使用 MIRNet 进行弱光图像增强
作者： Soumik Rakshit
创建日期 2021/09/11
最后修改日期 2023/07/15
描述： 实施 MIRNet 架构以进行弱光图像增强。
介绍
为了从其降级版本中恢复高质量图像内容，图像恢复在摄影、安全、医学成像和遥感等众多应用中得到了广泛应用。在本示例中，我们实现了用于弱光图像增强的 **MIRNet** 模型，这是一种全卷积架构，学习一组丰富的特征，这些特征结合了来自多个尺度的上下文信息，同时保留了高分辨率的空间细节。
参考
下载 LOLDataset
**LoL Dataset** 是为弱光图像增强而创建的。它提供了 485 张用于训练的图像和 15 张用于测试的图像。数据集中的每个图像对都包含一张弱光输入图像及其对应的曝光良好的参考图像。
import os
os.environ["KERAS_BACKEND"] = "tensorflow"
import random
import numpy as np
from glob import glob
from PIL import Image, ImageOps
import matplotlib.pyplot as plt
import keras
from keras import layers
import tensorflow as tf
!wget https://huggingface.co/datasets/geekyrakshit/LoL-Dataset/resolve/main/lol_dataset.zip
!unzip -q lol_dataset.zip && rm lol_dataset.zip
--2023-11-10 23:10:00-- https://hugging-face.cn/datasets/geekyrakshit/LoL-Dataset/resolve/main/lol_dataset.zip
Resolving huggingface.co (huggingface.co)... 3.163.189.74, 3.163.189.37, 3.163.189.114, ...
Connecting to huggingface.co (huggingface.co)|3.163.189.74|:443... connected.
HTTP request sent, awaiting response... 302 Found
Location: https://cdn-lfs.huggingface.co/repos/d9/09/d909ef7668bb417b7065a311bd55a3084cc83a1f918e13cb41c5503328432db2/419fddc48958cd0f5599939ee0248852a37ceb8bb738c9b9525e95b25a89de9a?response-content-disposition=attachment%3B+filename*%3DUTF-8%27%27lol_dataset.zip%3B+filename%3D%22lol_dataset.zip%22%3B&response-content-type=application%2Fzip&Expires=1699917000&Policy=eyJTdGF0ZW1lbnQiOlt7IkNvbmRpdGlvbiI6eyJEYXRlTGVzc1RoYW4iOnsiQVdTOkVwb2NoVGltZSI6MTY5OTkxNzAwMH19LCJSZXNvdXJjZSI6Imh0dHBzOi8vY2RuLWxmcy5odWdnaW5nZmFjZS5jby9yZXBvcy9kOS8wOS9kOTA5ZWY3NjY4YmI0MTdiNzA2NWEzMTFiZDU1YTMwODRjYzgzYTFmOTE4ZTEzY2I0MWM1NTAzMzI4NDMyZGIyLzQxOWZkZGM0ODk1OGNkMGY1NTk5OTM5ZWUwMjQ4ODUyYTM3Y2ViOGJiNzM4YzliOTUyNWU5NWIyNWE4OWRlOWE%7EcmVzcG9uc2UtY29udGVudC1kaXNwb3NpdGlvbj0qJnJlc3BvbnNlLWNvbnRlbnQtdHlwZT0qIn1dfQ__&Signature=xyZ1oUBOnWdy6-vCAFzqZsDMetsPu6OSluyOoTS%7EKRZ6lvAy8yUwQgp5WjcZGJ7Jnex0IdnsPiUzsxaxjM-eZjUcQGPdGj4WhSV5DUBxr8xkwTEospYSg1fX%7EE2I1KkP9gBsXvinsKIOAZzchbg9f28xxdlvTbZ0h4ndcUfbDPknwlU1CIZNa5qjU6NqLMH2bPQmI1AIVau2DgQC%7E1n2dgTZsMfHTVmoM2ivsAl%7E9XgQ3m247ke2aj5BmgssZF52VWKTE-vwYDtbuiem73pS6gS-dZlmXYPE1OSRr2tsDo1cgPEBBtuK3hEnYcOq8jjEZk3AEAbFAJoHKLVIERZ30g__&Key-Pair-Id=KVTP0A1DKRTAX [following]
--2023-11-10 23:10:00-- https://cdn-lfs.huggingface.co/repos/d9/09/d909ef7668bb417b7065a311bd55a3084cc83a1f918e13cb41c5503328432db2/419fddc48958cd0f5599939ee0248852a37ceb8bb738c9b9525e95b25a89de9a?response-content-disposition=attachment%3B+filename*%3DUTF-8%27%27lol_dataset.zip%3B+filename%3D%22lol_dataset.zip%22%3B&response-content-type=application%2Fzip&Expires=1699917000&Policy=eyJTdGF0ZW1lbnQiOlt7IkNvbmRpdGlvbiI6eyJEYXRlTGVzc1RoYW4iOnsiQVdTOkVwb2NoVGltZSI6MTY5OTkxNzAwMH19LCJSZXNvdXJjZSI6Imh0dHBzOi8vY2RuLWxmcy5odWdnaW5nZmFjZS5jby9yZXBvcy9kOS8wOS9kOTA5ZWY3NjY4YmI0MTdiNzA2NWEzMTFiZDU1YTMwODRjYzgzYTFmOTE4ZTEzY2I0MWM1NTAzMzI4NDMyZGIyLzQxOWZkZGM0ODk1OGNkMGY1NTk5OTM5ZWUwMjQ4ODUyYTM3Y2ViOGJiNzM4YzliOTUyNWU5NWIyNWE4OWRlOWE%7EcmVzcG9uc2UtY29udGVudC1kaXNwb3NpdGlvbj0qJnJlc3BvbnNlLWNvbnRlbnQtdHlwZT0qIn1dfQ__&Signature=xyZ1oUBOnWdy6-vCAFzqZsDMetsPu6OSluyOoTS%7EKRZ6lvAy8yUwQgp5WjcZGJ7Jnex0IdnsPiUzsxaxjM-eZjUcQGPdGj4WhSV5DUBxr8xkwTEospYSg1fX%7EE2I1KkP9gBsXvinsKIOAZzchbg9f28xxdlvTbZ0h4ndcUfbDPknwlU1CIZNa5qjU6NqLMH2bPQmI1AIVau2DgQC%7E1n2dgTZsMfHTVmoM2ivsAl%7E9XgQ3m247ke2aj5BmgssZF52VWKTE-vwYDtbuiem73pS6gS-dZlmXYPE1OSRr2tsDo1cgPEBBtuK3hEnYcOq8jjEZk3AEAbFAJoHKLVIERZ30g__&Key-Pair-Id=KVTP0A1DKRTAX
