Research And Implementation Of Ceramic Tile Defect Detection And Recognition Based On Deep Learning

In industry,defect detection is a very important application.At the same time,deep learning has made great achievements in feature extraction and location.Therefore,convolutional neural network has made great progress in the field of target detection,which makes ceramic tile surface defect detection a potential research direction.However,ceramic tile defect detection is still plagued by many small and medium-sized targets,changeable and irregular shapes,unclear characteristics and other factors.Enterprises still can not avoid producing ceramic tiles with various defects in the process of production and manufacturing.An urgent problem to be tackled is how to solve the foregoing problems and increase the accuracy of small target defect detection.This thesis mainly solves the above problems from two aspects:enhancing the original data set and enhancing the feature extraction ability of the network.Considering that the image size in the tile defect data set is large,but the target defect is small,it is mainly solved from the two aspects of data enhancement of the original data set and enhanced feature extraction ability of the network.Firstly,the data enhancement is carried out by means of image cropping,and the experimental verification is conducted out using both machine learning and deep learning techniques.respectively,and the results show the effectiveness of data enhancement.Secondly,Faster RCNN is selected as the target detection network,the original network structure is replaced by the distracting network,and the network feature extraction ability is strengthened,and in order to further extract multi-scale image features,the recursive feature pyramid network(RFPN)module is introduced to fuse the high and low layer feature information,and due to the variable and irregular shape of the target defect,the deformation convolution network(DCN)is introduced for target detection.Finally,after experimental verification,the mean average accuracy and accuracy of the three network models based on the replacement characteristics of Faster RCNN are scattered network(network 1),the replaced Faster RCNN fusion RFCNN feature pyramid(network 2)and the average accuracy and accuracy of the three network models based on the variable convolution Faster RCNN into the RFPN(network 3),and the experiment shows that after the optimized network,the m AP of network 2 is 1.61%higher than the m AP of network 1.This thesis shows that the introduction of the RFPN module extracts more image features,and the m AP value of network 3 is 1.23% higher than that of network 2,indicating that the detection accuracy of the network is improved after the introduction of DCN.Considering that the prior art is limited in traditional industrial manufacturing,the slow and subjective nature of manual testing products will affect the accuracy of the test results.Based on the proposed improved model,this thesis designs and implements the tile defect detection system,including login registration,image detection and background management,which can identify and record the ceramic tile defect data collection of the imported system and promote the intelligent transformation of enterprises.