風格遷移： 在內容上盡量與基准圖像保持一致，在風格上盡量與風格圖像保持一致。
- 1. 使用預訓練的VGG19網絡提取特征
- 2. 損失函數之一是“內容損失”（content loss），代表合成的圖像的特征與基准圖像的特征之間的L2距離，保證生成的圖像內容和基准圖像保持一致。
- 3. 損失函數之二是“風格損失”（style loss），代表合成圖像的特征與風格圖像的特征之間的Gram矩陣之間的差異，保證生成圖像的風格和風格圖像保持一致。
- 4. 損失函數之三是“差異損失”（variation loss），代表合成的圖像局部特征之間的差異，保證生成的圖像局部特征的一致性，整體看上去自然不突兀。
基於keras的代碼實現：
# coding: utf-8
from __future__ import print_function
from keras.preprocessing.image import load_img, img_to_array
import numpy as np
from scipy.optimize import fmin_l_bfgs_b
import time
import argparse
from scipy.misc import imsave
from keras.applications import vgg19
from keras import backend as K
import os
from PIL import Image, ImageFont, ImageDraw, ImageOps, ImageEnhance, ImageFilter
# 輸入參數
parser = argparse.ArgumentParser(description='基於Keras的圖像風格遷移.') # 解析器
parser.add_argument('--style_reference_image_path', metavar='ref', type=str,default = './style.jpg',
help='目標風格圖片的位置')
parser.add_argument('--base_image_path', metavar='ref', type=str,default = './base.jpg',
help='基准圖片的位置')
parser.add_argument('--iter', type=int, default=25, required=False,
help='迭代次數')
parser.add_argument('--pictrue_size', type=int, default=500, required=False,
help='圖片大小.')
# 獲取參數
args = parser.parse_args()
base_image_path = args.base_image_path
style_reference_image_path = args.style_reference_image_path
iterations = args.iter
pictrue_size = args.pictrue_size
source_image = Image.open(base_image_path)
source_image= source_image.resize((pictrue_size, pictrue_size))
width, height = pictrue_size, pictrue_size
def save_img(fname, image, image_enhance=True): # 圖像增強
image = Image.fromarray(image)
if image_enhance:
# 亮度增強
enh_bri = ImageEnhance.Brightness(image)
brightness = 1.2
image = enh_bri.enhance(brightness)
# 色度增強
enh_col = ImageEnhance.Color(image)
color = 1.2
image = enh_col.enhance(color)
# 銳度增強
enh_sha = ImageEnhance.Sharpness(image)
sharpness = 1.2
image = enh_sha.enhance(sharpness)
imsave(fname, image)
return
# util function to resize and format pictures into appropriate tensors
def preprocess_image(image):
"""
預處理圖片，包括變形到(1，width, height)形狀，數據歸一到0-1之間
:param image: 輸入一張圖片
:return: 預處理好的圖片
"""
image = image.resize((width, height))
image = img_to_array(image)
image = np.expand_dims(image, axis=0) # (width, height)->(1，width, height)
image = vgg19.preprocess_input(image) # 0-255 -> 0-1.0
return image
def deprocess_image(x):
"""
將0-1之間的數據變成圖片的形式返回
:param x: 數據在0-1之間的矩陣
:return: 圖片，數據都在0-255之間
"""
x = x.reshape((width, height, 3))
x[:, :, 0] += 103.939
x[:, :, 1] += 116.779
x[:, :, 2] += 123.68
# 'BGR'->'RGB'
x = x[:, :, ::-1]
x = np.clip(x, 0, 255).astype('uint8') # 以防溢出255范圍
return x
def gram_matrix(x): # Gram矩陣
assert K.ndim(x) == 3
if K.image_data_format() == 'channels_first':
features = K.batch_flatten(x)
else:
features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))
gram = K.dot(features, K.transpose(features))
return gram
# 風格損失，是風格圖片與結果圖片的Gram矩陣之差，並對所有元素求和
def style_loss(style, combination):
assert K.ndim(style) == 3
assert K.ndim(combination) == 3
S = gram_matrix(style)
C = gram_matrix(combination)
S_C = S-C
channels = 3
size = height * width
return K.sum(K.square(S_C)) / (4. * (channels ** 2) * (size ** 2))
#return K.sum(K.pow(S_C,4)) / (4. * (channels ** 2) * (size ** 2)) # 居然和平方沒有什么不同
#return K.sum(K.pow(S_C,4)+K.pow(S_C,2)) / (4. * (channels ** 2) * (size ** 2)) # 也能用，花后面出現了葉子
def eval_loss_and_grads(x): # 輸入x，輸出對應於x的梯度和loss
if K.image_data_format() == 'channels_first':
x = x.reshape((1, 3, height, width))
else:
x = x.reshape((1, height, width, 3))
outs = f_outputs([x]) # 輸入x，得到輸出
loss_value = outs[0]
if len(outs[1:]) == 1:
grad_values = outs[1].flatten().astype('float64')
else:
grad_values = np.array(outs[1:]).flatten().astype('float64')
return loss_value, grad_values
# an auxiliary loss function
# designed to maintain the "content" of the
# base image in the generated image
def content_loss(base, combination):
return K.sum(K.square(combination - base))