Resolving cdn-lfs.huggingface.co (cdn-lfs.huggingface.co)... 108.138.94.122, 108.138.94.14, 108.138.94.25, ...
Connecting to cdn-lfs.huggingface.co (cdn-lfs.huggingface.co)|108.138.94.122|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 347171015 (331M) [application/zip]
Saving to: ‘lol_dataset.zip’
lol_dataset.zip 100%[===================>] 331.09M 316MB/s in 1.0s
2023-11-10 23:10:01 (316 MB/s) - ‘lol_dataset.zip’ saved [347171015/347171015]
创建 TensorFlow 数据集
我们使用来自 LoL Dataset 训练集的 300 对图像进行训练，并使用剩余的 185 对图像进行验证。我们从图像对中生成大小为 128 x 128 的随机裁剪，用于训练和验证。
random.seed(10)
IMAGE_SIZE = 128
BATCH_SIZE = 4
MAX_TRAIN_IMAGES = 300
def read_image(image_path):
image = tf.io.read_file(image_path)
image = tf.image.decode_png(image, channels=3)
image.set_shape([None, None, 3])
image = tf.cast(image, dtype=tf.float32) / 255.0
return image
def random_crop(low_image, enhanced_image):
low_image_shape = tf.shape(low_image)[:2]
low_w = tf.random.uniform(
shape=(), maxval=low_image_shape[1] - IMAGE_SIZE + 1, dtype=tf.int32
)
low_h = tf.random.uniform(
shape=(), maxval=low_image_shape[0] - IMAGE_SIZE + 1, dtype=tf.int32
)
low_image_cropped = low_image[
low_h : low_h + IMAGE_SIZE, low_w : low_w + IMAGE_SIZE
]
enhanced_image_cropped = enhanced_image[
low_h : low_h + IMAGE_SIZE, low_w : low_w + IMAGE_SIZE
]
# in order to avoid `NONE` during shape inference
low_image_cropped.set_shape([IMAGE_SIZE, IMAGE_SIZE, 3])
enhanced_image_cropped.set_shape([IMAGE_SIZE, IMAGE_SIZE, 3])
return low_image_cropped, enhanced_image_cropped
def load_data(low_light_image_path, enhanced_image_path):
low_light_image = read_image(low_light_image_path)
enhanced_image = read_image(enhanced_image_path)
low_light_image, enhanced_image = random_crop(low_light_image, enhanced_image)
return low_light_image, enhanced_image
def get_dataset(low_light_images, enhanced_images):
dataset = tf.data.Dataset.from_tensor_slices((low_light_images, enhanced_images))
dataset = dataset.map(load_data, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.batch(BATCH_SIZE, drop_remainder=True)
return dataset
train_low_light_images = sorted(glob("./lol_dataset/our485/low/*"))[:MAX_TRAIN_IMAGES]
train_enhanced_images = sorted(glob("./lol_dataset/our485/high/*"))[:MAX_TRAIN_IMAGES]
val_low_light_images = sorted(glob("./lol_dataset/our485/low/*"))[MAX_TRAIN_IMAGES:]
val_enhanced_images = sorted(glob("./lol_dataset/our485/high/*"))[MAX_TRAIN_IMAGES:]
test_low_light_images = sorted(glob("./lol_dataset/eval15/low/*"))
test_enhanced_images = sorted(glob("./lol_dataset/eval15/high/*"))
train_dataset = get_dataset(train_low_light_images, train_enhanced_images)
val_dataset = get_dataset(val_low_light_images, val_enhanced_images)
