| Summary: | 碩士 === 國立中山大學 === 機械與機電工程學系研究所 === 107 === In recent years, due to the rapid development and popularization of computer software and hardware and image capture technology, Automatic Optical Inspection (AOI) technology is also widely used in various industries, and with the change of industrial structure, the first one is In order to reduce the demand for manpower, the issue related to the automatic detection of computers has attracted more attention. It is hoped that the work of manual visual inspection can be replaced by computer automation, which reduces the cost and improves the detection speed and detection. confirmation rate.
The main purpose of this paper is to establish a neural network model that can effectively assist the optical automatic inspection system for defect detection. Combined with the highly scalable and easy-to-access literal language of Python software, the Convolutional Neural Network (CNN) is used. In the main architecture, for the detection system to establish the problem of insufficient sample size in the early stage, the concept of Transfer Learning and the data enhancement function built in Python software is proposed to improve the number of picture samples used for model training. To make the training process faster,achieve convergence and significantly reduce model training time,are also performed in the proposed system.
In order to improve the prediction accuracy, classic models such as VGG, ResNet and DenseNet, which have been developed in recent years, are used as the basic model of training. Finally, the prediction results of the three models are comprehensively compared to obtain the final prediction results. The obtained results can obtain a prediction accuracy of 98% to 99% in distinguishing between two classifications with or without defects and multi-classifications for distinguishing defect types.
|
|---|
