# modified version of https://github.com/assafshocher/BlindSR_dataset_generator # Kai Zhang # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var # max_var = 2.5 * sf
(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0)
| 142 | |
| 143 | |
| 144 | def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): |
| 145 | """" |
| 146 | # modified version of https://github.com/assafshocher/BlindSR_dataset_generator |
| 147 | # Kai Zhang |
| 148 | # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var |
| 149 | # max_var = 2.5 * sf |
| 150 | """ |
| 151 | # Set random eigen-vals (lambdas) and angle (theta) for COV matrix |
| 152 | lambda_1 = min_var + np.random.rand() * (max_var - min_var) |
| 153 | lambda_2 = min_var + np.random.rand() * (max_var - min_var) |
| 154 | theta = np.random.rand() * np.pi # random theta |
| 155 | noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 |
| 156 | |
| 157 | # Set COV matrix using Lambdas and Theta |
| 158 | LAMBDA = np.diag([lambda_1, lambda_2]) |
| 159 | Q = np.array([[np.cos(theta), -np.sin(theta)], |
| 160 | [np.sin(theta), np.cos(theta)]]) |
| 161 | SIGMA = Q @ LAMBDA @ Q.T |
| 162 | INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] |
| 163 | |
| 164 | # Set expectation position (shifting kernel for aligned image) |
| 165 | MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) |
| 166 | MU = MU[None, None, :, None] |
| 167 | |
| 168 | # Create meshgrid for Gaussian |
| 169 | [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) |
| 170 | Z = np.stack([X, Y], 2)[:, :, :, None] |
| 171 | |
| 172 | # Calcualte Gaussian for every pixel of the kernel |
| 173 | ZZ = Z - MU |
| 174 | ZZ_t = ZZ.transpose(0, 1, 3, 2) |
| 175 | raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) |
| 176 | |
| 177 | # shift the kernel so it will be centered |
| 178 | # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) |
| 179 | |
| 180 | # Normalize the kernel and return |
| 181 | # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) |
| 182 | kernel = raw_kernel / np.sum(raw_kernel) |
| 183 | return kernel |
| 184 | |
| 185 | |
| 186 | def fspecial_gaussian(hsize, sigma): |