基于卷积神经网络的图像去噪（基础篇）

import numpy as np import os import matplotlib.pyplot as plt from keras_preprocessing import image import cv2 import re def resize_images(src_path, sizeX, sizeY):#对图像进行resize images = [] imagesresize = [] width, height = sizeX, sizeY for filename in os.listdir(src_path): img = cv2.imread(src_path + filename) if img is not None: images.append(img) cnt = 1 for img in images: im2 = cv2.resize(img, (width, height), interpolation=cv2.INTER_CUBIC) ext = ".png" imagesresize.append(im2) cnt += 1 print(len(imagesresize)) return imagesresize def gaussian_noise(x_train, mean, sigma):#添加高斯噪声 x_train_noisy = [] for i in range(x_train.shape[0]): image = x_train[i] noise = np.random.normal(mean, sigma/255.0, x_train[0].shape) x_train_noisy.append(image + noise) x_train_noisy = np.clip(x_train_noisy, 0., 1.) return np.array(x_train_noisy) x = resize_images('data/Test/Set12/', 256,256) image_data = [] for file in x:#读取文件夹下的图片并保存到列表中 gray_img = cv2.cvtColor(file, cv2.COLOR_BGR2GRAY) img_array = image.img_to_array(gray_img) image_data.append(img_array) img_data = np.array(image_data, dtype='float32') / 255.0 print(img_data.shape) noisy_set = gaussian_noise(img_data,0,25) ''' for i in range(noisy_set.shape[0]): cv2.imwrite("./noise/" + str(i) + ".jpg", noisy_set[i] * 255.) ''' print(noisy_set.shape) from keras.layers import * from keras.layers import Lambda from keras.models import Model from keras import backend as K def get_cnn_9conv(model_path=None): input_img = Input(shape=((None, None, 1))) # adapt this if using `channels_first` image data format x = Conv2D(56, (5, 5), activation='relu', padding='same')(input_img) x = Conv2D(12, (1, 1), activation='relu', padding='same')(x) x = Conv2D(12, (3, 3), activation='relu', padding='same')(x) x = Conv2D(12, (3, 3), activation='relu', padding='same')(x) x = Conv2D(12, (3, 3), activation='relu', padding='same')(x) x = Conv2D(12, (3, 3), activation='relu', padding='same')(x) x = Conv2D(56, (1, 1), activation='relu', padding='same')(x) x = Conv2D(28, (9, 9), activation='relu', padding='same')(x) decoded = Conv2D(1, (5, 5), activation='relu', padding='same')(x) autoencoder = Model(input_img, decoded) autoencoder.compile(optimizer='sgd', loss='mean_squared_error', metrics=['accuracy', 'mean_squared_error']) # print(autoencoder.summary())# return autoencoder def get_gated_connections(gatePercentageFactor, inputLayer): gateFactor = Input(tensor=K.variable([gatePercentageFactor])) fractionG = Lambda(lambda x: x[0] * x[1])([inputLayer, gateFactor]) complement = Lambda(lambda x: x[0] - x[1])([inputLayer, fractionG]) return gateFactor, fractionG, complement def get_cnn_architecture(weights_path=None): input_img = Input(shape=(None, None, 1)) # adapt this if using `channels_first` image data format x1 = Conv2D(64, (3, 3), activation='relu', padding='same')(input_img)#180 180 64 gf1, fg1, c1 = get_gated_connections(0.1, x1) x = MaxPooling2D((2, 2), padding='same')(fg1)#90 90 64 x2 = Conv2D(64, (3, 3), activation='relu', padding='same')(x) #90 90 64 gf2, fg2, c2 = get_gated_connections(0.2, x2) x = MaxPooling2D((2, 2), padding='same')(fg2) #45 45 64 x3 = Conv2D(128, (3, 3), activation='relu', padding='same')(x)#45 45 128 gf3, fg3, c3 = get_gated_connections(0.3, x3) x = MaxPooling2D((2, 2), padding='same')(x3)#22.5 22.5 128 x4 = Conv2D(256, (3, 3), activation='relu', padding='same')(x)#22.5 22.5 256 gf4, fg4, c4 = get_gated_connections(0.4, x4) x = MaxPooling2D((2, 2), padding='same')(x4) x5 = Conv2D(512, (3, 