Image Classification for Machine Maintenance by CNN with Polarized Filtered Images

Abstract

The application of AI to the maintenance of various types of machinery is expanding. In image-based diagnosis, convolutional neural networks (CNNs) are sometimes used, in which a digital filter is convolved with an input image to filter it, creating multiple images with different characteristics for analysis by the CNN. In this study, we investigate the possibility of improving the accuracy of NNs by using analog filters in addition to digital filters. Specifically, three types of analog filters were created by placing a cellophane film between polarizing plates at different angles. The filters were placed in front of a camera, and multiple images obtained through the filters were used as input images for the CNN. In this study, stickers imitating rust were placed on walls and pillars, and a CNN was created to determine the presence or absence of rust. We compared the accuracy of the NNs when the images were taken without filters and analyzed using only digital filters, and when the images taken through analog filters were also input and analyzed in conjunction with digital filters. As a result, we confirmed that accuracy was improved when an analog filter was used and the pooling layer was not placed immediately after the analog filter input.