Simple Itty-Koch-Style Saliency Maps

saliency.py

 # coding=utf-8
#
#!/usr/bin/python
#
# Copyright 2013 Johannes Bauer, Universitaet Hamburg
#
# This file is free software.  Do with it whatever you like.
# It comes with no warranty, explicit or implicit, whatsoever.
#
# This python script implements an early version of Itti and Koch's
# saliency model.  Specifically, it was written according to the
# information contained in the following paper:
#
#   Laurent Itti, Christof Koch, and Ernst Niebur. A model of
#   Saliency-Based visual attention for rapid scene analysis. IEEE
#   Transactions on Pattern Analysis and Machine Intelligence,
#   20(11):1254–1259, 1998.
#
# If you find it useful or if you have any questions, do not
# hesitate to contact me at 
#   bauer at informatik dot uni dash hamburg dot de.
#
# For information on how to use this script, type 
#   > python saliency.py -h
# on the command line.
#

import math
import logging
import cv2
import numpy
from scipy.ndimage.filters import maximum_filter

import os.path
import sys

if sys.version_info[0] != 2:
	raise Exception("This script was written for Python version 2.  You're running Python %s." % sys.version)

logger = logging.getLogger(__name__)


def features(image, channel, levels=9, start_size=(640,480), ):
	"""
		Extracts features by down-scaling the image levels times,
		transforms the image by applying the function channel to 
		each scaled version and computing the difference between 
		the scaled, transformed	versions.
			image : 		the image
			channel : 		a function which transforms the image into
							another image of the same size
			levels : 		number of scaling levels
			start_size : 	tuple.  The size of the biggest image in 
							the scaling	pyramid.  The image is first
							scaled to that size and then scaled by half
							levels times.  Therefore, both entries in
							start_size must be divisible by 2^levels.
	"""
	image = channel(image)
	if image.shape != start_size:
		image = cv2.resize(image, dsize=start_size)

	scales = [image]
	for l in xrange(levels - 1):
		logger.debug("scaling at level %d", l)
		scales.append(cv2.pyrDown(scales[-1]))

	features = []
	for i in xrange(1, levels - 5):
		big = scales[i]
		for j in (3,4):
			logger.debug("computing features for levels %d and %d", i, i + j)
			small = scales[i + j]
			srcsize = small.shape[1],small.shape[0]
			dstsize = big.shape[1],big.shape[0]
			logger.debug("Shape source: %s, Shape target :%s", srcsize, dstsize)
			scaled = cv2.resize(src=small, dsize=dstsize)
			features.append(((i+1,j+1),cv2.absdiff(big, scaled)))

	return features

def intensity(image):
	"""
		Converts a color image into grayscale.
		Used as `channel' argument to function `features'
	"""
	return cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)

def makeGaborFilter(dims, lambd, theta, psi, sigma, gamma):
	"""
		Creates a Gabor filter (an array) with parameters labmbd, theta,
		psi, sigma, and gamma of size dims.  Returns a function which 
		can be passed to `features' as `channel' argument.

		In some versions of OpenCV, sizes greater than (11,11) will lead
		to segfaults (see http://code.opencv.org/issues/2644).
	"""
	def xpf(i,j):
		return i*math.cos(theta) + j*math.sin(theta)
	def ypf(i,j):
		return -i*math.sin(theta) + j*math.cos(theta)
	def gabor(i,j):
		xp = xpf(i,j)
		yp = ypf(i,j)
		return math.exp(-(xp**2 + gamma**2*yp**2)/2*sigma**2) * math.cos(2*math.pi*xp/lambd + psi)
	
	halfwidth = dims[0]/2
	halfheight = dims[1]/2

	kernel = numpy.array([[gabor(halfwidth - i,halfheight - j) for j in range(dims[1])] for i in range(dims[1])])

	def theFilter(image):
		return cv2.filter2D(src = image, ddepth = -1, kernel = kernel, )

	return theFilter

def intensityConspicuity(image):
	"""
		Creates the conspicuity map for the channel `intensity'.
	"""
	fs = features(image = im, channel = intensity)
	return sumNormalizedFeatures(fs)

def gaborConspicuity(image, steps):
	"""
		Creates the conspicuity map for the channel `orientations'.
	"""
	gaborConspicuity = numpy.zeros((60,80), numpy.uint8)
	for step in range(steps):
		theta = step * (math.pi/steps)
		gaborFilter = makeGaborFilter(dims=(10,10), lambd=2.5, theta=theta, psi=math.pi/2, sigma=2.5, gamma=.5)
		gaborFeatures = features(image = intensity(im), channel = gaborFilter)
		summedFeatures = sumNormalizedFeatures(gaborFeatures)
		gaborConspicuity += N(summedFeatures)
	return gaborConspicuity

