Anomaly detection of Logo images in the mobile phone using convolutional autoencoder
Muyuan Ke, Chun-Yi Lin, Qinghua Huang · 2017
There are Logo images printed in the mobile phone. Anomaly detection of Logo image is an important quality control task in the intelligent manufacture. In real applications, there are not enough negative samples in the production line for us to study their difference from normal samples. In this paper, we propose an unsupervised learning method based on convolutional autoencoder (CAE) to generate the template of sample and detect the abnormal information through comparing test images with the adaptive template. Firstly, several methods of data augmentation are introduced to expand the scale of positive samples, aiming to improve the performance of CAE. Secondly, the topology of proposed CAE model is introduced. Thirdly, we introduce the image processing methods to detect and locate the abnormal information in the Logo image. A series of experiments on three group of different Logo image have shown that the method we proposed can effectively detect most of the anomalies in the image and achieve the average accuracy of 98.9%.