This software uses neural networks to process detector images, counting and classifying the defects it finds.
The motivation and full description of the project can be found here.
labelling_tools contains tools to annotate the data manually, marking the position of the defects in the images, as well as classifying them. The Matlab software code is under ClassLabelingTool.
generate_mirror_csv.mconverts the labels to csv format.
nn contains scripts to auxiliate training.
matlab_integration contains python scripts to get predictions over a set of images, and also matlab scripts that run the python scripts.
utils contains utility tools for other classes to use.
csv contains csv files used in training and testing.
Our approach to detect defects is based on keras-retinanet 0.2. It uses Keras (2.1.3) on top of Tensorflow (1.5.0). Other requirements: tox (2.9.1), numpy (>= 1.14), OpenCV (3.3.0), Pillow, keras-resnet, cython, matplotlib, h5py, pandas, setGPU.
A detailed tutorial on submitting training to Ukko2 and using the labeling tool can be found in docs/tutorials.md.
