Image Restoration For Motion Blur In Real-Time Detection System Of Veneer Sheets Surface Defects

Based on machine vision,the real-time detection system of veneer sheets surface defects,whose types and areas are detected with the veneer sheets image,evaluates the grading of veneer sheets.The moving veneer sheets on the conveyor make the collected image blurred,resulting in the inaccuracy of the real-time detection system.Image restoration is an effective solution to this problem without costly hardware.However,the complex noises and the small motion blur kernel in moving veneer sheets images cause disappointed results of the traditional restoration.In order to efficiently restore blurred image of moving veneer sheets,this paper studies the image restoration based on neural network.The main work is as follows:1.The image dataset of veneer sheets was established.After analyzing the demand of industrial production and scientific research,the image acquisition system of the veneer sheets was designed.Following the hardware selection and software development according to system design,the veneer acquisition system sheets images was set up.The veneer sheets for the experiment was produced by the rotary veneer sheets production line.Then the images of veneer sheets were collected by the image acquisition system.Processed by cropping,rotating,labeling and classifying,the images composed the dataset including 2080 pairs of blurred and sharp images of veneer sheets.2.Traditional image restoration and neural network-based image restoration were studied.The traditional image restoration based on the approximate L0 norm solves the blur kernel iteratively and then obtains the restored image by deconvolution.Whereas solving blur kernel is an ill-posed problem because of complex noises in the images of real scenario.Therefore it is difficult to have expected results in restoring images of veneer sheets with small blur kernel.But the multi-scale convolutional neural network(MSCNN)directly solves the nonlinear mapping between the blurred image and the sharp image in an end-to-end method,which is indicated as experimental comparison that the effect in restoring images of veneer sheets at different moving speeds is improved.3.A modified image restoration based on MSCNN was proposed.The stacking of Res Blocks at each scale in the MSCNN was optimized by encoder-decoder network based on Res Block,which expand the receptive field.Meanwhile Conv LSTM was added in each scale to improve the utilization of features.The experimental results showed that the modified method could improve the effect in restoring images of veneer sheets at different moving speeds.In this paper,a modified image restoration based on MSCNN was proposed to solve the blur of moving veneer sheets image in the real-time detection system of veneer sheets surface defects,which provides a theoretical support for improving the production quality of wood products.