昇思25天学习打卡营第12天|FCN图像语义分割

昇思MindSpore应用实践

本系列文章主要用于记录昇思25天学习打卡营的学习心得。

基于MindSpore的FCN图像语义分割

1、FCN 图像分割简介

全卷积网络（Fully Convolutional Networks，FCN）是UC Berkeley的Jonathan Long等人于2015年在Fully Convolutional Networks for Semantic Segmentation^[1]一文中提出的用于图像语义分割的一种框架。

FCN是首个端到端（end to end）进行像素级（pixel level）预测分类（即语义分割的工作原理）的全卷积网络。

语义分割（Semantic Segmentation）的定义：每个像素对应一个类标签。同一类会被定义成一个区域块，不区分其中单个物体。

•任务：将图像中的每个像素分配到预定义的类别中，从而实现对图像的像素级别理解和分类。

•特点：不区分不同物体的实例，只关注像素所属的语义类别，例如人、车、树等。

•应用：自动驾驶、医学图像分析、视频分析等领域。

输入一张图片，经过前向传播和反向传播训练学习网络中的参数，可以实现端到端地输出语义分割的推理结果，像一个黑盒一样无需额外预处理。

2、构建 FCN 模型

FCN主要用于图像分割领域，是一种端到端的分割方法，是深度学习应用在图像语义分割的开山之作。通过进行像素级的预测直接得出与原图大小相等的label map。因FCN丢弃全连接层替换为全卷积层，网络所有层均为卷积层，故称为全卷积网络。

全卷积神经网络主要使用以下三种技术：

1. 卷积化（Convolutional）

   使用VGG-16作为FCN的骨干网络（backbone）。VGG-16的输入为224*224的RGB图像，输出为1000个预测值。
   VGG-16只能接受固定大小的输入，丢弃了空间坐标，产生非空间输出。VGG-16中共有三个全连接层，全连接层也可视为带有覆盖整个区域的卷积。将全连接层转换为卷积层能使网络输出由一维非空间输出变为二维矩阵，利用输出能生成输入图片映射的heatmap。

   

2. 上采样（Upsample）

   在卷积过程的卷积操作和池化操作会使得特征图的尺寸变小，为得到原图的大小的稠密图像预测，需要对得到的特征图进行上采样操作。使用双线性插值的参数来初始化上采样逆卷积的参数，后通过反向传播来学习非线性上采样。在网络中执行上采样，以通过像素损失的反向传播进行端到端的学习。

   

3. 跳跃结构（Skip Layer）

   利用上采样技巧对最后一层的特征图进行上采样得到原图大小的分割是步长为32像素的预测，称之为FCN-32s。
   由于最后一层的特征图太小，损失过多细节，采用skips结构将更具有全局信息的最后一层预测和更浅层的预测结合，使预测结果获取更多的局部细节。将底层（stride 32）的预测（FCN-32s）进行2倍的上采样得到原尺寸的图像，并与从pool4层（stride 16）进行的预测融合起来（相加），这一部分的网络被称为FCN-16s。随后将这一部分的预测再进行一次2倍的上采样并与从pool3层得到的预测融合起来，这一部分的网络被称为FCN-8s。 Skips结构将深层的全局信息与浅层的局部信息相结合。

FCN 网络特点：

1. 不含全连接层(fc)的全卷积(fully conv)网络，可适应任意尺寸输入。
2. 增大数据尺寸的反卷积(deconv)层，能够输出精细的结果。
3. 结合不同深度层结果的跳跃连接(skip)结构，同时确保鲁棒性和精确性。

