Edge Detection with Machine Vision Using Derivative of Gaussian Filters with q-Gaussian Kernels
Jos F. E. Bruggeman, Kurt Stockman, Jeroen D. M. De Kooning · 2025
Edge detection is a fundamental operation in a multitude of computer vision applications, with significant relevance in mechatronic systems for tasks such as object identification, image segmentation, shape recognition and feature extraction from images. This process is contingent upon the observation that edges typically arise at boundaries between regions exhibiting notable intensity fluctuations in luma images. This study examines the Derivative of Gaussian (DoG) filter, which identifies edges in luma images captured with a low-cost vision system inside a mechatronic application, by convolving the filter with the luma image. This process serves to highlight regions where rapid intensity changes occur, thus enabling precise edge localisation. A novel advancement is presented in this article which improves these filters using Tsallis statistics, particularly the q-Gaussian probability density function. This approach incorporates an additional parameter, the entropic index$q$, into the filters, thereby providing enhanced flexibility in shaping them. Subsequently, the efficacy of conventional DoG filters is evaluated in comparison to the novel q-Gaussian-based counterparts (DoqG) and their employement on a low-cost embedded system within a mechatronic system. The results demonstrate that the latter exhibits superior edge detection accuracy, enabling precise tracking of material following a trajectory while passing through a mechatronic system.