Drop swarm analysis in dispersions with incident-light and transmitted-light illumination

摘要

The new developed Optical Multimode Online Probe (OMOP) can process images from either incident-light illumination (also called epi-illumination) or transmitted-light illumination (also called trans-illumination). The probe has an outer diameter of 38 mm and the illumination is achieved by high performance Light Emitting Diodes (LEDs) with specifications of 1.96 mm~2 and 493 lm (251.53 lm/mm~2) at an angular deviation of 0.37°. A camera probe is used with either an object-space telecentric (telecentricity < 0.2°, 2437 mm virtual pupil distance) or entocentric objective (Kohler based illumination, 6238 mm virtual pupil distance). Using the telecentric mode, the particle distance independency is located within 20 mm while the focal depth is approximately 5 mm. The local resolution is between 20 and 30 μm, according to the used optics, with a standard deviation less than 4.5%. Maximum particle diameter is up to 5 mm while particles can reach up to 2 m/s as function of exposure. The basic distance transform approach with watershed segmentation for analysis of transmitted-light images gives deviations less than 5% at high particle densities and less than 2% at low ones. The error of false positives typically is below 5% while the error of wrong radiuses is below 1% for up to 90% of all droplets and below 5% otherwise. Up to five images per core and second (trans-illumination) can be analyzed automatically and online at densities up to 25% (trans-illumination, gap width less than 5 mm) 40% (object side telecentric epi-illumination, single probe) respectively. The advanced pre-segmentation approach based on the Random Forest Classifier (RFC) is used to perform the more complex image analysis with epi-illumination. As long as the quality of pre-segmentation is high enough, the classification results in images, which can be analyzed in the following distance transform approach. This is considerably depending on the quality of training the algorithm and recurring image features. Compared to the distance transform analysis at low densities the deviation increases. The RFC pre-segmented image gives an additional deviation of 1.1% (both in regard to the total amount of evaluated pixels) and a deviation of 12.9% in regard to the mean particle diameter. Below a particle size of 50 pixels the image analysis overestimates the actual number of particles due to the sensitivity of the Euclidian distance approach.