基于MindSpore的FCN网络构建代码如下：

import mindspore.nn as nnclass FCN8s(nn.Cell):def __init__(self, n_class):super().__init__()self.n_class = n_class# conv1，输入通道数=输入图像通道数，输出通道数为64，卷积核为3*3，默认滑动步长为1self.conv1 = nn.SequentialCell(  nn.Conv2d(in_channels=3, out_channels=64,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(64),nn.ReLU(),nn.Conv2d(in_channels=64, out_channels=64,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(64),nn.ReLU())# pool1，采用最大池化，池化不改变通道数，卷积核为2*2，滑动步长为2self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)# conv2，输入通道数为64，输出通道数128，卷积核3*3self.conv2 = nn.SequentialCell(nn.Conv2d(in_channels=64, out_channels=128,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(128),nn.ReLU(),nn.Conv2d(in_channels=128, out_channels=128,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(128),nn.ReLU())# pool2，采用最大池化，池化不改变通道数，卷积核为2*2，滑动步长为2self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)# conv3，输入通道数为128，输出通道数256，卷积核3*3self.conv3 = nn.SequentialCell(nn.Conv2d(in_channels=128, out_channels=256,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(256),nn.ReLU(),nn.Conv2d(in_channels=256, out_channels=256,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(256),nn.ReLU(),nn.Conv2d(in_channels=256, out_channels=256,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(256),nn.ReLU())# pool3，采用最大池化，池化不改变通道数，卷积核为2*2，滑动步长为2self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)# conv4，输入通道数为256，输出通道数512，卷积核3*3self.conv4 = nn.SequentialCell(nn.Conv2d(in_channels=256, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(512),nn.ReLU(),nn.Conv2d(in_channels=512, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(512),nn.ReLU(),nn.Conv2d(in_channels=512, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(512),nn.ReLU())# pool4，采用最大池化，池化不改变通道数，卷积核为2*2，滑动步长为2self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2)# conv5，输入通道数为512，输出通道数1024，卷积核3*3self.conv5 = nn.SequentialCell(nn.Conv2d(in_channels=512, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(512),nn.ReLU(),nn.Conv2d(in_channels=512, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),nn.BatchNorm2d(512),nn.ReLU(),nn.Conv2d(in_channels=512, out_channels=512,kernel_size=3, weight_init='xavier_uniform'),# pool5，采用最大池化，池化不改变通道数，卷积核为2*2，滑动步长为2self.pool5 = nn.MaxPool2d(kernel_size=2, stride=2)# conv6，输入通道数为512，输出通道数4096，卷积核为7*7self.conv6 = nn.SequentialCell(nn.Conv2d(in_channels=512, out_channels=4096,kernel_size=7, weight_init='xavier_uniform'),nn.BatchNorm2d(4096),nn.ReLU(),)# conv7，输入通道数为4096，输出通道数4096，卷积核为1*1（像素级）self.conv7 = nn.SequentialCell(nn.Conv2d(in_channels=4096, out_channels=4096,kernel_size=1, weight_init='xavier_uniform'),nn.BatchNorm2d(4096),nn.ReLU(),)# 将特征图从较高的通道数（4096，通常接在一个CNN网络末端的特征维度）减少到分类的类别数（self.n_class）self.score_fr = nn.Conv2d(in_channels=4096, out_channels=self.n_class,kernel_size=1, weight_init='xavier_uniform')# 转置卷积层（反卷积，将特征图的尺寸从较小的尺寸增加到较大的尺寸），用来将特征图的尺寸上采样两倍（stride=2）# 将低分辨率的输出映射回较高分辨率，为最终恢复到原始输入图像的尺寸做准备self.upscore2 = nn.Conv2dTranspose(in_channels=self.n_class, out_channels=self.n_class,kernel_size=4, stride=2, weight_init='xavier_uniform')# 将通道数为512的特征图转换为类别数维度的输出self.score_pool4 = nn.Conv2d(in_channels=512, out_channels=self.n_class,kernel_size=1, weight_init='xavier_uniform')# 转置卷积层，用于将通过score_pool4层处理后的特征图再次上采样两倍self.upscore_pool4 = nn.Conv2dTranspose(in_channels=self.n_class, out_channels=self.n_class,kernel_size=4, stride=2, weight_init='xavier_uniform')# 将通道数为256的特征图转换为类别数维度的输出self.score_pool3 = nn.Conv2d(in_channels=256, out_channels=self.n_class,kernel_size=1, weight_init='xavier_uniform')# 最后一个转置卷积层，使用16x16的卷积核和8倍上采样，将特征图从较低分辨率放大到与原始输入图像相同的尺寸self.upscore8 = nn.Conv2dTranspose(in_channels=self.n_class, out_channels=self.n_class,kernel_size=16, stride=8, weight_init='xavier_uniform')def construct(self, x):x1 = self.conv1(x)p1 = self.pool1(x1)x2 = self.conv2(p1)p2 = self.pool2(x2)x3 = self.conv3(p2)p3 = self.pool3(x3)x4 = self.conv4(p3)p4 = self.pool4(x4)x5 = self.conv5(p4)p5 = self.pool5(x5)x6 = self.conv6(p5)x7 = self.conv7(x6)sf = self.score_fr(x7)u2 = self.upscore2(sf)s4 = self.score_pool4(p4)f4 = s4 + u2u4 = self.upscore_pool4(f4)s3 = self.score_pool3(p3)f3 = s3 + u4out = self.upscore8(f3)return out 

