Parameter-learning for color sorting of bulk materials using genetic algorithms
Sensor based sorting finds broad applications in mining, recycling and quality control. Digital image processing and pattern recognition are key components, as they determine whether to keep or discard an object under inspection. In many scenarios, the color of a material stands out as the primary sorting criterion. In this paper, we present a flexible system for color sorting of bulk materials based on semantic color features. The features are constructed in a three stages: the color occurrence frequencies of different materials are estimated and then fused to a small number of color classes, which in turn are used to map each color to a discrete attribute. A compact object descriptor composed of the fractions of foreground pixels that share the same attribute characterizes the objects under inspection. This descriptor has many advantages: it has a very clear, intuitive interpretation, is invariant to rotation and scale of the object and requires very little computation. However, a major drawback are the many variables that govern the construction process. Manual fine tuning requires a large amount of time and experience. Subtle changes in the parameters can have strong effects on the classification performance. To overcome this shortcoming, we propose a method to automatically learn the parameters by a genetic algorithm. We apply our method to wine grape sorting problems to show that this approach performs at least as good a human expert. At the same time, it takes considerably less effort on the human part and frees the operator to attend to other tasks.