Image segmentation

From P2P Food Lab
Jump to: navigation, search

To have good yield, it is specially important to avoid invasion by weeds in the first weeks during which th crops are grown. Algorithms are necessary to identify thecrops and the grounds. An appropriate colour index is often sufficient [1]. More advanced processing using neural networks trained on large databases have also shown some interesting in recent times.

Segmentation based on color thresholds

  • RGB image

To identify plants, a commonly employed quantity is based on the excess green index (EGI) which in normalized form is:

<math>EGI=3*G/L-1.</math>

EGI under vairious ligh conditions
  • HSV image

Green may may also be detected from the hue channel after conversion to the HSV representation. Selecting the hue values close to green with suffisant saturation gives the following index:

<math>HGI=S/(1+exp(-||H-H_0||^2/d))</math>.

  • CieL*a*b* image

In CieL*a*b* space, a is a direct measure of the green/red balance.


Additional filtering

Bilateral filters and anisotopic diffusion help in getting rid of glitches.

Superpixels

Superpixels enable to perform computations on coarse regoions so that it runs faster.

Morphological snakes

Details may vanish from filtering or coarse graining, morphological snakes recover some of those details.

Semantic segmentation

Labeling of all pixels can be implemented with neural networks for example. To be implemented...

References

  1. A survey of image processing techniques for plant extraction and segmentation in the field Esmael Hamuda , Martin Glavin, Edward Jones, Computers and Electronics in Agriculture 2016