3、数据预处理

由于PASCAL VOC 2012数据集中图像的分辨率大多不一致，无法放在一个tensor中，故输入前需做标准化处理。

将PASCAL VOC 2012数据集与SDB数据集进行混合：

import numpy as np
import cv2
import mindspore.dataset as dsclass SegDataset:def __init__(self,image_mean,image_std,data_file='',batch_size=32,crop_size=512,max_scale=2.0,min_scale=0.5,ignore_label=255,num_classes=21,num_readers=2,num_parallel_calls=4):self.data_file = data_fileself.batch_size = batch_sizeself.crop_size = crop_sizeself.image_mean = np.array(image_mean, dtype=np.float32)self.image_std = np.array(image_std, dtype=np.float32)self.max_scale = max_scaleself.min_scale = min_scaleself.ignore_label = ignore_labelself.num_classes = num_classesself.num_readers = num_readersself.num_parallel_calls = num_parallel_callsmax_scale > min_scaledef preprocess_dataset(self, image, label):image_out = cv2.imdecode(np.frombuffer(image, dtype=np.uint8), cv2.IMREAD_COLOR)label_out = cv2.imdecode(np.frombuffer(label, dtype=np.uint8), cv2.IMREAD_GRAYSCALE)sc = np.random.uniform(self.min_scale, self.max_scale)new_h, new_w = int(sc * image_out.shape[0]), int(sc * image_out.shape[1])image_out = cv2.resize(image_out, (new_w, new_h), interpolation=cv2.INTER_CUBIC)label_out = cv2.resize(label_out, (new_w, new_h), interpolation=cv2.INTER_NEAREST)image_out = (image_out - self.image_mean) / self.image_stdout_h, out_w = max(new_h, self.crop_size), max(new_w, self.crop_size)pad_h, pad_w = out_h - new_h, out_w - new_wif pad_h > 0 or pad_w > 0:image_out = cv2.copyMakeBorder(image_out, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=0)label_out = cv2.copyMakeBorder(label_out, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=self.ignore_label)offset_h = np.random.randint(0, out_h - self.crop_size + 1)offset_w = np.random.randint(0, out_w - self.crop_size + 1)image_out = image_out[offset_h: offset_h + self.crop_size, offset_w: offset_w + self.crop_size, :]label_out = label_out[offset_h: offset_h + self.crop_size, offset_w: offset_w+self.crop_size]if np.random.uniform(0.0, 1.0) > 0.5:image_out = image_out[:, ::-1, :]label_out = label_out[:, ::-1]image_out = image_out.transpose((2, 0, 1))image_out = image_out.copy()label_out = label_out.copy()label_out = label_out.astype("int32")return image_out, label_outdef get_dataset(self):ds.config.set_numa_enable(True)dataset = ds.MindDataset(self.data_file, columns_list=["data", "label"],shuffle=True, num_parallel_workers=self.num_readers)transforms_list = self.preprocess_datasetdataset = dataset.map(operations=transforms_list, input_columns=["data", "label"],output_columns=["data", "label"],num_parallel_workers=self.num_parallel_calls)dataset = dataset.shuffle(buffer_size=self.batch_size * 10)dataset = dataset.batch(self.batch_size, drop_remainder=True)return dataset# 定义创建数据集的参数
IMAGE_MEAN = [103.53, 116.28, 123.675]
IMAGE_STD = [57.375, 57.120, 58.395]
DATA_FILE = "dataset/dataset_fcn8s/mindname.mindrecord"# 定义模型训练参数
train_batch_size = 4
crop_size = 512
min_scale = 0.5
max_scale = 2.0
ignore_label = 255
num_classes = 21# 实例化Dataset
dataset = SegDataset(image_mean=IMAGE_MEAN,image_std=IMAGE_STD,data_file=DATA_FILE,batch_size=train_batch_size,crop_size=crop_size,max_scale=max_scale,min_scale=min_scale,ignore_label=ignore_label,num_classes=num_classes,num_readers=2,num_parallel_calls=4)dataset = dataset.get_dataset()