3), activation='relu', padding='same')(x) gf5, fg5, c5 = get_gated_connections(0.5, x5) x = MaxPooling2D((2, 2), padding='same')(x5) x6 = Conv2D(1024, (3, 3), activation='relu', padding='same')(x) x = UpSampling2D((2, 2))(x6) y1 = Conv2DTranspose(512, (3, 3), activation='relu', padding='same')(x) jt4 = Add()([y1, c5]) x = UpSampling2D((2, 2))(jt4) y2 = Conv2DTranspose(256, (3, 3), activation='relu', padding='same')(x) jt3 = Add()([y2, c4]) x = UpSampling2D((2, 2))(jt3) y3 = Conv2DTranspose(128, (3, 3), activation='relu', padding='same')(x) jt2 = Add()([y3, c3]) x = UpSampling2D((2, 2))(jt2)#90 90 128 y4 = Conv2DTranspose(64, (3, 3), activation='relu', padding='same')(x)#90 90 64 jt1 = Add()([y4, c2]) x = UpSampling2D((2, 2))(jt1)#180 180 64 jt0 = Add()([x, c1]) y5 = Conv2DTranspose(64, (3, 3), activation='relu', padding='same')(jt0)#180 180 64 y6 = Conv2DTranspose(1, (3, 3), activation='relu', padding='same')(y5)#180 180 3 layers = y6 #model = Model([input_img,gf1,gf2], layers) model = Model([input_img, gf1, gf2, gf3, gf4, gf5], layers) model.compile(optimizer='sgd', loss='mean_squared_error', metrics=['accuracy'])#编译 #model.summary() return model def DnCNN(depth, filters=64, image_channels=1, use_bnorm=True): layer_count = 0 inpt = Input(shape=(None, None, image_channels), name='input' + str(layer_count)) # 1st layer, Conv+relu layer_count += 1 x = Conv2D(filters=filters, kernel_size=(3, 3), strides=(1, 1), kernel_initializer='Orthogonal', padding='same', name='conv' + str(layer_count))(inpt) layer_count += 1 x = Activation('relu', name='relu' + str(layer_count))(x) # depth-2 layers, Conv+BN+relu for i in range(depth - 2): layer_count += 1 x = Conv2D(filters=filters, kernel_size=(3, 3), strides=(1, 1), kernel_initializer='Orthogonal', padding='same', use_bias=False, name='conv' + str(layer_count))(x) if use_bnorm: layer_count += 1 # x = BatchNormalization(axis=3, momentum=0.1,epsilon=0.0001, name = 'bn'+str(layer_count))(x) x = BatchNormalization(axis=3, momentum=0.0, epsilon=0.0001, name='bn' + str(layer_count))(x) layer_count += 1 x = Activation('relu', name='relu' + str(layer_count))(x) # last layer, Conv layer_count += 1 x = Conv2D(filters=image_channels, kernel_size=(3, 3), strides=(1, 1), kernel_initializer='Orthogonal', padding='same', use_bias=False, name='conv' + str(layer_count))(x) layer_count += 1 x = Subtract(name='subtract' + str(layer_count))([inpt, x]) # input - noise model = Model(inputs=inpt, outputs=x) return model if __name__ == '__main__': #bmodel = get_cnn_9conv() #bmodel = get_cnn_architecture() bmodel = DnCNN(depth=17,filters=64,image_channels=1,use_bnorm=True) #bmodel.load_weights('./save_weights/get_cnn_9conv25.h5') #bmodel.load_weights('./save_weights/get_cnn_architecture25.h5') bmodel.load_weights('./save_weights/DnCNN25.h5') pic = bmodel.predict(noisy_set) from skimage.measure import compare_psnr psnrs = [] psnr1 = [] print(img_data[1].shape,pic[1].shape) for x in range(len(noisy_set)): ''' psnr_x = skimage.metrics.peak_signal_noise_ratio(img_data[x] * 255., pic[x] * 255.,data_range = 255) psnr_y = skimage.metrics.peak_signal_noise_ratio(img_data[x] * 255., noisy_set[x] * 255.,data_range = 255) ''' psnr_x = compare_psnr(img_data[x] * 255., pic[x] * 255., data_range=255) psnr_y = compare_psnr(img_data[x] * 255., noisy_set[x] * 255., data_range=255) psnrs.append(psnr_x) psnr1.append(psnr_y) psnr_avg = np.mean(psnrs) psnr_avg1 = np.mean(psnr1) for y in psnrs: print(y, end=' ') print(psnr_avg) for z in psnr1: prin