Spaces:
Running
Running
File size: 9,007 Bytes
a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 |
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torch
import numpy as np
import os, time, random
import argparse
from torch.utils.data import Dataset, DataLoader
from PIL import Image as PILImage
from glob import glob
from tqdm import tqdm
import rawpy
import colour_demosaicing
from .InvISP.model.model import InvISPNet
from .utils.common import Notify
from datasets.noise import (
camera_params,
addGStarNoise,
addPStarNoise,
addQuantNoise,
addRowNoise,
sampleK,
)
class NoiseSimulator:
def __init__(self, device, ckpt_path="./datasets/InvISP/pretrained/canon.pth"):
self.device = device
# load Invertible ISP Network
self.net = (
InvISPNet(channel_in=3, channel_out=3, block_num=8).to(self.device).eval()
)
self.net.load_state_dict(torch.load(ckpt_path), strict=False)
print(
Notify.INFO, "Loaded ISPNet checkpoint: {}".format(ckpt_path), Notify.ENDC
)
# white balance parameters
self.wb = np.array([2020.0, 1024.0, 1458.0, 1024.0])
# use Canon EOS 5D4 noise parameters provided by ELD
self.camera_params = camera_params
# random specify exposure time ratio from 50 to 150
self.ratio_min = 50
self.ratio_max = 150
pass
# inverse demosaic
# input: [H, W, 3]
# output: [H, W]
def invDemosaic(self, img):
img_R = img[::2, ::2, 0]
img_G1 = img[::2, 1::2, 1]
img_G2 = img[1::2, ::2, 1]
img_B = img[1::2, 1::2, 2]
raw_img = np.ones(img.shape[:2])
raw_img[::2, ::2] = img_R
raw_img[::2, 1::2] = img_G1
raw_img[1::2, ::2] = img_G2
raw_img[1::2, 1::2] = img_B
return raw_img
# demosaic - nearest ver
# input: [H, W]
# output: [H, W, 3]
def demosaicNearest(self, img):
raw = np.ones((img.shape[0], img.shape[1], 3))
raw[::2, ::2, 0] = img[::2, ::2]
raw[::2, 1::2, 0] = img[::2, ::2]
raw[1::2, ::2, 0] = img[::2, ::2]
raw[1::2, 1::2, 0] = img[::2, ::2]
raw[::2, ::2, 2] = img[1::2, 1::2]
raw[::2, 1::2, 2] = img[1::2, 1::2]
raw[1::2, ::2, 2] = img[1::2, 1::2]
raw[1::2, 1::2, 2] = img[1::2, 1::2]
raw[::2, ::2, 1] = img[::2, 1::2]
raw[::2, 1::2, 1] = img[::2, 1::2]
raw[1::2, ::2, 1] = img[1::2, ::2]
raw[1::2, 1::2, 1] = img[1::2, ::2]
return raw
# demosaic
# input: [H, W]
# output: [H, W, 3]
def demosaic(self, img):
return colour_demosaicing.demosaicing_CFA_Bayer_bilinear(img, "RGGB")
# load rgb image
def path2rgb(self, path):
return torch.from_numpy(np.array(PILImage.open(path)) / 255.0)
# InvISP
# input: rgb image [H, W, 3]
# output: raw image [H, W]
def rgb2raw(self, rgb, batched=False):
# 1. rgb -> invnet
if not batched:
rgb = rgb.unsqueeze(0)
rgb = rgb.permute(0, 3, 1, 2).float().to(self.device)
with torch.no_grad():
reconstruct_raw = self.net(rgb, rev=True)
pred_raw = reconstruct_raw.detach().permute(0, 2, 3, 1)
pred_raw = torch.clamp(pred_raw, 0, 1)
if not batched:
pred_raw = pred_raw[0, ...]
