1 from __future__ import division 2 import os, time, scipy.io 3 import tensorflow as tf 4 import numpy as np 5 import rawpy 6 import glob 7 from model_sid_latest import network_my_unet 8 import platform 9 from PIL import Image 10 11 # BBF100-2 12 bbf_w = 4032 // 2 13 bbf_h = 3024 // 2 14 15 input_dir = 'D:/data/Sony/dataset/illu_detect/' 16 17 18 def preprocess(raw, bl, wl): 19 im = raw.raw_image_visible.astype(np.float32) 20 im = np.maximum(im - bl, 0) 21 return im / (wl - bl) 22 23 24 def pack_raw_bbf(path): 25 raw = rawpy.imread(path) 26 bl = 64 27 wl = 1023 28 im = preprocess(raw, bl, wl) 29 im = np.expand_dims(im, axis=2) 30 H = im.shape[0] 31 W = im.shape[1] 32 if raw.raw_pattern[0, 0] == 0: # CFA=RGGB 33 out = np.concatenate((im[0:H:2, 0:W:2, :], 34 im[0:H:2, 1:W:2, :], 35 im[1:H:2, 1:W:2, :], 36 im[1:H:2, 0:W:2, :]), axis=2) 37 elif raw.raw_pattern[0,0] == 2: # BGGR 38 out = np.concatenate((im[1:H:2, 1:W:2, :], 39 im[0:H:2, 1:W:2, :], 40 im[0:H:2, 0:W:2, :], 41 im[1:H:2, 0:W:2, :]), axis=2) 42 elif raw.raw_pattern[0,0] == 1 and raw.raw_pattern[0,1] == 0: # GRBG 43 out = np.concatenate((im[0:H:2, 1:W:2, :], 44 im[0:H:2, 0:W:2, :], 45 im[1:H:2, 0:W:2, :], 46 im[1:H:2, 1:W:2, :]), axis=2) 47 elif raw.raw_pattern[0,0] == 1 and raw.raw_pattern[0,1] == 2: # GBRG 48 out = np.concatenate((im[1:H:2, 0:W:2, :], 49 im[0:H:2, 0:W:2, :], 50 im[0:H:2, 1:W:2, :], 51 im[1:H:2, 1:W:2, :]), axis=2) 52 else: 53 assert False 54 # wb = np.array(raw.camera_whitebalance) 55 # wb[3] = wb[1] 56 # wb = wb / wb[1] 57 # out = np.minimum(out * wb, 1.0) 58 return out 59 60 61 in_im = tf.placeholder(tf.float32, [1, bbf_h, bbf_w, 4], name='input') 62 63 # Obtain the noise level map 64 radius = 3 65 strides = 32 66 67 # in_patches_r = tf.extract_image_patches( 68 # tf.expand_dims( 69 # in_im[:, :, :, 0], axis=3), 70 # (1, radius, radius, 1), 71 # (1, strides, strides, 1), 72 # (1, 1, 1, 1), 73 # 'VALID') 74 # miu_r = tf.maximum(tf.reduce_mean(in_patches_r, axis=-1, keepdims=True), 1e-7) 75 # sigma_r = tf.reduce_sum(tf.abs((in_patches_r - miu_r) / miu_r), axis=-1, keepdims=True)/(radius * radius - 1) 76 77 in_patches_g = tf.extract_image_patches( 78 tf.expand_dims( 79 in_im[:, :, :, 3], axis=3), 80 (1, radius, radius, 1), 81 (1, strides, strides, 1), 82 (1, 1, 1, 1), 83 'VALID') 84 miu_g = tf.maximum(tf.reduce_mean(in_patches_g, axis=-1, keepdims=True), 1e-7) 85 sigma_g = tf.reduce_sum(tf.abs((in_patches_g - miu_g) / miu_g), axis=-1, keepdims=True)/(radius * radius - 1) 86 87 # in_patches_b = tf.extract_image_patches( 88 # tf.expand_dims( 89 # in_im[:, :, :, 2], axis=3), 90 # (1, radius, radius, 1), 91 # (1, strides, strides, 1), 92 # (1, 1, 1, 1), 93 # 'VALID') 94 # miu_b = tf.maximum(tf.reduce_mean(in_patches_b, axis=-1, keepdims=True), 1e-7) 95 # sigma_b = tf.reduce_sum(tf.abs((in_patches_b - miu_b) / miu_b), axis=-1, keepdims=True)/(radius * radius - 1) 96 # 97 # sigma_level = tf.reduce_mean(tf.minimum(tf.minimum(sigma_r, sigma_g), sigma_b)) 98 99 sigma_level = tf.reduce_mean(sigma_g) 100 illu_level = tf.reduce_mean(miu_g) 101 print(in_patches_g.shape) 102 103 sess = tf.Session() 104 # print('============= DAY ============') 105 # files = glob.glob(input_dir + '/day/*.dng') 106 # files.sort() 107 # for i in range(len(files)): 108 # noise_level = sess.run(sigma_level, feed_dict={in_im: np.expand_dims(pack_raw_bbf(files[i]), axis=0)}) 109 # illu = 0.3 / noise_level 110 # print(illu) 111 # 112 # print('============= NIGHT ============') 113 # files = glob.glob(input_dir + '/night/*.dng') 114 # files.sort() 115 # for i in range(len(files)): 116 # noise_level = sess.run(sigma_level, feed_dict={in_im: np.expand_dims(pack_raw_bbf(files[i]), axis=0)}) 117 # illu = 0.3 / noise_level 118 # print(illu) 119 120 print('============= NIGHT ============') 121 files = glob.glob('C:/Users/Administrator/Desktop/ILLU/*.dng') 122 files.sort() 123 for i in range(len(files)): 124 sigma_, illu_ = sess.run([sigma_level, illu_level], feed_dict={in_im: np.expand_dims(pack_raw_bbf(files[i]), axis=0)}) 125 illu = 0.3 / sigma_ 126 print(illu) 127 print(illu_) 128 print('--------------')
通过检测每个采样像素的邻域均方差来确定画面的整体噪声水平。
这里需要注意的是每个点的明暗程度不一致，因此需要对均方差进行亮度归一化。
【上述程序有个BUG】
在存在明显的光源时，会出现照度测量不正确的情形。为了解决这一问题，引入全局画面亮度作为参考，即：平均照度应该考量总体像素值波动幅度的同时对全局亮度的归一。
1 from __future__ import division 2 import os, time, scipy.io 3 import tensorflow as tf 4 import numpy as np 5 import rawpy 6 import glob 7 from model_sid_latest import network_my_unet 8 import platform 9 from PIL import Image 10 11 # BBF100-2 12 bbf_w = 4032 // 2 13 bbf_h = 3024 // 2 14 15 input_dir = 'D:/data/Sony/dataset/illu_detect/' 16 17 18 def preprocess(raw, bl, wl): 19 im = raw.raw_image_visible.astype(np.float32) 20 im = np.maximum(im - bl, 0) 21 return im / (wl - bl) 22 23 24 def pack_raw_bbf(path): 25 raw = rawpy.imread(path) 26 bl = 64 27 wl = 1023 28 im = preprocess(raw, bl, wl) 29 im = np.expand_dims(im, axis=2) 30 H = im.shape[0] 31 W = im.shape[1] 32 if raw.raw_pattern[0, 0] == 0: # CFA=RGGB 33 out = np.concatenate((im[0:H:2, 0:W:2, :], 34 im[0:H:2, 1:W:2, :], 35 im[1:H:2, 1:W:2, :], 36 im[1:H:2, 0:W:2, :]), axis=2) 37 elif raw.raw_pattern[0,0] == 2: # BGGR 38 out = np.concatenate((im[1:H:2, 1:W:2, :], 39 im[0:H:2, 1:W:2, :], 40 im[0:H:2, 0:W:2, :], 41 im[1:H:2, 0:W:2, :]), axis=2) 42 elif raw.raw_pattern[0,0] == 1 and raw.raw_pattern[0,1] == 0: # GRBG 43 out = np.concatenate((im[0:H:2, 1:W:2, :], 44 im[0:H:2, 0:W:2, :], 45 im[1:H:2, 0:W:2, :], 46 im[1:H:2, 1:W:2, :]), axis=2) 47 elif raw.raw_pattern[0,0] == 1 and raw.raw_pattern[0,1] == 2: # GBRG 48 out = np.concatenate((im[1:H:2, 0:W:2, :], 49 im[0:H:2, 0:W:2, :], 50 im[0:H:2, 1:W:2, :], 51 im[1:H:2, 1:W:2, :]), axis=2) 52 else: 53 assert False 54 # wb = np.array(raw.camera_whitebalance) 55 # wb[3] = wb[1] 56 # wb = wb / wb[1] 57 # out = np.minimum(out * wb, 1.0) 58 return out 59 60 61 in_im = tf.placeholder(tf.float32, [1, bbf_h, bbf_w, 4], name='input') 62 63 # Obtain the noise level map 64 radius = 3 65 strides = 32 66 67 in_patches_g = tf.extract_image_patches( 68 tf.expand_dims( 69 in_im[:, :, :, 3], axis=3), 70 (1, radius, radius, 1), 71 (1, strides, strides, 1), 72 (1, 1, 1, 1), 73 'VALID') 74 miu_g = tf.maximum(tf.reduce_mean(in_patches_g, axis=-1, keepdims=True), 1e-7) 75 illu_level = tf.reduce_mean(miu_g) 76 # sigma_g = tf.reduce_sum(tf.abs((in_patches_g - miu_g) / miu_g), axis=-1, keepdims=True)/(radius * radius - 1) 77 sigma_g = tf.reduce_sum(tf.abs((in_patches_g - miu_g) / illu_level), axis=-1, keepdims=True)/(radius * radius - 1) 78 79 sigma_level = tf.reduce_mean(sigma_g) 80 print(in_patches_g.shape) 81 82 sess = tf.Session() 83 84 files = glob.glob('C:/Users/Administrator/Desktop/ILLU/*.dng') 85 86 files.sort() 87 for i in range(len(files)): 88 sigma_, illu_ = sess.run([sigma_level, illu_level], feed_dict={in_im: np.expand_dims(pack_raw_bbf(files[i]), axis=0)}) 89 illu = 0.25 / sigma_ 90 print(illu)