Surface Defect Detection Method Of Valve Core Porcelain Sheet Based On Deep Learning

The ceramic valve core is an important part of the faucet,which plays the role of water volume control and water temperature control.Defects such as breakage,chipped corners,chipped edges,cracks,dirt or incomplete polishing will occur in all links of its production,resulting in water leakage in the final product during use.Therefore,it is necessary to detect surface defects before assembly.However,there is still a big gap between the existing traditional visual detection algorithms and the actual production requirements in the aspects of miss rate,false detection rate and robustness.Aiming at the shortcomings of existing detection algorithms,a visual detection method for surface defects of ceramic valve core based on deep learning technology was proposed.First,this work proposes the strategy of pre-extracting suspicious regions and then using convolutional neural networks for their defect identification to suppress the occurrence of false detections.We align the target to a standard pose by aligning a large number of positive sample images so that the target has the same position and orientation in each image.The probability of the possible occurrence of grayscale values at each pixel location is analyzed for all positively sampled images after positional adjustment.On this basis we set the interval of normal variation of gray value at each pixel position to provide a basis for identifying suspicious pixels during detection.It not only reduces the computational effort of detection,but also suppresses the chance of easy false detection near surface edge features.In order to provide training set,validation set and test set data for the convolutional neural network.Our work designs an auxiliary labeling algorithm,which greatly reduces the workload of manual labeling while ensuring the correctness of label data.In addition,the labelled data is expanded by means of flipping and adding noise to the sample images to enhance the generalisation ability of the model.The VGG16 network is selected as the pre-training model,which is improved and migrated.The migrated model is further trained and optimised using the labelled dataset to obtain the final defect recognition model.Finally,we performed experimental verification on the proposed detection method.This thesis designed and built an image acquisition platform for the valve core tiles,and carried out detection experiments on the collected 400 sample images containing defects.The experimental results show that the false detection rate is 1.8%,the missed detection rate is 1.3%.It shows that it basically meets the requirements of actual production.The research results of this paper have practical value and can provide reference for similar research.