Yolov5網絡結構詳解

Yolov5是目標檢測領域中一種高效的神經網絡結構，是Yolov系列的最新版本。本文將以Yolov5網絡結構為中心，從多個方面對其進行詳細闡述。

一、骨幹網絡

骨幹網絡是指網絡的主幹部分，用於提取圖像的特徵表示。Yolov5的骨幹網絡採用CSPNet（Cross Stage Partial Network）架構，相較於傳統的ResNet等網絡，CSPNet可以顯著減小網絡的參數量和運算量。該網絡結構在既保證檢測精度的情況下，顯著提高了訓練和推理的效率。

import torch.nn as nn
class CSPDarknet(nn.Module):
    def __init__(self, layers):
        super(CSPDarknet, self).__init__()
        self.conv1 = nn.Conv2d(3, 32, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(32)
        self.relu = nn.LeakyReLU(0.1, inplace=True)
        self.layer1 = self.make_layers(32, layers[0])
        self.layer2 = self.make_layers(64, layers[1], stride=2)
        self.layer3 = self.make_layers(128, layers[2], stride=2)
        self.layer4 = self.make_layers(256, layers[3], stride=2)
        self.layer5 = self.make_layers(512, layers[4], stride=2)
        self.layer6 = self.make_layers(1024, layers[5], stride=2)
        self._initialize_weights()

    def make_layers(self, in_channels, num_blocks, stride=1):
        layers = []
        layers.append(('res0', ResBlock(in_channels, in_channels * 2, shortcut=False)))
        for i in range(num_blocks):
            layers.append(('residual_%d' % i, ResBlock(in_channels * 2, in_channels, stride)))
        return nn.Sequential(OrderedDict(layers))

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x1 = self.layer1(x)
        x2 = self.layer2(x1)
        x3 = self.layer3(x2)
        x4 = self.layer4(x3)
        x5 = self.layer5(x4)
        x6 = self.layer6(x5)

        return x4, x5, x6

二、特徵金字塔

目標檢測任務中，不同大小不同層次的目標需要被檢測到，並且需要提取多尺度的特徵。Yolov5使用FPN（Feature Pyramid Network）特徵金字塔結構，通過特徵上採樣和特徵拼接的方式實現多層次、多尺度特徵的融合。它可以同時處理不同尺度的目標，提高模型的檢測效果。

class YOLOv5(nn.Module):
    def __init__(self, cfg, ch=3):
        super(YOLOv5, self).__init__()
        self.ch = ch
        self.model, self.save = parse_model(cfg)
        self.nc = int(self.model[-1]['filters'])
        self.nl = len(self.model)
        self.stem = Focus(ch, 80, 3)  
        self.m = nn.Sequential(*self.model[1:])
        self.init_weights()

    def forward(self, x):
        x = self.stem(x)
        yolo_out, _, _ = [], [], []
        for i in range(self.nl):
            x = self.m[i](x)
            if i in [2, 4, 6]:
                yolo_out.append(x)
            elif i == 8:
                x = self.m[i](x, yolo_out[-1])
                yolo_out.append(x)
        return yolo_out

三、激活函數

激活函數在神經網絡中扮演着至關重要的角色，Yolov5使用的激活函數是Mish。Mish激活函數在保持與ReLU相同的計算速度的同時，提高了模型的精度。

class Mish(nn.Module):
    def __init__(self):
        super(Mish, self).__init__()

    def forward(self, x):
        return x * torch.tanh(F.softplus(x))

class MishModule(nn.Module):
    def __init__(self, parent):
        super(MishModule, self).__init__()
        self.model = parent.model
        for i, m in enumerate(self.model.children()):
            self.model[i] = Mish() if type(m) == nn.ReLU else m

    def forward(self, x):
        return self.model(x)

四、預測頭

Yolov5的預測頭由三個卷積層構成，用於對特徵圖進行輸出通道的降維，並且進行邊界框和目標類別的預測。預測頭可以預測多種不同尺度下的目標，實現多尺度目標檢測。

class Conv(nn.Module):
    def __init__(self, in_channel, out_channel, kernel_size=1, stride=1):
        super().__init__()
        self.conv = nn.Conv2d(in_channel, out_channel, kernel_size, stride, kernel_size // 2, bias=False)
        self.bn = nn.BatchNorm2d(out_channel)
        self.act = nn.LeakyReLU(0.1, inplace=True)

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.act(x)
        return x

class PredictionLayer(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.pred = nn.Sequential(
            Conv(in_channels, in_channels * 2),
            Conv(in_channels * 2, in_channels),
            nn.Conv2d(in_channels, out_channels, kernel_size=1)
        )

    def forward(self, x):
        x = self.pred(x)
        return x

五、總結

Yolov5是目標檢測領域中一種高效的神經網絡結構，採用了CSPNet骨幹網絡和FPN特徵金字塔結構。同時，使用Mish激活函數和預測頭實現多尺度目標檢測。該網絡結構在保證檢測精度的同時，大大提高了訓練和推理的效率，主要應用於實時目標檢測和視頻分析等領域。