Figure 5
From: Wide Field-of-View Fluorescence Imaging of Coral Reefs
Computer-aided segmentation of the live areas of coral colonies using fluorescence images.
In (a) and (b), yellow lines indicate user-designated foreground strokes, blue lines indicate user-designated background strokes and red-green lines indicate automatically generated contours. (a) Supervised segmentation done with the same strokes on (left to right) reflectance, nighttime fluorescence and daytime (mixed with ambient light) fluorescence images, compared to manual segmentation that was done by outlining the colony on a reflectance image. The automated segmentation of the reflectance image mistakenly includes some part of the sand as the coral, along with other anomalies, while the nighttime fluorescence image enables accurate automatic segmentation. The daytime fluorescence image (without ambient light subtraction) also demonstrated improved contrast between the coral and its background due to the red chlorophyll-a fluorescence component. The automatic contour is almost optimal and it only missed a small part of the colony in the lower right corner. These images were taken in Bocas Del Toro, Panama, at 5 m depth. (b) Segmentation of a wide angle scene during daytime, using reflectance and the red channel of Iday (daytime fluorescence image mixed with ambient). As in (a), segmentation using the reflectance image included part of the background as a coral. These images were taken in Moorea, French Polynesia, at 10 m depth, at mid-day. Red arrows point to a coral that blends in with the background rock in the reflectance image, but stands out in the fluorescence image. (c) Segmentation of four additional corals (actual images not shown for clarity). For each coral, semi-automated segmentation was done using a fluorescence image (blue curves) and compared to a manual segmentation done using a reflectance image (black curve). Images were not registered so the contours were manually aligned. The contours closely match, but the semi-automated segmentation using fluorescence images required only 20% of the time of manual segmentation.
