Hi,
I implemented a code to gabor filter cifar10 data but the images after being filtered and stacked are not clear like the original images. I think the problem is in the way I am using the values for gabor filter's parameters. These values should work because I got them form this paper 'Adding biological constraints to CNNs makes image classification more human-like and robust'
Here are the values:
kernel size = 31*31
thatas = 0, pi/8, pi/4, pi*3/8, pi/2, pi*5/8, pi*3/4, pi*7/8
offset (psi)= 0, pi/2, pi, pi*3/2
gamma = 0.5, 1
lambda = 5
bandwidth = 1,1.4,1.8
For sigmas I used equation (3) from paper 'Adding biological constraints to CNNs makes image classification more human-like and robust'
sigma_x = (lambda/pi)*(sqrt(np.log(2)/2))*((2**b + 1)/(2**b - 1))
sigma_y = sigma_x/gamma
I do not think the problem is in the code itself because I am using the ready gabor filter function provided from tensorflow:
tfio.experimental.filter.gabor( input, freq, sigma=None, theta=0, nstds=3, offset=0, mode=None, constant_values=None, name=None)
I used it such that
freq = 1/lambda
theta = theta
nstds = kernel size
offset = psi
sigma = [sigma_x, sigma_y]
- Badges
- Science topic
- Similar topics
- Cognitive Science
- Thinking