Image Filtering

Learn how to filter images through a pipeline.

# Make sure we have the image plugin
import numpy as np
import imagepypelines as ip
ip.require('image')

Construct our filtering blocks

# Frequency Filter
@ip.blockify()
def freq_filter(src, kernel):
    return src * kernel

# Easy "frequency space" normalized lowpass filter
@ip.blockify()
def circular_pass_filter(shape, radius=0.1, type='low'):
    x, y = np.meshgrid(np.linspace(-1, 1, shape[0]), np.linspace(-1, 1, shape[1]))
    circle = np.sqrt(x*x + y*y)

    if type=='low':
        return np.where(circle > radius, 0, 1)

    elif type=='high':
        return np.where(circle > radius, 1, 0)

    else:
        raise ValueError("Filter 'type' must be either 'low' or 'high'")

Define our tasks and build the pipeline

tasks = {
        # set an entry point for images into the pipeline
        'images': ip.Input(),
        # get the image dimensions for later use
        'shape': (ip.image.Dimensions(order="WHC"), 'images'),
        # take the fourier transform
        'fft': (ip.image.ImageFFT(order="WHC"), 'images'),
        # produce 'frequency space' circular pass filters
        'circles': (circular_pass_filter, 'shape'),
        # add an extra dimension to the filters for broadcasting the arrays
        'unsqueezed': (ip.image.Unsqueeze(-1), 'circles'),
        # the convolution of your image and filter is element-wise
        # multiplication in frequency space!!!
        'filtered': (freq_filter, 'fft', 'unsqueezed'),
        # take the inverse fourier transform of your filtered images
        'ifft': (ip.image.ImageIFFT(order="WHC"), 'filtered'),
        # Scale the images, cast dtype, and view the filtered images
        # in sequence! Note: simple filtering in RGB results in artifacts!
        'safe': (ip.image.DisplaySafe(), 'ifft'),
        'null' : (ip.image.CompareView(pause_for=500), 'images','safe')
         }

im_filt = ip.Pipeline(tasks)

Let’s process some data!

Let’s grab some example data from the ImagePypelines standard set

images = [ip.image.panda(), ip.image.gecko(), ip.image.redhat()]
# Number and view the images!
processed = im_filt.process(images)