|
1 | | -# Unsupervised-Defect-Segmentation |
2 | | -Unsupervised-Defect-Segmentation anomaly detection |
| 1 | +# DeeCamp-10Group-Unsupervised-Defect-Segmentation |
| 2 | + |
| 3 | +# Abstract |
| 4 | +Deep learning is well established in the mainstream of computer vision, but it relies on large amounts of data. |
| 5 | +especially in the industrial field, such as product surface defect detection task. Mainstream neural networks framework |
| 6 | +(detection, semantic segmentation) must have defect samples (all defect types to be detected) for training, and the number |
| 7 | +of defect samples should not be too small. Although defect learning is relatively easy, heavy defect collection and labeling |
| 8 | +work is introduced. Moreover, in some production processes in the industrial field, it is difficult to provide sufficient |
| 9 | +defect samples for training with a low defect rate, and it is difficult to meet the 'data-hungry deep'learning model.In contrast, |
| 10 | +there are plenty of positive samples in industrial production that have no defects at all. We use this characteristic |
| 11 | +of industrial production to gain the ability to detect bad samples by learning only good samples. This defect detection technology |
| 12 | +based on positive samples can be applied to AI at low cost in more and more complex industrial scenarios, so it has great practical significance. |
| 13 | + |
| 14 | +# Introduction |
| 15 | +In this work, we focus on unsupervised defect segmentation for visual inspection. |
| 16 | +The goal is to segment defective regions in images after having trained exclusively on |
| 17 | +non-defective samples. It has been shown that architectures based on convolutional neural networks (CNNs) such |
| 18 | +as autoencoders or generative adversarial networks can be used for this task. |
| 19 | +The input picture x into the autoencoder or GAN, get the restored image y. |
| 20 | +Then compare the features of each pixel of x and y, where the feature difference between x and y is large, that is |
| 21 | +the defect. |
| 22 | + |
| 23 | +# Architecture |
| 24 | + |
| 25 | + |
| 26 | +# Directory |
| 27 | +- `/config`: This repo includes configuration file for training and evaluation. The parameters are saved in `.jason` file. |
| 28 | +- `/db`: This repo includes codes of data set tools |
| 29 | +- `/model`: Our networks codes are in `model/netwroks` and segmentation codes are in `mode/segmentation`. |
| 30 | +- `/tools`: This repo includes codes of auxiliary function just like timer, log writer and so on. |
0 commit comments