import numpy as np
import matplotlib.pyplot as pltplt.figure(figsize=(16, 8))# 对训练集中的数据进行展示
for i in range(1, 9):plt.subplot(2, 4, i)show_data = next(dataset.create_dict_iterator())show_images = show_data["data"].asnumpy()show_images = np.clip(show_images, 0, 1)
# 将图片转换HWC格式后进行展示plt.imshow(show_images[0].transpose(1, 2, 0))plt.axis("off")plt.subplots_adjust(wspace=0.05, hspace=0)
plt.show()

4、模型训练

import mindspore
from mindspore import Tensor
import mindspore.nn as nn
from mindspore.train import ModelCheckpoint, CheckpointConfig, LossMonitor, TimeMonitor, Modeldevice_target = "GPU"
mindspore.set_context(mode=mindspore.PYNATIVE_MODE, device_target=device_target)train_batch_size = 4
num_classes = 21
# 初始化模型结构
net = FCN8s(n_class=21)
# 导入vgg16预训练参数
load_vgg16()
# 计算学习率
min_lr = 0.0005
base_lr = 0.05
train_epochs = 1
iters_per_epoch = dataset.get_dataset_size()
total_step = iters_per_epoch * train_epochslr_scheduler = mindspore.nn.cosine_decay_lr(min_lr,base_lr,total_step,iters_per_epoch,decay_epoch=2)
lr = Tensor(lr_scheduler[-1])# 定义损失函数
# 语义分割是对图像中每个像素点进行分类，仍是分类问题，故损失函数选择交叉熵损失函数来计算FCN网络输出与mask之间的交叉熵损失
# mindspore中使用mindspore.nn.CrossEntropyLoss()作为交叉熵损失函数。loss = nn.CrossEntropyLoss(ignore_index=255)
# 定义优化器
optimizer = nn.Momentum(params=net.trainable_params(), learning_rate=lr, momentum=0.9, weight_decay=0.0001)
# 定义loss_scale
scale_factor = 4
scale_window = 3000
loss_scale_manager = ms.amp.DynamicLossScaleManager(scale_factor, scale_window)
# 初始化模型
if device_target == "Ascend":model = Model(net, loss_fn=loss, optimizer=optimizer, loss_scale_manager=loss_scale_manager, metrics={"pixel accuracy": PixelAccuracy(), "mean pixel accuracy": PixelAccuracyClass(), "mean IoU": MeanIntersectionOverUnion(), "frequency weighted IoU": FrequencyWeightedIntersectionOverUnion()})
else:model = Model(net, loss_fn=loss, optimizer=optimizer, metrics={"pixel accuracy": PixelAccuracy(), "mean pixel accuracy": PixelAccuracyClass(), "mean IoU": MeanIntersectionOverUnion(), "frequency weighted IoU": FrequencyWeightedIntersectionOverUnion()})# 设置ckpt文件保存的参数
time_callback = TimeMonitor(data_size=iters_per_epoch)
loss_callback = LossMonitor()
callbacks = [time_callback, loss_callback]
save_steps = 330
keep_checkpoint_max = 5
config_ckpt = CheckpointConfig(save_checkpoint_steps=10,keep_checkpoint_max=keep_checkpoint_max)
ckpt_callback = ModelCheckpoint(prefix="FCN8s",directory="./ckpt",config=config_ckpt)
callbacks.append(ckpt_callback)
model.train(train_epochs, dataset, callbacks=callbacks)IMAGE_MEAN = [103.53, 116.28, 123.675]
IMAGE_STD = [57.375, 57.120, 58.395]
DATA_FILE = "dataset/dataset_fcn8s/mindname.mindrecord"# 下载已训练好的权重文件
url = "https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/datasets/FCN8s.ckpt"
download(url, "FCN8s.ckpt", replace=True)
net = FCN8s(n_class=num_classes)ckpt_file = "FCN8s.ckpt"
param_dict = load_checkpoint(ckpt_file)
load_param_into_net(net, param_dict)if device_target == "Ascend":model = Model(net, loss_fn=loss, optimizer=optimizer, loss_scale_manager=loss_scale_manager, metrics={"pixel accuracy": PixelAccuracy(), "mean pixel accuracy": PixelAccuracyClass(), "mean IoU": MeanIntersectionOverUnion(), "frequency weighted IoU": FrequencyWeightedIntersectionOverUnion()})
else:model = Model(net, loss_fn=loss, optimizer=optimizer, metrics={"pixel accuracy": PixelAccuracy(), "mean pixel accuracy": PixelAccuracyClass(), "mean IoU": MeanIntersectionOverUnion(), "frequency weighted IoU": FrequencyWeightedIntersectionOverUnion()})# 实例化Dataset
dataset = SegDataset(image_mean=IMAGE_MEAN,image_std=IMAGE_STD,data_file=DATA_FILE,batch_size=train_batch_size,crop_size=crop_size,max_scale=max_scale,min_scale=min_scale,ignore_label=ignore_label,num_classes=num_classes,num_readers=2,num_parallel_calls=4)
dataset_eval = dataset.get_dataset()
model.eval(dataset_eval)

从训练过程可以看出，基于VGG作为Backbone的FCN网路运算量是比较大的，仅1个epoch还未训练出良好的收敛状态：

由于FCN网络在训练的过程中需要大量的训练数据和训练轮数，这里只提供了小数据单个epoch的训练来演示loss收敛的过程，下文中使用已训练好的权重文件进行模型评估和推理效果的展示。

自定义评价指标 Metrics

这一部分主要对训练出来的模型效果进行评估，为了便于解释，假设如下：共有 $k+1$ 个类（从 $L_0$ 到 $L_k$, 其中包含一个空类或背景）， $p_{ij}$ 表示本属于$i$类但被预测为$j$类的像素数量。即, $p_{ii}$ 表示真正的数量, 而 $p_{ij}{p}_{ji}$ 则分别被解释为假正和假负, 尽管两者都是假正与假负之和。