print("Train Dataset:", train_dataset.element_spec)
print("Val Dataset:", val_dataset.element_spec)
Train Dataset: (TensorSpec(shape=(4, 128, 128, 3), dtype=tf.float32, name=None), TensorSpec(shape=(4, 128, 128, 3), dtype=tf.float32, name=None))
Val Dataset: (TensorSpec(shape=(4, 128, 128, 3), dtype=tf.float32, name=None), TensorSpec(shape=(4, 128, 128, 3), dtype=tf.float32, name=None))
MIRNet 模型
以下是 MIRNet 模型的主要特征
- 一种特征提取模型，它在多个空间尺度上计算互补特征集，同时保持原始高分辨率特征以保留精确的空间细节。
- 一种定期重复的信息交换机制，其中跨多分辨率分支的特征被逐步融合在一起以提高表示学习。
- 一种使用选择性核网络融合多尺度特征的新方法，该网络动态地组合可变感受野并忠实地保留每个空间分辨率处的原始特征信息。
- 一种递归残差设计，它逐步分解输入信号以简化整体学习过程，并允许构建非常深的网络。
选择性核特征融合
选择性核特征融合或 SKFF 模块通过两个操作执行感受野的动态调整：**融合** 和 **选择**。融合操作通过组合来自多分辨率流的信息来生成全局特征描述符。选择操作使用这些描述符来重新校准特征图（来自不同流的特征图），然后进行聚合。
**融合**：SKFF 从三个并行的卷积流接收输入，这些流承载着不同尺度的信息。我们首先使用逐元素求和将这些多尺度特征组合起来，然后在空间维度上应用全局平均池化 (GAP)。接下来，我们应用一个通道降维卷积层以生成紧凑的特征表示，该特征表示通过三个并行的通道上采样卷积层（每个分辨率流一个）传递，并为我们提供了三个特征描述符。
**选择**：此操作对特征描述符应用 softmax 函数以获得相应的激活，这些激活用于自适应地重新校准多尺度特征图。聚合后的特征被定义为相应多尺度特征和特征描述符的乘积之和。
def selective_kernel_feature_fusion(
multi_scale_feature_1, multi_scale_feature_2, multi_scale_feature_3
):
channels = list(multi_scale_feature_1.shape)[-1]
combined_feature = layers.Add()(
[multi_scale_feature_1, multi_scale_feature_2, multi_scale_feature_3]
)
gap = layers.GlobalAveragePooling2D()(combined_feature)
channel_wise_statistics = layers.Reshape((1, 1, channels))(gap)
compact_feature_representation = layers.Conv2D(
filters=channels // 8, kernel_size=(1, 1), activation="relu"
)(channel_wise_statistics)
feature_descriptor_1 = layers.Conv2D(
channels, kernel_size=(1, 1), activation="softmax"
)(compact_feature_representation)
feature_descriptor_2 = layers.Conv2D(
channels, kernel_size=(1, 1), activation="softmax"
)(compact_feature_representation)
feature_descriptor_3 = layers.Conv2D(
channels, kernel_size=(1, 1), activation="softmax"
)(compact_feature_representation)
feature_1 = multi_scale_feature_1 * feature_descriptor_1
feature_2 = multi_scale_feature_2 * feature_descriptor_2
feature_3 = multi_scale_feature_3 * feature_descriptor_3
aggregated_feature = layers.Add()([feature_1, feature_2, feature_3])
return aggregated_feature
双重注意力单元
双重注意力单元 (DAU) 用于提取卷积流中的特征。虽然 SKFF 模块融合了多分辨率分支的信息，但我们还需要一种机制来共享特征张量内的信息，包括空间和通道维度，这由 DAU 模块完成。DAU 会抑制不太有用的特征，只允许更具信息量的特征继续传递。这种特征重新校准是通过使用 **通道注意力** 和 **空间注意力** 机制实现的。
**通道注意力** 分支通过应用挤压和激励操作来利用卷积特征图的通道间关系。给定一个特征图，挤压操作在空间维度上应用全局平均池化，以编码全局上下文，从而产生一个特征描述符。激励算子将此特征描述符传递给两个卷积层，然后进行 sigmoid 门控，并生成激活。最后，通道注意力分支的输出是通过使用输出激活对输入特征图进行重新缩放得到的。
**空间注意力** 分支旨在利用卷积特征的通道间依赖关系。空间注意力的目标是生成一个空间注意力图，并使用它来重新校准传入的特征。为了生成空间注意力图，空间注意力分支首先在通道维度上独立地对输入特征应用全局平均池化和最大池化操作，并将输出连接起来形成一个结果特征图，然后将其传递给卷积和 sigmoid 激活，以获得空间注意力图。然后使用此空间注意力图对输入特征图进行重新缩放。
class ChannelPooling(layers.Layer):
def __init__(self, axis=-1, *args, **kwargs):
super().__init__(*args, **kwargs)
self.axis = axis
self.concat = layers.Concatenate(axis=self.axis)
def call(self, inputs):
average_pooling = tf.expand_dims(tf.reduce_mean(inputs, axis=-1), axis=-1)
max_pooling = tf.expand_dims(tf.reduce_max(inputs, axis=-1), axis=-1)
return self.concat([average_pooling, max_pooling])
