yolov8neck端

YOLOv8 Neck 架构概述
YOLOv8 的颈部网络（Neck）作为连接骨干网（Backbone）和头部网络（Head）的关键组件，在目标检测任务中起着至关重要的作用。其设计旨在通过多尺度特征融合来提升模型性能，特别是在处理不同大小的目标时表现出色[^1]。
特征金字塔网络（FPN）
FPN 是一种经典的自上而下的结构，用于构建多层次的语义表示。在 YOLOv8 中，FPN 主要负责将高层抽象特征逐步下采样并与低层高分辨率特征相结合，从而形成一系列具有丰富空间信息和上下文理解能力的新特征图。这种机制有助于提高对小物体识别的效果[^3]。
class FPN(nn.Module):
def __init__(self, in_channels_list, out_channels):
super(FPN, self).__init__()
self.inner_blocks = nn.ModuleList()
self.layer_blocks = nn.ModuleList()

for idx, in_channels in enumerate(in_channels_list[::-1]):
inner_block_module = ConvBNReLU(
in_channels,
out_channels,
kernel_size=1,
stride=1
)
layer_block_module = ConvBNReLU(
out_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1
)

self.inner_blocks.append(inner_block_module)
self.layer_blocks.append(layer_block_module)

def forward(self, x):
last_inner = getattr(x[-1], "inner", None)
results = []

for feature, conv_in, conv_out in zip(
reversed(x),
self.inner_blocks,
self.layer_blocks
):
if last_inner is not None:
last_inner = conv_in(last_inner)
y = resize_like(last_inner, feature)
last_inner = torch.add(y, feature)

result = conv_out(last_inner)
results.insert(0, result)
last_inner = result

return tuple(results)
登录后复制
Path Aggregation Network (PANet)
为了进一步优化特征传播路径并减少计算成本，YOLOv8 引入了 PANet 结构。该方法不仅继承了传统 FPN 自顶向下的特性，还增加了横向连接以及额外的一条由底至上的通路，使得浅层特征能够更有效地参与最终预测过程。这有效解决了梯度消失问题，并增强了全局感受野范围内的交互效果。
class PANet(nn.Module):
def __init__(self, fpn, bottom_up_channels):
super(PANet, self).__init__()
self.fpn = fpn
num_levels = len(bottom_up_channels)
lateral_convs = []
output_convs = []

for i in range(num_levels - 1): # Only need N-1 levels.
lateral_conv = ConvBNReLU(
bottom_up_channels[i],
bottom_up_channels[i + 1],
kernel_size=1,
stride=1
)
output_conv = ConvBNReLU(
bottom_up_channels[i + 1],
bottom_up_channels[i + 1],
kernel_size=3,
stride=1,
padding=1
)
lateral_convs.append(lateral_conv)
output_convs.append(output_conv)

self.lateral_convs = nn.ModuleList(lateral_convs[::-1])
self.output_convs = nn.ModuleList(output_convs[::-1])

def forward(self, inputs):
outputs = list(inputs)
last_inner = self.fpn(outputs)[::-1][0]

for idx, (lateral_conv, output_conv) in enumerate(zip(
self.lateral_convs,
self.output_convs
)):
cur_top_down = lateral_conv(outputs[idx]) if idx != 0 else last_inner
last_inner = cur_top_down + F.interpolate(
last_inner,
scale_factor=2.,
mode='nearest'
) if idx != 0 else cur_top_down

last_inner = output_conv(last_inner)
outputs[idx] = last_inner

return tuple(outputs[:-1][::-1])
登录后复制