16‐3: A Visualization Method of Training Data Completeness in Array Defect Recognition Based on Deep Learning

Abstract

Feature visualization engineering of convolutional neural network is a basic research project in deep learning. The working principle of network, the image features extracted from network and classification basis of images can be revealed, by visualization the extracted features. In this paper, the de‐convolution and CAM method are used to visualize array defect features extracted by CNN. We find that the low‐layer network of CNN is unselective, and its main function is to separate all objects in the picture from their background, this is why the low‐layer networks’ parameters usually do not been trained in transfer learning. Instead, the selectivity is slowly emerging in high‐layer networks and only the object features related to classification are preserved. At the same time, we find that the classification of array defects is not only related to itself, but also to its background, this can also explain why array defect classification is more difficult. On the other hand, this paper verifies the importance of array defect attributes in defect classification by counting the accuracy of network identification after modified the defect attributes, and a defect recognition disturbance rate index was defined to quantify the dependence of defect classification results on its attributes. Finally, the images edited by defect attributes with high disturbance rate are added to the training set, and the defect recognition accuracy is improved by 3% ~ 5%. This method can be used to judge the importance of each array defect attribute, find out the crucial attributes and pinpoint the completeness of training data in CNN classification. At the same time, it can explain the classification principle of CNN and provide necessary guidance and help for attribute image collection and CNN classification accuracy improvement.

In‐line mura detection using machine learning and subspace method in display manufacturing

“…In recent years, the use of machine learning has also been progressing in the display field. The utilization of machine learning in the display field is divided into those used for test and evaluation, [7][8][9][10][11][12][13][14][15][16][17][18][19] and those used for development and design of display products. [20][21][22][23][24] Among these previous works, the following papers are the latest and closely related to this work, which deals with mura detection using machine leaning.…”