Weakly Supervised Learning for Industrial Optical Inspection

DAGM 2007 and the German Chapter of the European Neural Network Society (GNNS) offer a competition in which every interested participant of the Symposium can take part.

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.

Data description

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 ellipse-parameters are provided in a separate .txt-file with a format as shown below.

[filename] \t \n
[semi-major axis] \t [semi-minor axis] \t [rotation angle] \t
[x-position of the centre of the ellipsoid] \t [y-position of the centre of the ellipsoid] \n
[filename] \t ...

The rotation angle is measured counterclockwise (positive angle). x- and y-coordinates are given in MATLAB format, i.e. the origin is in the upper left corner of the image.

For the development of the algorithm, the participants will be provided with 6 different data sets, each simulated using a different texture and defect model. During the competition, the performance of the participants' algorithms will be tested on 4 different data sets, previously unknown to the users. Note that these data sets are generated by texture models and defect models different from the models of the first 6 data sets.

The results of the test phase have to be stored in a text file with the following format:

[filename] \t [date of test as] \t [time of test as] \t
[1 if a defect was found, 0 if no defect was found] \n
[filename] \t ...


The 4 competition data sets are separated into training data and test data, each consisting of 1000 images without defects and 150 images with defects. The defects of the training data are weakly labeled, the defects of the test data are not, of course. Note again that no manual manipulation is allowed during the training or testing phase.

Participants bring their own computing equipment. At the time of the competition, the participants are first given a CD-ROM with the training data. They then have 5 minutes to start the training phase on their computers, under supervision from the panel of judges. The computers then have to be left running unattended in a separate locked room, with no internet access. 24 hours later, access to the computers is granted again to start the test phase. The corresponding data is again given on a CD-ROM and the participants have 5 minutes to start their evaluation, again under supervision. 12 hours later, access to the computers is granted again and a file in the format detailed above containing the test results has to be given to the judges.

For the comparison of the algorithms the following loss matrix will be employed.

Prediction\Truth No defect Defect
No defect 0 20
Defect 1 0


Only those competitors whose total loss count:=(number of false negatives)*20 + (number of false positives)*1 does not exceed 200 qualify for a prize. The total prize money is 3000 EUR. It will be divided among the three best qualifying participants at the ratio of 70/20/10%. In case there are only two participants that qualify for a prize, the ratio is 80/20%. If only one participant qualifies, he/she receives the entire prize money. If no competitor qualifies, no prize money will be distributed.

Scientists involved in the organization of the competition do not qualify for participation.

Exploitation rights

The source code does not have to be revealed. The participants keep all rights to their software. The winner is expected to give a short presentation of his/her approach after the awards ceremony at the end of the DAGM.


At least one member of each participating group must be registered for and has to be present at the DAGM 2007.


The data sets for the development of the algorithms can be downloaded below as of today (May 15th).

Deadline for registration of participation at the email address provided below: August 1st.
The participants have to register with their computers with the organizers of the competition at the DAGM 2007 in Heidelberg, Germany, on September 11th, noon.
The competition will take place from the evening of September 11th to September 13th.


For further questions please contact: Dr. Christian Perwass, Please use 'DAGM 2007 Competition' as subject for all emails; other emails will be considered as spam.

Design award for new DAGM logo

The DAGM solicits proposals for a new logo. The logo should be usable both in small scale (for letterheads, etc.) and in large scale (for banners, etc.); color submissions must be supplemented by a black-and-white version.

The proposals will be rated by a jury and the three best designs will be awarded Eur 500, Eur 350 and Eur 200, respectively. Submissions from non-DAGM members are explicitly welcome. All proposals should be emailed to before August 1st, 2007.