• Pixel Accuracy(PA, 像素精度)：这是最简单的度量，为标记正确的像素占总像素的比例。

$$PA=\frac{{\sum }_{i=0}^k{p}_{ii}}{{\sum }_{i=0}^k{\sum }_{j=0}^k{p}_{ij}}$$

• Mean Pixel Accuracy(MPA, 均像素精度)：是PA的一种简单提升，计算每个类内被正确分类像素数的比例，之后求所有类的平均。

$$MPA=\frac{1}{k+1}\sum _{i=0}^k\frac{p_{ii}}{{\sum }_{j=0}^k{p}_{ij}}$$

• Mean Intersection over Union(MloU, 均交并比)：为语义分割的标准度量。其计算两个集合的交集和并集之，在语义分割的问题中，这两个集合为真实值（ground truth) 和预测值（predicted segmentation）。这个比例可以变形为正真数 (intersection) 比上真正、假负、假正（并集）之和。在每个类上计算loU，之后平均。

$$MIoU=\frac{1}{k+1}\sum _{i=0}^k\frac{p_{ii}}{{\sum }_{j=0}^k{p}_{ij}+{\sum }_{j=0}^k{p}_{ji}-p_{ii}}$$

• Frequency Weighted Intersection over Union(FWIoU, 频权交井比)：为MloU的一种提升，这种方法根据每个类出现的频率为其设置权重。

$$FWIoU=\frac{1}{{\sum }_{i=0}^k{\sum }_{j=0}^k{p}_{ij}}\sum _{i=0}^k\frac{p_{ii}}{{\sum }_{j=0}^k{p}_{ij}+{\sum }_{j=0}^k{p}_{ji}-p_{ii}}$$

import numpy as np
import mindspore as ms
import mindspore.nn as nn
import mindspore.train as trainclass PixelAccuracy(train.Metric):def __init__(self, num_class=21):super(PixelAccuracy, self).__init__()self.num_class = num_classdef _generate_matrix(self, gt_image, pre_image):mask = (gt_image >= 0) & (gt_image < self.num_class)label = self.num_class * gt_image[mask].astype('int') + pre_image[mask]count = np.bincount(label, minlength=self.num_class**2)confusion_matrix = count.reshape(self.num_class, self.num_class)return confusion_matrixdef clear(self):self.confusion_matrix = np.zeros((self.num_class,) * 2)def update(self, *inputs):y_pred = inputs[0].asnumpy().argmax(axis=1)y = inputs[1].asnumpy().reshape(4, 512, 512)self.confusion_matrix += self._generate_matrix(y, y_pred)def eval(self):pixel_accuracy = np.diag(self.confusion_matrix).sum() / self.confusion_matrix.sum()return pixel_accuracyclass PixelAccuracyClass(train.Metric):def __init__(self, num_class=21):super(PixelAccuracyClass, self).__init__()self.num_class = num_classdef _generate_matrix(self, gt_image, pre_image):mask = (gt_image >= 0) & (gt_image < self.num_class)label = self.num_class * gt_image[mask].astype('int') + pre_image[mask]count = np.bincount(label, minlength=self.num_class**2)confusion_matrix = count.reshape(self.num_class, self.num_class)return confusion_matrixdef update(self, *inputs):y_pred = inputs[0].asnumpy().argmax(axis=1)y = inputs[1].asnumpy().reshape(4, 512, 