def rgConspicuity(image):
	"""
		Creates the conspicuity map for the sub channel `red-green conspicuity'.
		of the color channel.
	"""
	def rg(image):
		r,g,_,__ = cv2.split(image)
		return cv2.absdiff(r,g)
	fs = features(image = image, channel = rg)
	return sumNormalizedFeatures(fs)

def byConspicuity(image):
	"""
		Creates the conspicuity map for the sub channel `blue-yellow conspicuity'.
		of the color channel.
	"""
	def by(image):
		_,__,b,y = cv2.split(image)
		return cv2.absdiff(b,y)
	fs = features(image = image, channel = by)
	return sumNormalizedFeatures(fs)

def sumNormalizedFeatures(features, levels=9, startSize=(640,480)):
	"""
		Normalizes the feature maps in argument features and combines them into one.
		Arguments:
			features	: list of feature maps (images)
			levels		: the levels of the Gaussian pyramid used to
							calculate the feature maps.
			startSize	: the base size of the Gaussian pyramit used to
							calculate the feature maps.
		returns:
			a combined feature map.
	"""
	commonWidth = startSize[0] / 2**(levels/2 - 1)
	commonHeight = startSize[1] / 2**(levels/2 - 1)
	commonSize = commonWidth, commonHeight
	logger.info("Size of conspicuity map: %s", commonSize)
	consp = N(cv2.resize(features[0][1], commonSize))
	for f in features[1:]:
		resized = N(cv2.resize(f[1], commonSize))
		consp = cv2.add(consp, resized)
	return consp

def N(image):
	"""
		Normalization parameter as per Itti et al. (1998).
		returns a normalized feature map image.
	"""
	M = 8.	# an arbitrary global maximum to which the image is scaled.
			# (When saving saliency maps as images, pixel values may become
			# too large or too small for the chosen image format depending
			# on this constant)
	image = cv2.convertScaleAbs(image, alpha=M/image.max(), beta=0.)
	w,h = image.shape
	maxima = maximum_filter(image, size=(w/10,h/1))
	maxima = (image == maxima)
	mnum = maxima.sum()
	logger.debug("Found %d local maxima.", mnum)
	maxima = numpy.multiply(maxima, image)
	mbar = float(maxima.sum()) / mnum
	logger.debug("Average of local maxima: %f.  Global maximum: %f", mbar, M)
	return image * (M-mbar)**2

def makeNormalizedColorChannels(image, thresholdRatio=10.):
	"""
		Creates a version of the (3-channel color) input image in which each of
		the (4) channels is normalized.  Implements color opponencies as per 
		Itti et al. (1998).
		Arguments:
			image			: input image (3 color channels)
			thresholdRatio	: the threshold below which to set all color values
								to zero.
		Returns:
			an output image with four normalized color channels for red, green,
			blue and yellow.
	"""
	intens = intensity(image)
	threshold = intens.max() / thresholdRatio
	logger.debug("Threshold: %d", threshold)
	r,g,b = cv2.split(image)
	cv2.threshold(src=r, dst=r, thresh=threshold, maxval=0.0, type=cv2.THRESH_TOZERO)
	cv2.threshold(src=g, dst=g, thresh=threshold, maxval=0.0, type=cv2.THRESH_TOZERO)
	cv2.threshold(src=b, dst=b, thresh=threshold, maxval=0.0, type=cv2.THRESH_TOZERO)
	R = r - (g + b) / 2
	G = g - (r + b) / 2
	B = b - (g + r) / 2
	Y = (r + g) / 2 - cv2.absdiff(r,g) / 2 - b

	# Negative values are set to zero.
	cv2.threshold(src=R, dst=R, thresh=0., maxval=0.0, type=cv2.THRESH_TOZERO)
	cv2.threshold(src=G, dst=G, thresh=0., maxval=0.0, type=cv2.THRESH_TOZERO)
	cv2.threshold(src=B, dst=B, thresh=0., maxval=0.0, type=cv2.THRESH_TOZERO)
	cv2.threshold(src=Y, dst=Y, thresh=0., maxval=0.0, type=cv2.THRESH_TOZERO)

	image = cv2.merge((R,G,B,Y))
	return image

def markMaxima(saliency):
	"""
		Mark the maxima in a saliency map (a gray-scale image).
	"""
	maxima = maximum_filter(saliency, size=(20,20))
	maxima = numpy.array(saliency == maxima, dtype=numpy.float64) * 255
	r = cv2.max(saliency, maxima)
	g = saliency
	b = saliency
	marked = cv2.merge((b,g,r))
	return marked