pred_raw = pred_raw.cpu().numpy()
# 2. -> inv gamma
norm_value = np.power(16383, 1 / 2.2)
pred_raw *= norm_value
pred_raw = np.power(pred_raw, 2.2)
# 3. -> inv white balance
wb = self.wb / self.wb.max()
pred_raw = pred_raw / wb[:-1]
# 4. -> add black level
pred_raw += self.camera_params["black_level"]
# 5. -> inv demosaic
if not batched:
pred_raw = self.invDemosaic(pred_raw)
else:
preds = []
for i in range(pred_raw.shape[0]):
preds.append(self.invDemosaic(pred_raw[i]))
pred_raw = np.stack(preds, axis=0)
return pred_raw
def raw2noisyRaw(self, raw, ratio_dec=1, batched=False):
if not batched:
ratio = (random.uniform(self.ratio_min, self.ratio_max) - 1) * ratio_dec + 1
raw = raw.copy() / ratio
K = sampleK(self.camera_params["Kmin"], self.camera_params["Kmax"])
q = 1 / (
self.camera_params["max_value"] - self.camera_params["black_level"]
)
raw = addPStarNoise(raw, K)
raw = addGStarNoise(
raw,
K,
self.camera_params["G_shape"],
self.camera_params["Profile-1"]["G_scale"],
)
raw = addRowNoise(raw, K, self.camera_params["Profile-1"]["R_scale"])
raw = addQuantNoise(raw, q)
raw *= ratio
return raw
else:
raw = raw.copy()
for i in range(raw.shape[0]):
ratio = random.uniform(self.ratio_min, self.ratio_max)
raw[i] /= ratio
K = sampleK(self.camera_params["Kmin"], self.camera_params["Kmax"])
q = 1 / (
self.camera_params["max_value"] - self.camera_params["black_level"]
)
raw[i] = addPStarNoise(raw[i], K)
raw[i] = addGStarNoise(
raw[i],
K,
self.camera_params["G_shape"],
self.camera_params["Profile-1"]["G_scale"],
)
raw[i] = addRowNoise(
raw[i], K, self.camera_params["Profile-1"]["R_scale"]
)
raw[i] = addQuantNoise(raw[i], q)
raw[i] *= ratio
return raw
def raw2rgb(self, raw, batched=False):
# 1. -> demosaic
if not batched:
raw = self.demosaic(raw)
else:
raws = []
for i in range(raw.shape[0]):
raws.append(self.demosaic(raw[i]))
raw = np.stack(raws, axis=0)
# 2. -> substract black level
raw -= self.camera_params["black_level"]
raw = np.clip(
raw, 0, self.camera_params["max_value"] - self.camera_params["black_level"]
)
# 3. -> white balance
wb = self.wb / self.wb.max()
raw = raw * wb[:-1]
# 4. -> gamma
norm_value = np.power(16383, 1 / 2.2)
raw = np.power(raw, 1 / 2.2)
raw /= norm_value
# 5. -> ispnet
if not batched:
input_raw_img = (
torch.Tensor(raw)
.permute(2, 0, 1)
.float()
.to(self.device)[np.newaxis, ...]
)
else:
input_raw_img = (
torch.Tensor(raw).permute(0, 3, 1, 2).float().to(self.device)
)
with torch.no_grad():
reconstruct_rgb = self.net(input_raw_img)
reconstruct_rgb = torch.clamp(reconstruct_rgb, 0, 1)
pred_rgb = reconstruct_rgb.detach().permute(0, 2, 3, 1)
if not batched:
pred_rgb = pred_rgb[0, ...]
pred_rgb = pred_rgb.cpu().numpy()
return pred_rgb
def raw2packedRaw(self, raw, batched=False):
# 1. -> substract black level
raw -= self.camera_params["black_level"]
raw = np.clip(
raw, 0, self.camera_params["max_value"] - self.camera_params["black_level"]
)
raw /= self.camera_params["max_value"]
# 2. pack
if not batched:
im = np.expand_dims(raw, axis=2)
img_shape = im.shape
H = img_shape[0]
W = img_shape[1]
out = np.concatenate(
(
im[0:H:2, 0:W:2, :],
im[0:H:2, 1:W:2, :],
im[1:H:2, 1:W:2, :],
im[1:H:2, 0:W:2, :],
),
axis=2,
)
else:
im = np.expand_dims(raw, axis=3)
img_shape = im.shape
H = img_shape[1]
W = img_shape[2]
out = np.concatenate(
(
im[:, 0:H:2, 0:W:2, :],
im[:, 0:H:2, 1:W:2, :],
im[:, 1:H:2, 1:W:2, :],
im[:, 1:H:2, 0:W:2, :],
),
axis=3,
)
return out
def raw2demosaicRaw(self, raw, batched=False):
# 1. -> demosaic
if not batched:
raw = self.demosaic(raw)
else:
raws = []
for i in range(raw.shape[0]):
raws.append(self.demosaic(raw[i]))
raw = np.stack(raws, axis=0)
# 2. -> substract black level
raw -= self.camera_params["black_level"]
raw = np.clip(
raw, 0, self.camera_params["max_value"] - self.camera_params["black_level"]
)
raw /= self.camera_params["max_value"]
return raw
|