全卷積網絡FCN

全卷積網絡FCN

fcn是深度學習用於圖像分割的鼻祖.後續的不少網絡結構都是在此基礎上演進而來.

圖像分割即像素級別的分類.

語義分割的基本框架:
前端fcn(以及在此基礎上的segnet,deconvnet,deeplab等) + 後端crf/mrf

FCN是分割網絡的鼻祖,後面的不少網絡都是在此基礎上提出的.
論文地址

和傳統的分類網絡相比,就是將傳統分類網絡的全鏈接層用反捲積層替代.獲得一個和圖像大小一致的feature map。本篇文章用的網絡是VGG.

主要關注兩點

• 全鏈接層替換成卷積層.用反捲積的方式完成上採樣
• 不一樣layer的輸出要作相加.用以加強feature map的表達能力.

反捲積(deconvolutional)

關於反捲積(也叫轉置卷積)的詳細推導,能夠參考:<https://blog.csdn.net/LoseInVain/article/details/81098502＞

簡單滴說就是:卷積的反向操做．以4x4矩陣Ａ爲例,卷積核Ｃ(3x3,stride=1),經過卷積操做獲得一個2x2的矩陣B. 轉置卷積即已知B,要獲得A,咱們要找到卷積核C,使得B至關於A經過C作正向卷積,獲得B.

轉置卷積是一種上採樣的方法.

跳連(skip layer)

若是隻用特徵提取部分(也就是VGG全鏈接層以前的部分)獲得的feature map作上採樣將feature map還原到圖像輸入的size的話,feature不夠精確.因此採用不一樣layer的feature map作上採樣再組合起來.

代碼解析

源碼:https://github.com/pochih/FCN-pytorch

其中的核心代碼以下:

class FCNs(nn.Module):

    def __init__(self, pretrained_net, n_class):
        super().__init__()
        self.n_class = n_class
        self.pretrained_net = pretrained_net
        self.relu    = nn.ReLU(inplace=True)
        self.deconv1 = nn.ConvTranspose2d(512, 512, kernel_size=3, stride=2, padding=1, dilation=1, output_padding=1)
        self.bn1     = nn.BatchNorm2d(512)
        self.deconv2 = nn.ConvTranspose2d(512, 256, kernel_size=3, stride=2, padding=1, dilation=1, output_padding=1)
        self.bn2     = nn.BatchNorm2d(256)
        self.deconv3 = nn.ConvTranspose2d(256, 128, kernel_size=3, stride=2, padding=1, dilation=1, output_padding=1)
        self.bn3     = nn.BatchNorm2d(128)
        self.deconv4 = nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, dilation=1, output_padding=1)
        self.bn4     = nn.BatchNorm2d(64)
        self.deconv5 = nn.ConvTranspose2d(64, 32, kernel_size=3, stride=2, padding=1, dilation=1, output_padding=1)
        self.bn5     = nn.BatchNorm2d(32)
        self.classifier = nn.Conv2d(32, n_class, kernel_size=1)

    def forward(self, x):
        output = self.pretrained_net(x)
        x5 = output['x5']  # size=(N, 512, x.H/32, x.W/32)
        x4 = output['x4']  # size=(N, 512, x.H/16, x.W/16)
        x3 = output['x3']  # size=(N, 256, x.H/8,  x.W/8)
        x2 = output['x2']  # size=(N, 128, x.H/4,  x.W/4)
        x1 = output['x1']  # size=(N, 64, x.H/2,  x.W/2)

        score = self.bn1(self.relu(self.deconv1(x5)))     # size=(N, 512, x.H/16, x.W/16)
        score = score + x4                                # element-wise add, size=(N, 512, x.H/16, x.W/16)
        score = self.bn2(self.relu(self.deconv2(score)))  # size=(N, 256, x.H/8, x.W/8)
        score = score + x3                                # element-wise add, size=(N, 256, x.H/8, x.W/8)
        score = self.bn3(self.relu(self.deconv3(score)))  # size=(N, 128, x.H/4, x.W/4)
        score = score + x2                                # element-wise add, size=(N, 128, x.H/4, x.W/4)
        score = self.bn4(self.relu(self.deconv4(score)))  # size=(N, 64, x.H/2, x.W/2)
        score = score + x1                                # element-wise add, size=(N, 64, x.H/2, x.W/2)
        score = self.bn5(self.relu(self.deconv5(score)))  # size=(N, 32, x.H, x.W)
        score = self.classifier(score)                    # size=(N, n_class, x.H/1, x.W/1)

        return score  # size=(N, n_class, x.H/1, x.W/1)

train.py中

vgg_model = VGGNet(requires_grad=True, remove_fc=True)
fcn_model = FCNs(pretrained_net=vgg_model, n_class=n_class)

