative error and minimum splicing cost,so as to improve the effect of image synthesis,and the recommended value is 0.3;(4)Through subjective scoring(up to 80% recognition)and objective indicators(Tenengrad,Laplace,variance function,information entropy,PSNR and SSIM)to evaluate the quality of the generated image,it is proved that the wood texture reconstruction image obtained by the algorithm has better quality.(3)From the perspective of deep learning,this paper proposes a new model of SK?GAN based on improved SinGAN neural network.This model can carry out many image processing tasks,such as quantity amplification,image enhancement,image editing and so on,and has achieved very good results.(1)In the image generation subtask,“Real/Fake” AMT test and the FID of a single image are used to evaluate the experimental results subjectively and objectively.In the scale of n = N,the SIFID value obtained by SK?GAN model is half lower than that of SinGAN model,while in the scale of n =N-1,the value of this method is nearly 0.016;(2)In the image super-resolution subtask,PSNR,RMSE and NIQE are used to evaluate the quality of the generated image from different dimensions.Compared with SinGAN model,PSNR of SK?GAN model increased by 1.03 dB,RMSE and NIQE decreased by3.74 and 1.79 respectively;(3)In the image editing subtask,the image effect is evaluated subjectively by comparing the image results before and after editing.(4)The two methods proposed in this paper can obtain qualified images with less defects and high resolution.They all put forward solutions to the problem of data shortage of wood slice texture image,and successfully completed the reconstruction of wood texture image.Compared with the two methods,the former requires less system resources and time cost,while the latter can obtain higher quality reconstruction images and has a more powerful and comprehensive series of image processing capabilities.