Automated recognition of oil drops in images of multiphase dispersions via gradient direction pattern

A novel feature designed to enable the automated recognition of oil drops from other circular objects found in images obtained in-situ of a multiphase dispersion model (oil in water) is described. The conditions inside a stirring tank are highly dynamic; this is reflected in the complexity of the images obtained, which in turn makes the problem of automating the recognition of objects of interest a very difficult one. To the best of our knowledge this is the first reported attempt to achieve fully-automated recognition of oil drops in this type of images. The proposed feature synthesizes local gradient orientation patterns that are characteristic of the boundary of the oil drops. The feature was tested as part of a supervised recognition framework based on a Bayesian classifier and employing the Hough transform for circles as a pre-selector of objects with circular shapes. By using the proposed feature alone, the classifier obtained a sensitivity value of 85% and a false-positive reduction of 33% (in this context a false-positive is an image-artifact with an approximately circular shape that was detected by the Hough transform but that does not correspond to an oil drop).