Download notebook (.ipynb)

geom_imshow() and geom_raster()#

geom_imshow() displays an image specified by 2D or 3D Numpy array.

Whether the image is grayscale or color depends on the shape of the image array:

  • (M, N) - grey-scale image

  • (M, N, 3) - color RGB image

  • (M, N, 4) - color RGB image with alpha channel

For example, you can create an RGB image array from a PNG file using PIL and io packages.

geom_raster() on the other hand is a regular ggplot graphic layer wich takes a data-frame as the input and applies the aesthetic mappings as specified by the aes() function.

import numpy as np
import pandas as pd

from lets_plot import *

LetsPlot.setup_html()

from PIL import Image
import requests
from io import BytesIO

response = requests.get('https://github.com/JetBrains/lets-plot-docs/raw/master/source/examples/cookbook/images/fisher_boat.png')

image = Image.open(BytesIO(response.content))
img = np.asarray(image)
img.shape

(225, 225, 3)

RGB image#

M x N x 3 ndarray specifies a color RGB image.

ggplot() + geom_imshow(img)

Grayscale image#

M x N ndarray specifies a grayscale image.

We will generate a grayscale image from RGB image above by selecting only its R channel.

img_gs = img[:,:,0]
img_gs.shape

(225, 225)

ggplot() + geom_imshow(img_gs)

Grayscale image in pseudo colors (cmap)#

ggplot() + geom_imshow(img_gs, cmap="magma")

geom_raster()#

The geom_raster() function also can display image. But unlike geom_imshow(), geom_raster() doesn’t accept ndarray as an input.

As a regular ggplot graphic layer, geom_raster() requires an observations-style input data and aestetic mapping.

This is not very convenient when the data is originally in the form of ndarray. On the other hand, aesthetic mapping and color scales might come in handy when rendering an image.

Lets transform the image data to a data-frame suitable for geom_raster().

# reshape 3-dimentional ndarray to 2-dimentional long-form ndarray
# and then to data frame with columns x,y,r,g,b
cols, rows, _ = img.shape
img_lf = img.reshape(cols * rows, -1)
img_df = pd.DataFrame(img_lf,columns=['r','g','b'])
X_mesh, Y_mesh = np.meshgrid(np.arange(rows), np.arange(cols))
img_df['x'] = X_mesh.reshape(-1)
img_df['y'] = Y_mesh.reshape(-1)[::-1]

# Pack color components values to 24-bit RGB values 
c_fill = img_df.apply(lambda row: ((int(row['r'] * 255) << 16) + 
                                   (int(row['g'] * 255) << 8) +
                                   int(row['b'] * 255)),
                      axis=1)

# Show image in pseudo colors with only few gradations 
(ggplot(img_df) + geom_raster(aes('x', 'y', fill=c_fill), show_legend=False)
        + scale_fill_brewer()
        + ggtitle('Raster geom with brewer palette') + ggsize(800,500))