YOLO中backbone.

### YOLO 模型 Backbone 组件详解 #### 3.1 CBL (Convolution-BatchNormalization-LeakyReLU) YOLOv5 的 Backbone 中，CBL 是一种常见的基础构建单元。该结构由卷积层(Convolution)、批标准化(Batch Normalization) 和激活函数(Leaky ReLU) 构成[^1]。这种组合有助于加速收敛过程，并提高模型性能。 ```python import torch.nn as nn def cbl_block(in_channels, out_channels, kernel_size=3, stride=1, padding=None): if not padding: padding = (kernel_size - 1) // 2 return nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=False), nn.BatchNorm2d(out_channels), nn.LeakyReLU(0.1) ) ``` #### BottleneckCSP / C3 为了进一步优化特征提取能力，在某些版本中引入了改进版残差连接——BottleneckCSP 或者称为 C3 结构。它通过跨阶段部分链接的方式增强了信息传递效率，减少了梯度消失现象的发生概率。 ```python class CSPBlock(nn.Module): def __init__(self, in_channels, out_channels, num_blocks=1): super(CSPBlock, self).__init__() hidden_channels = int(out_channels/2) self.conv1 = cbl_block(in_channels, hidden_channels, 1, 1) self.conv2 = cbl_block(hidden_channels, hidden_channels, 3, 1) self.blocks = nn.Sequential(*[ ResidualBlock(hidden_channels) for _ in range(num_blocks) ]) def forward(self, x): y = self.conv1(x) y = self.blocks(y) z = self.conv2(y) return torch.cat([z,x], dim=1) ``` #### SPP(Spatial Pyramid Pooling)/SPPF 空间金字塔池化（Spatial Pyramid Pooling,SPP）及其变体SPPF被用于增强感受野范围内的上下文理解力。这些模块允许网络捕捉不同尺度下的视觉模式，从而提高了检测效果。 ```python class SPPLayer(nn.Module): """ Space to Depth layer """ def __init__(self, size=(5,9,13)): super(SPPLayer,self).__init__() self.maxpool_layers = nn.ModuleList([ nn.MaxPool2d(kernel_size=k,stride=1,padding=k//2) for k in size]) def forward(self, input_tensor): features = [layer(input_tensor) for layer in self.maxpool_layers] cat_features = torch.cat(features+[input_tensor],dim=1) return cat_features ```