ClassIcam: Entwicklung von Klassifikationsmodellen für intelligente Netzwerk-Kameras

Surface inspection is of particular importance in industrial production, since the quality of a product can be recognized not only by its functionality but also by its visual representation. The project aims to inspect polymer surfaces with different colors and textures without having to explicitly train a system for a specific surface (color, surface texture).

An intelligent camera with integrated RISC processor and TCP/IP interface was used as optical system.

Solution concept 

The concept is based on methods of texture analysis of homogeneous, weakly granular surfaces. A new algorithm was developed that imitates the behavior of the human visual system. It is assumed that the human observer performs a consistency check in terms of surface homogeneity when viewing and correspondingly perceiving surfaces and makes a decision with respect to object quality based on this result.

For this purpose, a method was investigated and developed which decomposes a surface into homogeneous and non-homogeneous (edges, etc.) regions using directed sum and difference images.

Subsequently, local statistics that are probability density based are extracted from said images. From these, features that are considered representative of the homogeneous surface are derived (median decision of local features). These features are used to adjust a Modified Fuzzy Pattern Classifier (MFPC). It is used for subsequent decision making (classification) in terms of a quality statement regarding a surface.