Machine vision based automatic separation of touching convex shaped objects

An automatic separation procedure of touching objects based on their curvature values was developed. Plastic bottles were used as model geometries. Using elliptic Fourier descriptors (EFDs), the image contours were smoothed avoiding the occurrence of local pseudo-corners resulted from the image acquisition inefficiencies. Then, nodal points were determined by evaluating the curvature along the smoothed boundary of the images. Nodal points are those points at which the curvature fell below a threshold. In situations where multiple nodal points were found, the ‘nearest-neighbor’ and ‘critical radial distance’ criteria were used to draw the segmentation lines. The algorithm was tested to separate bottles of different sizes and shapes with different touching scenarios and showed a success rate of more than 99%. At the current stage of development, the algorithm can effectively separate objects with smooth boundaries and is thus suitable for separating mechanical objects with well-defined geometries. Further work is being done to make it robust enough to handle biological entities as well as to deal with scenarios where one or more objects are surrounded by a multitude of objects.

► Automatic machine vision based separation procedure was developed.
► Fourier approximation ensured all derivatives existed and were continuous.
► The computational costs of determining nodal points reduced significantly.
► Our algorithm separated multiple touching scenarios successfully.