The detection and quantification of surface defects is important for monitoring the production of plastic parts. This is particularly the case for automotive interior parts at prominent locations and in a high-throughput production, where an automated inspection gets economically relevant. For a human observer with its cognitive pattern recognition it is easily possible to distinguish defects from superimposing scattering substructures even if they are changing significantly due to regular process variations. However, for an automated in-line inspection system defect detection under these conditions proves to be a major challenge and requires not only reliable detection but also evaluation in very short time. In particular, the defect inspection of transparent surfaces with scattering substructures like they occur i.e. at illuminant components is a major computational problem. In our new defect inspection setup surface images of the part are acquired by a CCD-camera under specifically conditioned illumination. The images are processed using a hessian based eigenvalue analysis creating an inherently strongly smoothed filter image where especially line structures are amplified. The bulk of regular structures is removed by superimposed filtered golden samples. This is only enabled by the smoothness of the filter images. The few remaining structures well delineated from each other, which are then morphologically analyzed. Based on a multi-criteria decision making process defect and regular structures can be distinguished. The visual perceptibility of defect structures can then be estimated by its contrast and its spatial expansion and shape. The procedure has been tested on a set of 56 parts with defects like flow lines, smears, etc. with a very positive result.