def get_config(self):
config = super().get_config()
config.update({"axis": self.axis})
def spatial_attention_block(input_tensor):
compressed_feature_map = ChannelPooling(axis=-1)(input_tensor)
feature_map = layers.Conv2D(1, kernel_size=(1, 1))(compressed_feature_map)
feature_map = keras.activations.sigmoid(feature_map)
return input_tensor * feature_map
def channel_attention_block(input_tensor):
channels = list(input_tensor.shape)[-1]
average_pooling = layers.GlobalAveragePooling2D()(input_tensor)
feature_descriptor = layers.Reshape((1, 1, channels))(average_pooling)
feature_activations = layers.Conv2D(
filters=channels // 8, kernel_size=(1, 1), activation="relu"
)(feature_descriptor)
feature_activations = layers.Conv2D(
filters=channels, kernel_size=(1, 1), activation="sigmoid"
)(feature_activations)
return input_tensor * feature_activations
def dual_attention_unit_block(input_tensor):
channels = list(input_tensor.shape)[-1]
feature_map = layers.Conv2D(
channels, kernel_size=(3, 3), padding="same", activation="relu"
)(input_tensor)
feature_map = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(
feature_map
)
channel_attention = channel_attention_block(feature_map)
spatial_attention = spatial_attention_block(feature_map)
concatenation = layers.Concatenate(axis=-1)([channel_attention, spatial_attention])
concatenation = layers.Conv2D(channels, kernel_size=(1, 1))(concatenation)
return layers.Add()([input_tensor, concatenation])
多尺度残差块
多尺度残差块能够通过保持高分辨率表示来生成空间精确的输出，同时接收来自低分辨率的丰富上下文信息。MRB 由多个（本文中为三个）并联连接的全卷积流组成。它允许在并行流之间交换信息，以便在低分辨率特征的帮助下巩固高分辨率特征，反之亦然。MIRNet 采用递归残差设计（带有跳跃连接），以简化学习过程中的信息流。为了保持我们架构的残差性质，使用残差调整模块执行多尺度残差块中使用的降采样和上采样操作。
# Recursive Residual Modules
def down_sampling_module(input_tensor):
channels = list(input_tensor.shape)[-1]
main_branch = layers.Conv2D(channels, kernel_size=(1, 1), activation="relu")(
input_tensor
)
main_branch = layers.Conv2D(
channels, kernel_size=(3, 3), padding="same", activation="relu"
)(main_branch)
main_branch = layers.MaxPooling2D()(main_branch)
main_branch = layers.Conv2D(channels * 2, kernel_size=(1, 1))(main_branch)
skip_branch = layers.MaxPooling2D()(input_tensor)
skip_branch = layers.Conv2D(channels * 2, kernel_size=(1, 1))(skip_branch)
return layers.Add()([skip_branch, main_branch])
def up_sampling_module(input_tensor):
channels = list(input_tensor.shape)[-1]
main_branch = layers.Conv2D(channels, kernel_size=(1, 1), activation="relu")(
input_tensor
)
main_branch = layers.Conv2D(
channels, kernel_size=(3, 3), padding="same", activation="relu"
)(main_branch)
main_branch = layers.UpSampling2D()(main_branch)
main_branch = layers.Conv2D(channels // 2, kernel_size=(1, 1))(main_branch)
skip_branch = layers.UpSampling2D()(input_tensor)
skip_branch = layers.Conv2D(channels // 2, kernel_size=(1, 1))(skip_branch)
return layers.Add()([skip_branch, main_branch])
# MRB Block
def multi_scale_residual_block(input_tensor, channels):
# features
level1 = input_tensor
level2 = down_sampling_module(input_tensor)
level3 = down_sampling_module(level2)
# DAU
level1_dau = dual_attention_unit_block(level1)