if __name__ == "__main__":
	logging.basicConfig(level=logging.DEBUG)

	import argparse
	import sys
	parser = argparse.ArgumentParser(description = "Simple Itti-Koch-style conspicuity.")
	parser.add_argument('--fileList', type=str, dest='fileList', action='store', help='Text file containing input file names, one per line.')
	parser.add_argument('--inputFile', type=str, dest='inputFile', action='store', help='File to compute compute saliency list for.  Need either --fileList or --inputFile.')
	parser.add_argument('--intensityOutput', type=str, dest='intensityOutput', action='store', help="Filename for intensity conspicuity map,")
	parser.add_argument('--gaborOutput', type=str, dest='gaborOutput', action='store', help="Filename for intensity conspicuity map,")
	parser.add_argument('--rgOutput', type=str, dest='rgOutput', action='store', help="Filename for rg conspicuity map,")
	parser.add_argument('--byOutput', type=str, dest='byOutput', action='store', help="Filename for by conspicuity map,")
	parser.add_argument('--cOutput', type=str, dest='cOutput', action='store', help="Filename for color conspicuity map,")
	parser.add_argument('--saliencyOutput', type=str, dest='saliencyOutput', action='store', help="Filename for saliency map,")
	parser.add_argument("--markMaxima", action='store_true', help="Mark maximum saliency in output image.")
	args = parser.parse_args()

	if args.fileList is None and args.inputFile is None:
		logger.error("Need either --fileList or --inputFile cmd line arguments.")
		sys.exit()
	elif args.fileList is not None and args.inputFile is not None:
		logger.error("Need only one of --fileList or --inputFile cmd line arguments.")
		sys.exit()
	else:
		
		if args.fileList:
			# we are reading filenames from a file.
			filenames = (filename[:-1] for filename in open(args.fileList)) # remove end-of line character
		else:
			# filenames were given on the command line.
			filenames = [args.inputFile]

		for filename in filenames:
			im = cv2.imread(filename, cv2.COLOR_BGR2RGB) # assume BGR, convert to RGB---more intuitive code.
			
			if im is None:
				logger.fatal("Could not load file \"%s.\"", filename)
				sys.exit()


			intensty = intensityConspicuity(im)
			gabor = gaborConspicuity(im, 4)

			im = makeNormalizedColorChannels(im)
			rg = rgConspicuity(im)
			by = byConspicuity(im)
			c = rg + by
			saliency = 1./3 * (N(intensty) + N(c) + N(gabor))

			if args.markMaxima:
				saliency = markMaxima(saliency)

			def writeCond(outFileName, image):
				name,_ = os.path.splitext(os.path.basename(filename))
				if outFileName and args.fileList:
					cv2.imwrite(outFileName % name, image)
				elif outFileName:
					cv2.imwrite(outFileName, image)

			writeCond(args.intensityOutput, intensty)
			writeCond(args.gaborOutput, gabor)
			writeCond(args.rgOutput, rg)
			writeCond(args.byOutput, by)
			writeCond(args.cOutput, .25 * c)
			writeCond(args.saliencyOutput, saliency)

Hi, Johannes. thank you for the quick response. I am studying on "Audio-Visual Saliency Prediction in 360-Degree Videos", and Itti-style normalization/quantization has worked well for producing the final saliency maps.

Hi @tatome,
I have another question regarding your code.
In the function sumNormalizedFeatures we compute the commonSize with

commonWidth = startSize[0] / 2**(levels/2 - 1)
commonHeight = startSize[1] / 2**(levels/2 - 1)

I do not quite understand why we devide by 2**(levels/2 - 1). Which would be 2**3.5. Shoudn't it be

commonWidth = startSize[0] / 2**((levels-1)/2)
commonHeight = startSize[1] / 2**((levels-1)/2)

Which would leave us with 2**4=16. This makes more sense to me since the reduction factor at scale 4 is 1:16. Or am I misunderstanding something?

Thanks in advance :)

Hi @maltesilber , I am not the owner of the code. @tatome will answer more correctly.
But you seem to have made a good point.

Hi @maltesilber,

thanks for your interested in this script and potentially highlighting a bug.

It's been a long time since I wrote it. At this point, your understanding is probably as good as mine.
Have you tried your version of the code? Does it work? Are the results different?

The "size" of the saliency patches change. One patch corresponds to a 16x16 patch now, which is also stated by the original paper.