Defect Detection Of Cover Glass Based On Deep Learning

The material of the cover glass is specially treated strengthened glass,which is widely used to protect LCD screens.The defect detection of the cover glass in the production process is still mainly using manual visual inspection method.Manual detection efficiency is low,and its false detection and miss detection rate are high.Therefore,using AOI(Automated Optical Inspection)for visual inspection is a better solution.In the process of using AOI visual inspection,digital image processing and machine learning are difficult to meet the classification accuracy requirements.Therefore,researching an efficient,high-accuracy,and highly adaptable cover glass defect detection method is of great significance for accelerating the full automation of cover glass production.Aiming at some process defects in the production process of the cover glass,this paper studies the cover glass defect detection method based on deep learning.A method based on digital image processing is designed to construct the sub-image data set,and the deep learning method is studied to classify the sub-image.The method studied in this paper guarantees real-time detection under the premise of high accuracy.The specific content includes:(1)Defect detection algorithm based on digital image processing and image classification:This paper studies the distribution of the cover glass’s main defects.Accordingly,a feature extraction algorithm based on digital image processing is designed to perform sub-image segmentation,and a sub-image data set containing 9 categories is obtained.Aiming at the classification problem of sub-image,the EfficentNetB0 model is trained and optimized.The classification accuracy obtained by this method is 97.03%.(2)Model optimization algorithm based on image augmentation:Aiming at the model over-fitting problem caused by insufficient data in the classification process of the cover glass sub-image,this paper studies the image augmentation method,and proposes the Mixup-geometric transformation fusion augmentation algorithm.Through the algorithm to increase and expand the sample during the training process,the classification accuracy of the model can be increased to 98.02%.(3)Model compression:The network pruning algorithm is studied.Aiming at the unique MBconv structure in EfficientNet,a corresponding network pruning algorithm is proposed.After applying this algorithm,the model has an accuracy of 97.38%.And the algorithm can compress the model size to about 1/4 of the original model,at the same time it can reduce nearly 3/4 of the parameter amount and about 1/5 of the amount of calculation.Subsequently,this paper performed the calculation graph optimization and FP16 quantization on the model.At the cost of 0.17%inference accuracy loss,the model’s forward inference speed is increased to 7.3 times.The final defect classification accuracy of the deep learning-based cover glass defect detection method can reach 97.21%.The model size is 4.01MB,the parameter amount is 1.04MB.On the computing platform of the 1060Ti graphics card,the average inference time of this model is 1.18ms/pic.