超分辨重建
通过卷积神经网络提升图像的分辨率，本文采用一个简单的模型来实现对图片画质提升，测试数据来自《office》中的部分剧照，由于画面原始尺寸较大，所以是对原始画面切片后的每一片进行分辨率提升，然后在重组,训练数据也是基于每个图片的切片(Patch)进行训练。
训练和数据制作过程如下：
降低分辨率： 切割图片，补丁之间有重复
训练模型，学习低分辨率 →\to→ 高分辨率的映射关系
完整代码如下：
import pathlibimport numpy as npimport matplotlib.pyplot as pltfrom glob import globfrom PIL import Imagefrom tqdm import tqdmfrom tensorflow.keras import Modelfrom tensorflow.keras.optimizers import Adamfrom tensorflow.keras.layers import Conv2D, ReLU, Inputfrom tensorflow.keras.preprocessing.image import img_to_array, load_img
def build_srcnn(height, width, depth):input = Input(shape=(height, width, depth))x = Conv2D(filters=64, kernel_size=(7, 7), strides=1, padding='valid')(input)x = ReLU()(x)x = Conv2D(filters=64, kernel_size=(5, 5), strides=1, padding='valid')(x)x = ReLU()(x)x = Conv2D(filters=32, kernel_size=(5, 5), strides=1, padding='valid')(x)x = ReLU()(x)x = Conv2D(filters=32, kernel_size=(5, 5), strides=1, padding='valid')(x)x = ReLU()(x)x = Conv2D(filters=32, kernel_size=(3, 3), strides=1, padding='valid')(x)x = ReLU()(x)x = Conv2D(filters=32, kernel_size=(3, 3), strides=1, padding='valid')(x)x = ReLU()(x)x = Conv2D(filters=32, kernel_size=(3, 3), strides=1, padding='valid')(x)x = ReLU()(x)output = Conv2D(filters=depth, kernel_size=(3, 3))(x)return Model(input, output)
缩放图像
SCALE = 2.5 # 图像缩放系数INPUT_DIM = 51# 输入图像Patch 的大小LABEL_SIZE = 25# 标签Patch的大小PAD = int((INPUT_DIM - LABEL_SIZE)/2.0)STRIDE = 15 # 图像剪裁间隔(步长)
def resize_image(image_array, factor):original_image = Image.fromarray(image_array)new_size = np.array(original_image.size) * factornew_size = new_size.astype(np.int32)new_size = tuple(new_size)resized = original_image.resize(new_size)resized = img_to_array(resized)resized = resized.astype(np.uint8)return resized
def crop_image(image):height, width = image.shape[:2]width -= int(width % SCALE)height -= int(height % SCALE)return image[:height, :width]
def downsize_upsize_image(image):'''get low resolution'''scaled = resize_image(image, 1.0/SCALE)scaled = resize_image(scaled, SCALE/1.0)return scaled
def crop_input(image, w, h):h_slice = slice(h, h+INPUT_DIM)w_slice = slice(w, w+INPUT_DIM)return image[h_slice, w_slice]def crop_output(image, w, h):h_slice = slice(h+PAD, h+PAD+LABEL_SIZE)w_slice = slice(w+PAD, w+PAD+LABEL_SIZE)return image[h_slice, w_slice]
制作训练数据
file_patten = str(pathlib.Path.home()/'DeepVision/frames'/'*.png')DataSetPaths = [*glob(file_patten)]dataset_paths = np.random.choice(DataSetPaths, 2000)
data = []labels = []for image_path in tqdm(dataset_paths, ncols=80):image = load_img(image_path)image = img_to_array(image)image = image.astype(np.uint8)image = crop_image(image)scaled = downsize_upsize_image(image)height, width = image.shape[:2]for h in range(0, height - INPUT_DIM + 1, STRIDE):for w in range(0, width - INPUT_DIM + 1, STRIDE):crop = crop_input(scaled, w, h)target = crop_output(image, w, h)data.append(crop)labels.append(target)data = np.array(data)labels = np.array(labels)
100%|███████████████████████████████████████| 2000/2000 [00:28<00:00, 70.73it/s]
训练模型
EPOCHS = 20BATCH_SIZE = 128optimizer = Adam(learning_rate=1e-3, decay=1e-3 / EPOCHS)model = build_srcnn(INPUT_DIM, INPUT_DIM, 3)pile(loss='mse', optimizer=optimizer)history = model.fit(data, labels, batch_size=BATCH_SIZE, epochs=EPOCHS)