level2_dau = dual_attention_unit_block(level2)
level3_dau = dual_attention_unit_block(level3)
# SKFF
level1_skff = selective_kernel_feature_fusion(
level1_dau,
up_sampling_module(level2_dau),
up_sampling_module(up_sampling_module(level3_dau)),
)
level2_skff = selective_kernel_feature_fusion(
down_sampling_module(level1_dau),
level2_dau,
up_sampling_module(level3_dau),
)
level3_skff = selective_kernel_feature_fusion(
down_sampling_module(down_sampling_module(level1_dau)),
down_sampling_module(level2_dau),
level3_dau,
)
# DAU 2
level1_dau_2 = dual_attention_unit_block(level1_skff)
level2_dau_2 = up_sampling_module((dual_attention_unit_block(level2_skff)))
level3_dau_2 = up_sampling_module(
up_sampling_module(dual_attention_unit_block(level3_skff))
)
# SKFF 2
skff_ = selective_kernel_feature_fusion(level1_dau_2, level2_dau_2, level3_dau_2)
conv = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(skff_)
return layers.Add()([input_tensor, conv])
MIRNet 模型
def recursive_residual_group(input_tensor, num_mrb, channels):
conv1 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(input_tensor)
for _ in range(num_mrb):
conv1 = multi_scale_residual_block(conv1, channels)
conv2 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(conv1)
return layers.Add()([conv2, input_tensor])
def mirnet_model(num_rrg, num_mrb, channels):
input_tensor = keras.Input(shape=[None, None, 3])
x1 = layers.Conv2D(channels, kernel_size=(3, 3), padding="same")(input_tensor)
for _ in range(num_rrg):
x1 = recursive_residual_group(x1, num_mrb, channels)
conv = layers.Conv2D(3, kernel_size=(3, 3), padding="same")(x1)
output_tensor = layers.Add()([input_tensor, conv])
return keras.Model(input_tensor, output_tensor)
model = mirnet_model(num_rrg=3, num_mrb=2, channels=64)
训练
- 我们使用 **Charbonnier 损失** 作为损失函数，并使用学习率为
1e-4的 **Adam 优化器** 来训练 MIRNet。 - 我们使用 **峰值信噪比** (PSNR) 作为指标，它表示信号的最大可能值（功率）与影响其表示质量的失真噪声功率之间的比率。
def charbonnier_loss(y_true, y_pred):
return tf.reduce_mean(tf.sqrt(tf.square(y_true - y_pred) + tf.square(1e-3)))
def peak_signal_noise_ratio(y_true, y_pred):
return tf.image.psnr(y_pred, y_true, max_val=255.0)
optimizer = keras.optimizers.Adam(learning_rate=1e-4)
model.compile(
optimizer=optimizer,
loss=charbonnier_loss,
metrics=[peak_signal_noise_ratio],
)
history = model.fit(
train_dataset,
validation_data=val_dataset,
epochs=50,
callbacks=[
keras.callbacks.ReduceLROnPlateau(
monitor="val_peak_signal_noise_ratio",
factor=0.5,
patience=5,
verbose=1,
min_delta=1e-7,
mode="max",
)
],
)
def plot_history(value, name):
plt.plot(history.history[value], label=f"train_{name.lower()}")
plt.plot(history.history[f"val_{value}"], label=f"val_{name.lower()}")
plt.xlabel("Epochs")
plt.ylabel(name)
plt.title(f"Train and Validation {name} Over Epochs", fontsize=14)
plt.legend()
plt.grid()
plt.show()
plot_history("loss", "Loss")
plot_history("peak_signal_noise_ratio", "PSNR")
Epoch 1/50
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1699658204.480352 77759 device_compiler.h:187] Compiled cluster using XLA! This line is logged at most once for the lifetime of the process.
