Research On The Ceramic Tile Surface Defect Detection Based On Deep Learning

Ceramic tile is one of the important decorative materials in the construction industry.It is widely used in daily life.Due to the complexity of the ceramic tile production process,various defects inevitably occur during the production process.Serious defects can affect the quality and aesthetics of ceramic tile products,resulting in damage to the interests of relevant enterprises and even potential safety issues.Currently,most ceramic tile quality inspections are conducted manually,which is costly,inefficient,and susceptible to subjective factors.Therefore,how to efficiently and accurately detect ceramic tile surface defects has become a technical issue that has plagued the industry for many years.With the development of computer vision technology,manufacturing industry is gradually moving towards the direction of intelligence.Target detection algorithms based on deep learning are gradually applied to the field of industrial defect detection,improving detection efficiency and reducing detection costs.Therefore,this thesis combines ceramic tile surface defect detection with target detection algorithms based on deep learning,and proposes a ceramic tile surface defect detection algorithm based on improved Faster R-CNN and a ceramic tile surface defect detection algorithm based on improved YOLOv5,respectively,to improve the detection accuracy of the model for ceramic tile surface defects.The main research work of this thesis is as follows:Firstly,analysis and processing of the ceramic tile surface defect dataset.For the problem of large image resolution in the original dataset,a sliding window slicing method is used to slice the original dataset,reducing the original image resolution,enhancing model robustness,and making the detection target more prominent.Secondly,research on the performance optimization of the Faster R-CNN algorithm.The anchor generation parameters are improved and optimized to fit the target scale more accurately and make the positioning more accurately based on the characteristics of the ceramic tile defect dataset.Rank & Sort Loss is introduced to optimize the loss function,which can reduce the number of hyperparameters and improve model performance.Deformable convolution is introduced in the last three stages of the feature extraction network,resnet101,to adaptively learn flaw features.After introducing the above three improved strategies,the model obtained the optimal results,with an average accuracy of 76.3%,which is 17.9 percentage points higher than the original model,verifying the effectiveness of the improved strategy.Thirdly,research on the performance optimization of the YOLOv5 algorithm.CBAM,CA,and Sim AM attention mechanisms are added to the middle of the SPPF module and its previous layer of C3 module,respectively,to make the model more focused on flaw target characteristics.After comparing and analyzing the performance of the model with adding the above three attention mechanisms,CBAM attention mechanism is selected as the basis for subsequent experimental optimization because it can make the model perform best.Due to the SPD-Conv with better feature learning ability,this thesis replaces the convolution with a step size of 2 in YOLOv5 with SPD-Conv to improve the detection improve the detection effect of the network on small defect targets.The positioning loss CIo U of YOLOv5 is replaced with SIo U to optimize the loss function,which can accelerate the regression convergence speed of the prediction frame and improve the model detection performance.Finally,after adding the above three improved strategies,the average detection accuracy of the model m AP is 84.1%,which is 3.6 percentage points higher than the original model.This verifies the effectiveness of the improved strategy,indicating that the improved algorithm based on YOLOv5 studied in this thesis can better detect ceramic tile surface defects and meet the requirements of ceramic tile surface defect detection.