The following sections provide greater details of the dataset, running training and inference, and the training results.

Command line options

To see the full list of available options and their descriptions, use the -h or --help command line option, for example:

python main.py --help

The following mandatory flags must be used to tune different aspects of the training:

general

--exec_mode=train_and_evaluate Which execution mode to run the model into.

--iter_unit=batch Will the model be run for X batches or X epochs ?

--num_iter=2500 Number of iterations to run.

--batch_size=16 Size of each minibatch per GPU.

--results_dir=/path/to/results Directory in which to write training logs, summaries and checkpoints.

--data_dir=/path/to/dataset Directory which contains the DAGM2007 dataset.

--dataset_name=DAGM2007 Name of the dataset used in this run (only DAGM2007 is supported atm).

--dataset_classID=1 ClassID to consider to train or evaluate the network (used for DAGM).

model

--amp Enable Automatic Mixed Precision to speedup FP32 computation using tensor cores.

--xla Enable TensorFlow XLA to maximise performance.

--use_auto_loss_scaling Use AutoLossScaling in TF-AMP

Dataset guidelines

The UNet model was trained with the Weakly Supervised Learning for Industrial Optical Inspection (DAGM 2007) dataset.

The competition is inspired by problems from industrial image processing. In order to satisfy their customers' needs, companies have to guarantee the quality of their products, which can often be achieved only by inspection of the finished product. Automatic visual defect detection has the potential to reduce the cost of quality assurance significantly.

The competitors have to design a stand-alone algorithm which is able to detect miscellaneous defects on various background textures.

The particular challenge of this contest is that the algorithm must learn, without human intervention, to discern defects automatically from a weakly labeled (i.e., labels are not exact to the pixel level) training set, the exact characteristics of which are unknown at development time. During the competition, the programs have to be trained on new data without any human guidance.

Source: https://resources.mpi-inf.mpg.de/conference/dagm/2007/prizes.html

The provided data is artificially generated, but similar to real world problems. It consists of multiple data sets, each consisting of 1000 images showing the background texture without defects, and of 150 images with one labeled defect each on the background texture. The images in a single data set are very similar, but each data set is generated by a different texture model and defect model.

Not all deviations from the texture are necessarily defects. The algorithm will need to use the weak labels provided during the training phase to learn the properties that characterize a defect.

Below are two sample images from two data sets. In both examples, the left images are without defects; the right ones contain a scratch-shaped defect which appears as a thin dark line, and a diffuse darker area, respectively. The defects are weakly labeled by a surrounding ellipse, shown in red.

The DAGM2007 challenge comes in propose two different challenges:

• A development set: public and available for download from here. The number of classes and sub-challenges for the development set is 6.
• A competition set: which requires an account to be downloaded from here. The number of classes and sub-challenges for the competition set is 10.

The challenge consists in designing a single model with a set of predefined hyper-parameters which will not change across the 10 different classes or sub-challenges of the competition set.

The performance shall be measured on the competition set which is normalized and more complex that the public dataset while offering the most unbiased evaluation method.

Training Process

Laplace Smoothing

We use this technique in the DICE loss to improve the training efficiency. This technique consists in replacing the epsilon parameter (used to avoid dividing by zero and very small: +/- 1e-7) by 1. You can find more information on: https://en.wikipedia.org/wiki/Additive_smoothing

Adaptive Loss

The DICE Loss is not able to provide a meaningful gradient at initialisation. This leads to a model instability which often push the model to diverge. Nonetheless, once the model starts to converge, DICE loss is able to very efficiently fully train the model. Therefore, we implemented an adaptive loss which is composed of two sub-losses:

• Binary Cross-Entropy (BCE)
• DICE Loss

The model is trained with the BCE loss until the DICE Loss reach a experimentally defined threshold (0.3). Thereafter, DICE loss is used to finish training.

Weak Labelling

This dataset is referred as weakly labelled. That means that the segmentation labels are not given at the pixel level but rather in an approximate fashion.