75/75 ━━━━━━━━━━━━━━━━━━━━ 445s 686ms/step - loss: 0.2162 - peak_signal_noise_ratio: 61.5549 - val_loss: 0.1358 - val_peak_signal_noise_ratio: 65.2699 - learning_rate: 1.0000e-04
Epoch 2/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 385ms/step - loss: 0.1745 - peak_signal_noise_ratio: 63.1785 - val_loss: 0.1237 - val_peak_signal_noise_ratio: 65.8360 - learning_rate: 1.0000e-04
Epoch 3/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 386ms/step - loss: 0.1681 - peak_signal_noise_ratio: 63.4903 - val_loss: 0.1205 - val_peak_signal_noise_ratio: 65.9048 - learning_rate: 1.0000e-04
Epoch 4/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 385ms/step - loss: 0.1668 - peak_signal_noise_ratio: 63.4793 - val_loss: 0.1185 - val_peak_signal_noise_ratio: 66.0290 - learning_rate: 1.0000e-04
Epoch 5/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1564 - peak_signal_noise_ratio: 63.9205 - val_loss: 0.1217 - val_peak_signal_noise_ratio: 66.1207 - learning_rate: 1.0000e-04
Epoch 6/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 384ms/step - loss: 0.1601 - peak_signal_noise_ratio: 63.9336 - val_loss: 0.1166 - val_peak_signal_noise_ratio: 66.6102 - learning_rate: 1.0000e-04
Epoch 7/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 385ms/step - loss: 0.1600 - peak_signal_noise_ratio: 63.9043 - val_loss: 0.1335 - val_peak_signal_noise_ratio: 65.5639 - learning_rate: 1.0000e-04
Epoch 8/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 382ms/step - loss: 0.1609 - peak_signal_noise_ratio: 64.0606 - val_loss: 0.1135 - val_peak_signal_noise_ratio: 66.9369 - learning_rate: 1.0000e-04
Epoch 9/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 384ms/step - loss: 0.1539 - peak_signal_noise_ratio: 64.3915 - val_loss: 0.1165 - val_peak_signal_noise_ratio: 66.9783 - learning_rate: 1.0000e-04
Epoch 10/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 43s 409ms/step - loss: 0.1536 - peak_signal_noise_ratio: 64.4491 - val_loss: 0.1118 - val_peak_signal_noise_ratio: 66.8747 - learning_rate: 1.0000e-04
Epoch 11/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1449 - peak_signal_noise_ratio: 64.6579 - val_loss: 0.1167 - val_peak_signal_noise_ratio: 66.9626 - learning_rate: 1.0000e-04
Epoch 12/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1501 - peak_signal_noise_ratio: 64.7929 - val_loss: 0.1143 - val_peak_signal_noise_ratio: 66.9400 - learning_rate: 1.0000e-04
Epoch 13/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 380ms/step - loss: 0.1510 - peak_signal_noise_ratio: 64.6816 - val_loss: 0.1302 - val_peak_signal_noise_ratio: 66.0576 - learning_rate: 1.0000e-04
Epoch 14/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1632 - peak_signal_noise_ratio: 63.9234 - val_loss: 0.1146 - val_peak_signal_noise_ratio: 67.0321 - learning_rate: 1.0000e-04
Epoch 15/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 379ms/step - loss: 0.1486 - peak_signal_noise_ratio: 64.7125 - val_loss: 0.1284 - val_peak_signal_noise_ratio: 66.2105 - learning_rate: 1.0000e-04
Epoch 16/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 379ms/step - loss: 0.1482 - peak_signal_noise_ratio: 64.8123 - val_loss: 0.1176 - val_peak_signal_noise_ratio: 66.8114 - learning_rate: 1.0000e-04
Epoch 17/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1459 - peak_signal_noise_ratio: 64.7795 - val_loss: 0.1092 - val_peak_signal_noise_ratio: 67.4173 - learning_rate: 1.0000e-04
Epoch 18/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 378ms/step - loss: 0.1482 - peak_signal_noise_ratio: 64.8821 - val_loss: 0.1175 - val_peak_signal_noise_ratio: 67.0296 - learning_rate: 1.0000e-04
Epoch 19/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1524 - peak_signal_noise_ratio: 64.7275 - val_loss: 0.1028 - val_peak_signal_noise_ratio: 67.8485 - learning_rate: 1.0000e-04
Epoch 20/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 379ms/step - loss: 0.1350 - peak_signal_noise_ratio: 65.6166 - val_loss: 0.1040 - val_peak_signal_noise_ratio: 67.8551 - learning_rate: 1.0000e-04
Epoch 21/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 380ms/step - loss: 0.1383 - peak_signal_noise_ratio: 65.5167 - val_loss: 0.1071 - val_peak_signal_noise_ratio: 67.5902 - learning_rate: 1.0000e-04
Epoch 22/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 379ms/step - loss: 0.1393 - peak_signal_noise_ratio: 65.6293 - val_loss: 0.1096 - val_peak_signal_noise_ratio: 67.2940 - learning_rate: 1.0000e-04
Epoch 23/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1399 - peak_signal_noise_ratio: 65.5146 - val_loss: 0.1044 - val_peak_signal_noise_ratio: 67.6932 - learning_rate: 1.0000e-04
Epoch 24/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 378ms/step - loss: 0.1390 - peak_signal_noise_ratio: 65.7525 - val_loss: 0.1135 - val_peak_signal_noise_ratio: 66.9891 - learning_rate: 1.0000e-04
Epoch 25/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 0s 326ms/step - loss: 0.1333 - peak_signal_noise_ratio: 65.8340
Epoch 25: ReduceLROnPlateau reducing learning rate to 4.999999873689376e-05.
