ICBV Course Final Project Sergey Tyrin Itamar Barkai

Preview:

Citation preview

ICBV Course Final ProjectSergey TyrinItamar Barkai

The main goal of this research is to develop an image analysis system capable of locating

grapes in a 2-dimensional image with no additional

information

1. Shape inference using circular Hough transform

2. Hue and Green colour separation3. Local minima search and Clustering

Circular Hough transform parameters:• The range of radii to search for .• The perfectness threshold of a circle.

Playing with perfectness threshold :

Low threshold High threshold

Combining thresholds :

Threshold combination

Playing with radii threshold :

Too high range of radii Many false detections

All of the vine grape berries have a yellow-greenish hue, which is easily separated by the human vision system from the reddish hue of the trunks and from the almost white colour of the grass in the background.

The leaves have commonly a very similar range of green hue, and a human observer wouldn’t possibly know to separate the grapes from the leaves if he hadn’t had an a-priori knowledge of the shape.

Green hue layer of the RGB

Hue layer of the HSV

Look for more information in other colour spaces separation...

Solution for this was found in the yellow hue layer of the CMYK representation. The sun-highlighted centres of the berries stand out.

I. Local minima extractionFinding areas of local minima, cutting

of singlepoint minima.

II. Removal of noise detections by spot sizeFiltering out where spot size > 300.95% probability of being part of the

background.

III. ClusteringLooking for condensed groups of spots.

III. ClusteringLooking for condensed groups of

spots.

IV. Removal of edge linesMaking segments more oval by

removing one-pixel-width lines.

V. Filtering of a cluster as one entity•Using RGB, Magenta, and Black colour spaces to calculate median values of clusters as total and comparing these values to each segment median values .•Removing segments where median max value and min value are too far away.•Removing segments where magenta is too high or black is too low (trunk or background).•Applying this filter to the cluster as whole.

VI. Calculation of a hullJoining segments into real clusters by

findingconvex hull – inserts non-grape parts into

clusters.

VI. Calculation of a hullSolution – calculating Delaunay

triangulation and cutting off too long edges.

VII. Generating final maskRemoval of non-grape triangles, hull

filling and final rounding of the segmentation

mask.

• Due to the highly saturated and unrestricted lighting conditions, it is impossible to detect each and every grape in the image.

• For normalized lighting conditions, a satisfying result can be achieved.

• For the different lighting conditions, using two different sets of threshold parameters, increases the successful detection rate significantly.

Is grapes detection solved?

• Bright/dark areas – apply noise filtering thresholds to smaller regions.

• Smarter cluster filtering – comprehensive filtering decisions.

• Learning process – clusters and grapes shape, lighting conditions.

Recommended