# the 3rd loss function, total variation loss,
# designed to keep the generated image locally coherent
def total_variation_loss(x,img_nrows=width, img_ncols=height):
assert K.ndim(x) == 4
if K.image_data_format() == 'channels_first':
a = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] - x[:, :, 1:, :img_ncols - 1])
b = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] - x[:, :, :img_nrows - 1, 1:])
else:
a = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] - x[:, 1:, :img_ncols - 1, :])
b = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] - x[:, :img_nrows - 1, 1:, :])
return K.sum(K.pow(a + b, 1.25))
# Evaluator可以只需要進行一次計算就能得到所有的梯度和loss
class Evaluator(object):
def __init__(self):
self.loss_value = None
self.grads_values = None
def loss(self, x):
assert self.loss_value is None
loss_value, grad_values = eval_loss_and_grads(x)
self.loss_value = loss_value
self.grad_values = grad_values
return self.loss_value
def grads(self, x):
assert self.loss_value is not None
grad_values = np.copy(self.grad_values)
self.loss_value = None
self.grad_values = None
return grad_values
# 得到需要處理的數據，處理為keras的變量（tensor），處理為一個(3, width, height, 3)的矩陣
# 分別是基准圖片，風格圖片，結果圖片
base_image = K.variable(preprocess_image(source_image)) # 基准圖像
style_reference_image = K.variable(preprocess_image(load_img(style_reference_image_path)))
if K.image_data_format() == 'channels_first':
combination_image = K.placeholder((1, 3, width, height))
else:
combination_image = K.placeholder((1, width, height, 3))
# 組合以上3張圖片，作為一個keras輸入向量
input_tensor = K.concatenate([base_image, style_reference_image, combination_image], axis=0) #組合
# 使用Keras提供的訓練好的Vgg19網絡,不帶3個全連接層
model = vgg19.VGG19(input_tensor=input_tensor,weights='imagenet', include_top=False)
model.summary() # 打印出模型概況
'''
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, None, None, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, None, None, 64) 1792 A
_________________________________________________________________
block1_conv2 (Conv2D) (None, None, None, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, None, None, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, None, None, 128) 73856 B
_________________________________________________________________
block2_conv2 (Conv2D) (None, None, None, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, None, None, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, None, None, 256) 295168 C
_________________________________________________________________
block3_conv2 (Conv2D) (None, None, None, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, None, None, 256) 590080
_________________________________________________________________
block3_conv4 (Conv2D) (None, None, None, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, None, None, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, None, None, 512) 1180160 D
_________________________________________________________________
block4_conv2 (Conv2D) (None, None, None, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, None, None, 512) 2359808
_________________________________________________________________
block4_conv4 (Conv2D) (None, None, None, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, None, None, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, None, None, 512) 2359808 E
_________________________________________________________________
block5_conv2 (Conv2D) (None, None, None, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, None, None, 512) 2359808
_________________________________________________________________
block5_conv4 (Conv2D) (None, None, None, 512) 2359808 F
_________________________________________________________________
block5_pool (MaxPooling2D) (None, None, None, 512) 0
=================================================================
'''
# Vgg19網絡中的不同的名字，儲存起來以備使用
outputs_dict = dict([(layer.name, layer.output) for layer in model.layers])
loss = K.variable(0.)
