Research On Optical Film Defect Detection Based On Deep Learning

Optical film is a part of the display.When making optical films,a small number of products will have some defects.Since the traditional image processing algorithms cannot accurately detect the defects in optical films,this paper chooses the optical film defect detection algorithm based on deep learning to solve two practical problems encountered in the process of detecting the defects of optical films.The specific work is as follows:(1)A pooling algorithm based on Gaussian function is proposed to solve the problem that the common pooling algorithms cannot take into account the correlation between each element in a pooling domain and the features contained in the pooling domain.The algorithm first calculates the three parameters of the Gaussian function according to the value of each element in the pooling domain,then uses the Gaussian function to calculate the weights of each element in the pooling domain,and finally calculate the weighted average of the elements in the pooling domain according to these weights.Experiments show that the accuracy of this algorithm is improved by 0.5 to 5percentage points compared with other pooling algorithms on different datasets and models.(2)A class incremental learning algorithm based on placeholder data is proposed to solve the problem that common class incremental learning algorithms are not accurate enough to detect categories.The algorithm first improves the traditional model by adding a binary classifier between the convolutional layer and the fully connected layer to determine whether the input data is a new type,and then aiming at the problem of incomplete data set during the training of the binary classifier,proposes and generates placeholder data to train the binary classifier,and finally the input data is processed differently according to the output of the binary classifier.Experiments show that the algorithm can recognize new defect types on different datasets and models and improves the recognition accuracy by 1 to 6 percentage points compared with other class incremental learning algorithms.The algorithm which is proposed in this paper effectively solves the problems in optical film defect detection,and also has good performance in other datasets,which has certain research significance and application value for optical film defect detection.