Motion-blurry Image Restoration Method for Detecting Surface Defects of Wood Veneer

Abstract

The detection of veneer surface defects is of great significance to wood veneer material evaluation, quality control, and product classification in the wood processing. When the high-speed moving veneer image is collected on the conveyor belt, the image appears blurred, making it difficult to accurately identify the defect type and estimate the defect area. To solve this problem, this study compared three image restoration methods including unnatural L0 sparse representation (L0), multi-scale convolutional neural network (MSCNN), and scale-recurrent convolutional neural network (SRCNN). To perform the comparison analysis, a wood veneer image acquisition system was developed and it provided a wood veneer image dataset with 2,080 groups of blur-clear veneer image pairs. Analysis results showed that the SRCNN method performed better than the other two methods. At four different wood moving speeds, the peak signal to noise ratio (PSNR) of the SRCNN was 4.64%, 14.63%, 18.48%, and 25.79%, higher than the other two methods and structural similarity (SSIM) was less than 2%. The average time for this algorithm to restore a blurred wood veneer image was 13.4 s. The findings of this study can lay the foundation for the industrialized detection of wood veneer defects.