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1332 - peak_signal_noise_ratio: 65.8348 - val_loss: 0.1252 - val_peak_signal_noise_ratio: 66.5684 - learning_rate: 1.0000e-04
Epoch 26/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1547 - peak_signal_noise_ratio: 64.8968 - val_loss: 0.1105 - val_peak_signal_noise_ratio: 67.0688 - learning_rate: 5.0000e-05
Epoch 27/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1269 - peak_signal_noise_ratio: 66.3882 - val_loss: 0.1035 - val_peak_signal_noise_ratio: 67.7006 - learning_rate: 5.0000e-05
Epoch 28/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 30s 405ms/step - loss: 0.1243 - peak_signal_noise_ratio: 66.5826 - val_loss: 0.1063 - val_peak_signal_noise_ratio: 67.2497 - learning_rate: 5.0000e-05
Epoch 29/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 383ms/step - loss: 0.1292 - peak_signal_noise_ratio: 66.1734 - val_loss: 0.1064 - val_peak_signal_noise_ratio: 67.3989 - learning_rate: 5.0000e-05
Epoch 30/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 0s 328ms/step - loss: 0.1304 - peak_signal_noise_ratio: 66.1267
Epoch 30: ReduceLROnPlateau reducing learning rate to 2.499999936844688e-05.
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 382ms/step - loss: 0.1304 - peak_signal_noise_ratio: 66.1294 - val_loss: 0.1109 - val_peak_signal_noise_ratio: 66.8935 - learning_rate: 5.0000e-05
Epoch 31/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1141 - peak_signal_noise_ratio: 67.1338 - val_loss: 0.1145 - val_peak_signal_noise_ratio: 66.8367 - learning_rate: 2.5000e-05
Epoch 32/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1141 - peak_signal_noise_ratio: 66.9369 - val_loss: 0.1132 - val_peak_signal_noise_ratio: 66.9264 - learning_rate: 2.5000e-05
Epoch 33/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1184 - peak_signal_noise_ratio: 66.7723 - val_loss: 0.1090 - val_peak_signal_noise_ratio: 67.1115 - learning_rate: 2.5000e-05
Epoch 34/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 380ms/step - loss: 0.1243 - peak_signal_noise_ratio: 66.4147 - val_loss: 0.1080 - val_peak_signal_noise_ratio: 67.2300 - learning_rate: 2.5000e-05
Epoch 35/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 0s 325ms/step - loss: 0.1230 - peak_signal_noise_ratio: 66.7113
Epoch 35: ReduceLROnPlateau reducing learning rate to 1.249999968422344e-05.
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1229 - peak_signal_noise_ratio: 66.7121 - val_loss: 0.1038 - val_peak_signal_noise_ratio: 67.5288 - learning_rate: 2.5000e-05
Epoch 36/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1181 - peak_signal_noise_ratio: 66.9202 - val_loss: 0.1030 - val_peak_signal_noise_ratio: 67.6249 - learning_rate: 1.2500e-05
Epoch 37/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1086 - peak_signal_noise_ratio: 67.5034 - val_loss: 0.1016 - val_peak_signal_noise_ratio: 67.6940 - learning_rate: 1.2500e-05
Epoch 38/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1127 - peak_signal_noise_ratio: 67.3735 - val_loss: 0.1004 - val_peak_signal_noise_ratio: 68.0042 - learning_rate: 1.2500e-05
Epoch 39/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 379ms/step - loss: 0.1135 - peak_signal_noise_ratio: 67.3436 - val_loss: 0.1150 - val_peak_signal_noise_ratio: 66.9541 - learning_rate: 1.2500e-05
Epoch 40/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 381ms/step - loss: 0.1152 - peak_signal_noise_ratio: 67.1675 - val_loss: 0.1093 - val_peak_signal_noise_ratio: 67.2030 - learning_rate: 1.2500e-05
Epoch 41/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 378ms/step - loss: 0.1191 - peak_signal_noise_ratio: 66.7586 - val_loss: 0.1095 - val_peak_signal_noise_ratio: 67.2424 - learning_rate: 1.2500e-05
Epoch 42/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 30s 405ms/step - loss: 0.1062 - peak_signal_noise_ratio: 67.6856 - val_loss: 0.1092 - val_peak_signal_noise_ratio: 67.2187 - learning_rate: 1.2500e-05
Epoch 43/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 0s 323ms/step - loss: 0.1099 - peak_signal_noise_ratio: 67.6400
Epoch 43: ReduceLROnPlateau reducing learning rate to 6.24999984211172e-06.
