Cannabis Bud Area

Baby Steps to Tokenized Cannabis

https://tonsoffun.github.io/greenthumb/2018/05/22/greenthumb-bud-area/

Cannabis Computer Vision

I gleamed my inspiration from PyImageSearch.com yet again. I’m basically counting bright green spots, by combing some of the tricks from the last post on Leaf Area Analysis.

Detecting multiple bright spots in an image with Python and OpenCV - PyImageSearch

Dependencies

In this docker image I have all of the dependencies for both Leaf Area and Bud Area including working copies and works in progress. We’re going to be using the same setup as with the Leaf Area notebook. OpenCV will be used for preprocessing the image and pixel selection by range. Matplotlib is used from rendering images and graphs. We take advantage of skimage.measure.label for taking measurements and label them. We take these components and use OpenCV to find and annotate the contours.

HSV Range Detector

I pulled the first few lines of the script below from this comment. It was need to run the range_detector used in this and the previous blog post.

Running GUI applications using Docker for Mac

You can checkout this gif to get an idea of how I go about defining the HSV boundary values.

Creating the initial HSV Mask

Simple range mask like we did with the leaf detector

Preview the mask

We can easily mask the original image with OpenCV’s bitwise_and which takes the original image and the mask and outputs a masked image.

Cannabis Computer Vision

This is a side-by-side view of the HSV colorspace which the computer uses to isolate the buds and the masked preview. We use the numpy.vstack to vertically stack the images. I chose vstack for the gists here and hstack in the notebook, but it really is a matter of developer preference.

Refining the Mask

We can refine the mask using the same binary threshold, erode, and dilate technique used on the Leaf Area. This will clean up the specs of frost leave from the actual Bud Area.

Before and After Cleaning Up the Contours

You can see the vstack comparing the contours pre and post cleanup.

You can still see a few smaller contours that escape this sweep, but they’re technically smaller bud sites below the top layer of canopy. Either way we’ll filter them in the next pass below.

Enhance Mask with Labeled Components

Using skimage.measure.label we are further filtering out noise by ignoring blobs smaller than 100px. This means buds have to be this size or greater to get added to the labeled mask.

Label the Buds

Now we have what we need to label the bud sites on the preview image.

You can see the smaller spots, which actually appear to be smaller bud sites below the canopy are not labeled. We can remove or lower the threshold of 100px to further refine the camera’s calibration, but this is a great result for this stage.

A Full Copy of the Notebook Below

Below is the full copy of the notebook. I also have a docker image tonsoffun/tensorflow-notebook which I’ve simply added OpenCV 3.3 via anaconda. This docker image supports both Bud and Leaf Area Analysis as well. Keep following along as I show you how we can use our OpenCV masks to train a Mask RCNN.

top: HSV Original, bottom: Final Mask Before Contour Filtering