Defect_Segmentation

This repository contains a model to detect defects by Industrial Optical Inspection on Textured Surfaces.

Data

• The first six out of ten data sets are used for training while the last 4 are used for testing.
• The training set consists of 1000 non-defective and 150 defective images while the test set consists of 2000 non-defective and 300 defective images.
• The images are saved in grayscale 8-bit PNG format.
• The code makes use of a compressed version of the data which can be found here
• Actual data can be downloaded from here

Model

• The underlying model is based on the following paper U-net: Convolutional Networks for Biomedical Image Segmentation