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 377ms/step - loss: 0.1099 - peak_signal_noise_ratio: 67.6378 - val_loss: 0.1079 - val_peak_signal_noise_ratio: 67.4591 - learning_rate: 1.2500e-05
Epoch 44/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 378ms/step - loss: 0.1155 - peak_signal_noise_ratio: 67.0911 - val_loss: 0.1019 - val_peak_signal_noise_ratio: 67.8073 - learning_rate: 6.2500e-06
Epoch 45/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 377ms/step - loss: 0.1145 - peak_signal_noise_ratio: 67.1876 - val_loss: 0.1067 - val_peak_signal_noise_ratio: 67.4283 - learning_rate: 6.2500e-06
Epoch 46/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 384ms/step - loss: 0.1077 - peak_signal_noise_ratio: 67.7168 - val_loss: 0.1114 - val_peak_signal_noise_ratio: 67.1392 - learning_rate: 6.2500e-06
Epoch 47/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 377ms/step - loss: 0.1117 - peak_signal_noise_ratio: 67.3210 - val_loss: 0.1081 - val_peak_signal_noise_ratio: 67.3622 - learning_rate: 6.2500e-06
Epoch 48/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 0s 326ms/step - loss: 0.1074 - peak_signal_noise_ratio: 67.7986
Epoch 48: ReduceLROnPlateau reducing learning rate to 3.12499992105586e-06.
75/75 ━━━━━━━━━━━━━━━━━━━━ 29s 380ms/step - loss: 0.1074 - peak_signal_noise_ratio: 67.7992 - val_loss: 0.1101 - val_peak_signal_noise_ratio: 67.3376 - learning_rate: 6.2500e-06
Epoch 49/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 380ms/step - loss: 0.1081 - peak_signal_noise_ratio: 67.5032 - val_loss: 0.1121 - val_peak_signal_noise_ratio: 67.0685 - learning_rate: 3.1250e-06
Epoch 50/50
75/75 ━━━━━━━━━━━━━━━━━━━━ 28s 378ms/step - loss: 0.1077 - peak_signal_noise_ratio: 67.6709 - val_loss: 0.1084 - val_peak_signal_noise_ratio: 67.6183 - learning_rate: 3.1250e-06
推理
def plot_results(images, titles, figure_size=(12, 12)):
fig = plt.figure(figsize=figure_size)
for i in range(len(images)):
fig.add_subplot(1, len(images), i + 1).set_title(titles[i])
_ = plt.imshow(images[i])
plt.axis("off")
plt.show()
def infer(original_image):
image = keras.utils.img_to_array(original_image)
image = image.astype("float32") / 255.0
image = np.expand_dims(image, axis=0)
output = model.predict(image, verbose=0)
output_image = output[0] * 255.0
output_image = output_image.clip(0, 255)
output_image = output_image.reshape(
(np.shape(output_image)[0], np.shape(output_image)[1], 3)
)
output_image = Image.fromarray(np.uint8(output_image))
original_image = Image.fromarray(np.uint8(original_image))
return output_image
测试图像上的推理
我们将 MIRNet 增强的 LOLDataset 中的测试图像与通过 PIL.ImageOps.autocontrast() 函数增强的图像进行比较。
您可以使用托管在 Hugging Face Hub 上的训练模型，并在 Hugging Face Spaces 上尝试演示。
for low_light_image in random.sample(test_low_light_images, 6):
original_image = Image.open(low_light_image)
enhanced_image = infer(original_image)
plot_results(
[original_image, ImageOps.autocontrast(original_image), enhanced_image],
["Original", "PIL Autocontrast", "MIRNet Enhanced"],
(20, 12),
)