這裏咱們重點看FCN的forward函數

def forward(self, x):
        output = self.pretrained_net(x)
        x5 = output['x5']  # size=(N, 512, x.H/32, x.W/32)
        x4 = output['x4']  # size=(N, 512, x.H/16, x.W/16)
        x3 = output['x3']  # size=(N, 256, x.H/8,  x.W/8)
        x2 = output['x2']  # size=(N, 128, x.H/4,  x.W/4)
        x1 = output['x1']  # size=(N, 64, x.H/2,  x.W/2)

        score = self.bn1(self.relu(self.deconv1(x5)))     # size=(N, 512, x.H/16, x.W/16)
        score = score + x4                                # element-wise add, size=(N, 512, x.H/16, x.W/16)
        score = self.bn2(self.relu(self.deconv2(score)))  # size=(N, 256, x.H/8, x.W/8)
        score = score + x3                                # element-wise add, size=(N, 256, x.H/8, x.W/8)
        score = self.bn3(self.relu(self.deconv3(score)))  # size=(N, 128, x.H/4, x.W/4)
        score = score + x2                                # element-wise add, size=(N, 128, x.H/4, x.W/4)
        score = self.bn4(self.relu(self.deconv4(score)))  # size=(N, 64, x.H/2, x.W/2)
        score = score + x1                                # element-wise add, size=(N, 64, x.H/2, x.W/2)
        score = self.bn5(self.relu(self.deconv5(score)))  # size=(N, 32, x.H, x.W)
        score = self.classifier(score)                    # size=(N, n_class, x.H/1, x.W/1)

        return score  # size=(N, n_class, x.H/1, x.W/1)

可見FCN的輸入爲(batch_size,c,h,w),輸出爲(batch_size,class,h,w).
首先是通過vgg的特徵提取層,能夠獲得feature map. 5個max_pool後的feature map的size分別爲

x5 = output['x5']  # size=(N, 512, x.H/32, x.W/32)
        x4 = output['x4']  # size=(N, 512, x.H/16, x.W/16)
        x3 = output['x3']  # size=(N, 256, x.H/8,  x.W/8)
        x2 = output['x2']  # size=(N, 128, x.H/4,  x.W/4)
        x1 = output['x1']  # size=(N, 64, x.H/2,  x.W/2)

以後每個pool layer的feature map都通過一次2倍上採樣,並與前一個pool layer的輸出進行element-wise add.(resnet也有相似操做).從而使得上採樣後的feature map信息更充分更精準,模型的魯棒性會更好.
例如以輸入圖片尺寸爲224x224爲例,pool4的輸出爲(,512,14,14),pool5的輸出爲(,512,7,7),反捲積後獲得(,512,14,14),再與pool4的輸出作element-wise add。獲得的仍然是(,512,14,14). 對這個輸出作上採樣獲得(,256,28,28)再與pool3的輸出相加. 依次類推,最終獲得(,64,112,112).

此後,再作一次反捲積上採樣獲得(,32,224,224),以後卷積獲得(,n_class,224,224)。即獲得n_class張224x224的feature map。

下圖顯示了隨着上採樣的進行,獲得的feature map細節愈來愈豐富.

損失函數

criterion = nn.BCEWithLogitsLoss()

損失函數採用二分類交叉熵.torch中有2個計算二分類交叉熵的函數

• BCELoss()
• BCEWithLogitsLoss()

後者只是在前者的基礎上,對輸入先作一個sigmoid將輸入轉換到0-1之間.即BCEWithLogitsLoss = Sigmoid + BCELoss

一個具體的例子能夠參考:https://blog.csdn.net/qq_22210253/article/details/85222093