layer_features = outputs_dict['block5_conv2']
base_image_features = layer_features[0, :, :, :]
combination_features = layer_features[2, :, :, :]
content_weight = 0.08
loss += content_weight * content_loss(base_image_features,
combination_features)
feature_layers = ['block1_conv1','block2_conv1','block3_conv1','block4_conv1','block5_conv1']
feature_layers_w = [0.1,0.1,0.4,0.3,0.1]
# feature_layers = ['block5_conv1']
# feature_layers_w = [1]
for i in range(len(feature_layers)):
# 每一層的權重以及數據
layer_name, w = feature_layers[i], feature_layers_w[i]
layer_features = outputs_dict[layer_name] # 該層的特征
style_reference_features = layer_features[1, :, :, :] # 參考圖像在VGG網絡中第i層的特征
combination_features = layer_features[2, :, :, :] # 結果圖像在VGG網絡中第i層的特征
loss += w * style_loss(style_reference_features, combination_features) # 目標風格圖像的特征和結果圖像特征之間的差異作為loss
loss += total_variation_loss(combination_image)
# 求得梯度，輸入combination_image，對loss求梯度, 每輪迭代中combination_image會根據梯度方向做調整
grads = K.gradients(loss, combination_image)
outputs = [loss]
if isinstance(grads, (list, tuple)):
outputs += grads
else:
outputs.append(grads)
f_outputs = K.function([combination_image], outputs)
evaluator = Evaluator()
x = preprocess_image(source_image)
img = deprocess_image(x.copy())
fname = '原始圖片.png'
save_img(fname, img)
# 開始迭代
for i in range(iterations):
start_time = time.time()
print('迭代', i,end=" ")
x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(), fprime=evaluator.grads, maxfun=20, epsilon=1e-7)
# 一個scipy的L-BFGS優化器
print('目前loss:', min_val,end=" ")
# 保存生成的圖片
img = deprocess_image(x.copy())
fname = 'result_%d.png' % i
end_time = time.time()
print('耗時%.2f s' % (end_time - start_time))
if i%5 == 0 or i == iterations-1:
save_img(fname, img, image_enhance=True)
print('文件保存為', fname)基准圖像：
風格圖像：
合成的藝術風格圖像：
訓練時候整體的loss是3個loss的和，每個loss都有一個系數，調整不同的系數，對應不同的效果。
“內容損失”（content loss）
以下圖片分別對應內容損失系數為0.1、1、5、10的效果：
隨着內容損失系數的增大，迭代優化會更加側重於調整合成圖像的內容，使得圖像跟原始圖像越來越接近。
“風格損失”（style loss）
風格損失是VGG網絡5個CNN層的特征的融合，單純增大風格損失系數對圖像最終風格影響不大，以下是系數是1和100的對比：
系數相差100倍，但是圖像風格並沒有明顯的改變。可能調整5個卷積特征不同的比例系數會有效果。
以下是單純使用第1、2、3、4、5個卷積層特征的效果：
可見 5個卷積層特征里第3和第4個卷積層對圖像的風格影響較大。
以下調整第3和第4個卷積層的系數，5個系數比為1：1：1：1：1和0.5：0.5：0.4：0.4：1
增大第3、4層比例之后，圖像風格更加接近風格圖像。
“差異損失”（variation loss）
圖像差異損失衡量的是圖像本身的局部特征之間的差異，系數越大，圖像局部越接近，表現在圖像上就是圖像像素間過度自然，以下是系數是1、5、10的效果：
以上。