512)self.confusion_matrix += self._generate_matrix(y, y_pred)def clear(self):self.confusion_matrix = np.zeros((self.num_class,) * 2)def eval(self):mean_pixel_accuracy = np.diag(self.confusion_matrix) / self.confusion_matrix.sum(axis=1)mean_pixel_accuracy = np.nanmean(mean_pixel_accuracy)return mean_pixel_accuracyclass MeanIntersectionOverUnion(train.Metric):def __init__(self, num_class=21):super(MeanIntersectionOverUnion, self).__init__()self.num_class = num_classdef _generate_matrix(self, gt_image, pre_image):mask = (gt_image >= 0) & (gt_image < self.num_class)label = self.num_class * gt_image[mask].astype('int') + pre_image[mask]count = np.bincount(label, minlength=self.num_class**2)confusion_matrix = count.reshape(self.num_class, self.num_class)return confusion_matrixdef update(self, *inputs):y_pred = inputs[0].asnumpy().argmax(axis=1)y = inputs[1].asnumpy().reshape(4, 512, 512)self.confusion_matrix += self._generate_matrix(y, y_pred)def clear(self):self.confusion_matrix = np.zeros((self.num_class,) * 2)def eval(self):mean_iou = np.diag(self.confusion_matrix) / (np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) -np.diag(self.confusion_matrix))mean_iou = np.nanmean(mean_iou)return mean_iouclass FrequencyWeightedIntersectionOverUnion(train.Metric):def __init__(self, num_class=21):super(FrequencyWeightedIntersectionOverUnion, self).__init__()self.num_class = num_classdef _generate_matrix(self, gt_image, pre_image):mask = (gt_image >= 0) & (gt_image < self.num_class)label = self.num_class * gt_image[mask].astype('int') + pre_image[mask]count = np.bincount(label, minlength=self.num_class**2)confusion_matrix = count.reshape(self.num_class, self.num_class)return confusion_matrixdef update(self, *inputs):y_pred = inputs[0].asnumpy().argmax(axis=1)y = inputs[1].asnumpy().reshape(4, 512, 512)self.confusion_matrix += self._generate_matrix(y, y_pred)def clear(self):self.confusion_matrix = np.zeros((self.num_class,) * 2)def eval(self):freq = np.sum(self.confusion_matrix, axis=1) / np.sum(self.confusion_matrix)iu = np.diag(self.confusion_matrix) / (np.sum(self.confusion_matrix, axis=1) + np.sum(self.confusion_matrix, axis=0) -np.diag(self.confusion_matrix))frequency_weighted_iou = (freq[freq > 0] * iu[freq > 0]).sum()return frequency_weighted_iou

