Research On The Detection Method Of Small Scratches On The Surface Of The Chassis Based On Deep Learning

Scratches are one of the most common product appearance defects.Due to factors such as production environment and processing technology,some products may have scratch defects in various shapes.Because of the small scratch size and the low contrast between scratch gray level and background gray level,manual detection is often difficult to accurately detect.Defect detection technology based on machine vision has the characteristics of fast speed,high accuracy,strong stability and low cost.It is widely used in the field of product appearance defect detection.However,the traditional machine vision detection technology is difficult to extract appropriate features for products with complex texture and fine scratch defects,and can not achieve satisfactory detection results.Deep learning extracts the deep features of images through the self-learning ability of neural networks,which improves the accuracy of detection of fine scratches in complex texture products.This dissertation mainly studies the problem of fine scratch detection on the surface of the chassis,and proposes a method for detecting the fine scratches on the surface of the chassis based on deep learning.The main work of this dissertation are as follows:First,the hardware such as camera,lens and light source are selected and compared with different illumination methods.A set of image acquisition system is designed to collect the image of the chassis containing fine scratch defects.Secondly,the histogram characteristics of the scratch image are analyzed,and the traditional segmentation algorithm is studied.It is verified by experiments that the traditional machine vision technology is difficult to solve the problem of fine scratch detection on the surface of the chassis.Then,the structure and characteristics of the deep convolutional neural network are studied.Combining the characteristics of U-net,Leaky-ReLU activation function and residual learning module,a simplified U-network(SU-net),U-Residual-net(URes-net)are proposed.Data samples are augmented using a stochastic elastic deformation method.The comparison experiments between SU-net,Ures-net and U-net proposed in this dissertation verify the feasibility and accuracy of the proposed method in segmentation of scratch images.Finally,the scratch contour information is used to achieve the scratch region localization.K3M algorithm is selected to extract the skeleton of scratches to obtain the scratch length,and the advantages of K3M algorithm are verified by comparative experiments.The error analysis of scratch length detected by this method shows that the relative error of URes-net method is 0.26%,which realizes the qualitative and quantitative detection of the micro-scratch defects on the chassis surface.