Optical quality control of food is quite a challenge for pattern recognition systems because of the high variability of the items to be recognized. One example for the necessity of flexible algorithms is the automatic placement of food into aluminium fast food packages which requires fast changes of the system because of fast changes of the products. Usually the system needs a training phase. A learned menu can be classified in three steps: In the first step a Split-and-Merge-Methods finds connected regions with similar color. These colors are then compared with the taught-in prototypes of the different food items. In the last step the region sizes and positions are scored. The algorithm is highly optimized in speed and only needs one second to process a single image of about 1MByte size.
To speed up the training phase due to a high rate of menue changes a supervised training method is desirable that allows an easy handling and merely relies on "good" or "bad" answers of the teaching person. Our method involves the following steps:
Clustering in location domain: Like in the classification step the image is anlyzed by a Split-and-Merge method that describes the most important parts of the input images by very few clusters.
Clustering in color domain: A k-Means clustering is used to determine the most important colors of the images.
Generation of the feature space: Based on the number and size of pairs of neighboring regions with certain colors we build feature vectors for each image that are saved with the classification of the user.
Checking convergence: After the manual classification of a miminum number of images the system can do a pre-classification by a simple Nearest-Neighbor method. If the classification matches the manual classification a reasonable number of times the training phase was successful.
The described method can recognize both color and positioning errors.
It can be implemented on a standard PC (Windows) and fulfills real-time
demands. The simple user interface ensures that people can easily
learn to train and use the system even when frequent menu changes
Application report (in German)")?>