5、模型推理结果

import cv2
import matplotlib.pyplot as pltnet = FCN8s(n_class=num_classes)
# 设置超参
ckpt_file = "FCN8s.ckpt"
param_dict = load_checkpoint(ckpt_file)
load_param_into_net(net, param_dict)
eval_batch_size = 4
img_lst = []
mask_lst = []
res_lst = []
# 推理效果展示(上方为输入图片，下方为推理效果图片)
plt.figure(figsize=(8, 5))
show_data = next(dataset_eval.create_dict_iterator())
show_images = show_data["data"].asnumpy()                 # 原始图像
mask_images = show_data["label"].reshape([4, 512, 512])   # 语义分割的掩码图像
show_images = np.clip(show_images, 0, 1)                  # 限制show_images数组中的像素值在归一化后的0和1之间
for i in range(eval_batch_size):img_lst.append(show_images[i])mask_lst.append(mask_images[i])
res = net(show_data["data"]).asnumpy().argmax(axis=1)
for i in range(eval_batch_size):plt.subplot(2, 4, i + 1)plt.imshow(img_lst[i].transpose(1, 2, 0))plt.axis("off")plt.subplots_adjust(wspace=0.05, hspace=0.02)plt.subplot(2, 4, i + 5)plt.imshow(res[i])plt.axis("off")plt.subplots_adjust(wspace=0.05, hspace=0.02)
plt.show()

在推理时，训练好的FCN对输入图像进行前向传播，图像经过FCN网络中的多层卷积、激活和池化后，通过上采样层恢复到原始尺寸，每个像素的输出通常经过 softmax 函数转换成为概率，表明每个像素属于各个类别的概率。选择概率最高的类别作为该像素的类别标签，每个像素对应一个类标签。同一类会被定义成一个区域块，从而实现整个图像的语义分割。

因此，FCN通过这种结构能够有效地在像素级别上对图像进行分类，实现精确的语义分割。

Reference

[1] Long, Jonathan, Evan Shelhamer, and Trevor Darrell. “Fully convolutional networks for Semantic Segmentation.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015.
[2] 昇思大模型平台
[3] 昇思官方文档-FCN图像语义分割
[4] 图像分割概述 - 语义分割、实例分割、全景分割、一键抠图（FCN, U-Net,Mask R-CNN,UPSNet）