def getSaliencyFrame(frame):
widthFrame = len(frame[0])
heightFrame = len(frame)
cv2.imshow('frame', frame)
#compress the frame
frame = cv2.resize(frame, (64, 64), interpolation=cv2.INTER_CUBIC)
# fourier analyse
imageFourier = np.fft.fft2(frame)
# show imageFourier
showPlot(imageFourier)
cv2.imshow('fourier', np.abs(imageFourier))
# we now have the spectrum analyse
# by using convolve we can blur the frame
imageConvolve = convolutions.convolve(frame, np.abs(kernel(9)))
# take the fourier analyse of the convolve
fourierConvolve = np.fft.fft2(imageConvolve)
showPlot(fourierConvolve)
cv2.imshow('averagedSpectrum', np.abs(fourierConvolve))
#substract two images
spectralResidual = np.subtract(fourierConvolve, imageFourier)
showPlot(spectralResidual)
# inverse fourier analyse
saliencyImage = np.fft.ifft2(spectralResidual)
showPlot(saliencyImage)
saliencyImage = cv2.resize(np.abs(saliencyImage),
(widthFrame, heightFrame))
cv2.imshow('saliency', saliencyImage)
# while True:
# if cv2.waitKey(1) & 0xFF == ord('q'):
# break
return saliencyImage
def main():
image = scipy.misc.imread('image.jpg')
width, height, num_colors = image.shape
im = np.zeros([num_colors, width, height])
for i in xrange(num_colors):
im[i, :, :] = image[:, :, i]
w1 = np.array([
[0, -1, 0],
[-1, 4, -1],
[0, -1, 0],
])
w2 = np.array([
[-1, 0, 1],
[-2, 0, 2],
[-1, 0, 1],
])
w3 = np.array([
[-1, -2, -1],
[0, 0, 0],
[1, 2, 1],
])
w = np.zeros([3, 3, 3])
w[0, :, :] = w1
w[1, :, :] = w2
w[2, :, :] = w3
kernel_size = w.shape[1]
padding = kernel_size / 2
w_flip = w[:, ::-1, ::-1]
result = convolutions.convolve(im[np.newaxis, :, :, :],
w_flip[:, np.newaxis, :, :],
padding=padding)
plt.subplot(2, 2, 1)
plt.imshow(image)
for k in xrange(3):
res = result[k, :, :, :]
res -= res.min()
res /= res.max()
res *= 255
r = np.zeros([width, height, num_colors], dtype=np.uint8)
for i in xrange(num_colors):
r[:, :, i] = res[i, :, :]
plt.subplot(2, 2, 2 + k)
plt.imshow(r)
plt.show()
scaleFrameX = widthFrame / 64
scaleFrameY = int((64 / widthFrame) * heightFrame)
cv2.imshow('frame', frame)
frame = cv2.resize(frame, (64, 64), interpolation=cv2.INTER_CUBIC)
# frame = convolutions.convolve(frame, np.abs(kernel(5)))
showPlot(np.zeros((5, 5)))
imageFourier = np.fft.fft2(frame)
logImageFourier = np.log(imageFourier)
showPlot(imageFourier)
cv2.imshow('fourier', np.abs(imageFourier))
# we now have the spectrum analyse
# wich needs to be logaritmic:
imageConvolve = convolutions.convolve(frame, np.abs(kernel(9)))
cv2.imshow('org', np.abs(frame))
cv2.imshow('blur', np.abs(imageConvolve))
fourierConvolve = np.fft.fft2(imageConvolve)
logFourierConvolve = np.log(fourierConvolve)
showPlot(fourierConvolve)
cv2.imshow('averagedSpectrum', np.abs(fourierConvolve))
# averagedSpectrum = np.multiply(imageConvolve, logSpectrum)
# showPlot(averagedSpectrum)
phase = np.angle(imageFourier)
spectralResidual = np.subtract(logFourierConvolve, logImageFourier)
spectralResidual = np.exp(spectralResidual + 1j * phase)**2
showPlot(spectralResidual)
saliencyImage = np.fft.ifft2(spectralResidual)