Yolov8有效涨点：YOLOv8-AM，添加多种注意力模块提高检测精度，含代码，超详细

kay_5452024-02-28 10:20
前言

2023 年，Ultralytics 推出了最新版本的 YOLO 模型。注意力机制是提高模型性能最热门的方法之一。

本次介绍的是YOLOv8-AM，它将注意力机制融入到原始的YOLOv8架构中。具体来说，我们分别采用四个注意力模块：卷积块注意力模块（CBAM）、全局注意力机制（GAM）、高效通道注意力（ECA）和随机注意力（SA）来设计改进模型并在数据集上进行测试。实验结果表明，基于ResBlock + CBAM（ResCBAM）的YOLOv8-AM模型在IoU 50（mAP 50）下的平均精度提到了2.2%，达到了state-of-the-art（SOTA）表现。相反，结合GAM的YOLOv8-AM模型获得了的mAP @50并不是一个令人满意的增强。因此，我们将ResBlock和GAM结合起来，引入ResGAM设计另一个新的YOLOv8-AM模型，获得一个较为满意的结果。

目录

前言

注意力机制：

[Convolutional Block Attention Module](#Convolutional Block Attention Module)

[Efficient Channel Attention](#Efficient Channel Attention)

[Shuffle Attention](#Shuffle Attention)

[Global Attention Mechanism](#Global Attention Mechanism)

实验结果（供参考）

[可论文指导--------->v jiabei-545](#可论文指导--------->v jiabei-545)

[改进代码(失效+ v 👆)](#改进代码(失效+ v 👆))

注意力机制：

带有YOLOv8-AM的结构图

YOLOv8 架构由四个关键组件组成：Backbone、Neck、Head 和 Loss Function。 Backbone 融合了 Cross Stage Partial (CSP) 概念，具有减少计算负载、同时增强 CNN 学习能力的优势。如图所示，YOLOv8与采用C3模块的YOLOv5不同，采用C2f模块，该模块集成了C3模块和YOLOv7中的扩展ELAN（E-ELAN）概念。
YOLOv8-AM模型架构详解，其中注意力模块为Shuffle Attention（SA）、Efficient Channel Attention（ECA）、Global Attention Mechanism（GAM）、ResBlock + Convolutional Block Attention Module（ResCBAM）

Convolutional Block Attention Module

CBAM架构

CBAM 包括通道注意力（C-Attention）和空间注意力（S-Attention），如图所示。给定一个中间特征图，CBAM 通过等式依次推断出 1D 通道注意力图 和 2D 空间注意力图 。
ResBlock + Convolutional Block Attention Module

原理和resnet一样

# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 9  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]  # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
  s: [0.33, 0.50, 1024]  # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
  m: [0.67, 0.75, 768]   # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
  l: [1.00, 1.00, 512]   # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
  x: [1.00, 1.25, 512]   # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs

# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]]  # 9

# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  - [-1, 3, C2f, [512]]  # 12
  - [-1, 1, ResBlock_CBAM, [512]]

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  - [-1, 3, C2f, [256]]  # 16 (P3/8-small)
  - [-1, 1, ResBlock_CBAM, [256]]

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]]  # cat head P4
  - [-1, 3, C2f, [512]]  # 20 (P4/16-medium)
  - [-1, 1, ResBlock_CBAM, [512]]

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]]  # cat head P5
  - [-1, 3, C2f, [1024]]  # 24 (P5/32-large)
  - [-1, 1, ResBlock_CBAM, [1024]]

  - [[17, 21, 25], 1, Detect, [nc]]  # Detect(P3, P4, P5)
Efficient Channel Attention

Efficient Channel Attention

ECA 主要包含跨通道交互和具有自适应卷积核的一维卷积，如图 所示。跨通道交互代表了一种组合特征的新方法，增强了特定语义的特征表达。

# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 9  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]  # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
  s: [0.33, 0.50, 1024]  # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
  m: [0.67, 0.75, 768]   # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
  l: [1.00, 1.00, 512]   # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
  x: [1.00, 1.25, 512]   # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs

# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]]  # 9

# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  - [-1, 3, C2f, [512]]  # 12
  - [-1, 1, ECAAttention, [512]]

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  - [-1, 3, C2f, [256]]  # 16 (P3/8-small)
  - [-1, 1, ECAAttention, [256]]

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]]  # cat head P4
  - [-1, 3, C2f, [512]]  # 20 (P4/16-medium)
  - [-1, 1, ECAAttention, [512]]

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]]  # cat head P5
  - [-1, 3, C2f, [1024]]  # 24 (P5/32-large)
  - [-1, 1, ECAAttention, [1024]]

  - [[17, 21, 25], 1, Detect, [nc]]  # Detect(P3, P4, P5)
Shuffle Attention

Shuffle Attention

SA将输入特征图分为不同的组，利用Shuffle Unit将通道注意力和空间注意力整合到每个组的一个块中，如图所示。随后，子特征被聚合，并且" ShuffleNetV2 中使用的"Channel Shuffle"算子用于促进各种子特征之间的信息通信。对于通道注意力，SA 采用 GAP 来捕获和嵌入子特征。此外，使用带有 sigmoid 函数的简单门控机制来创建紧凑的函数，以促进精确和自适应的选择。

# SA.yaml
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 9  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]  # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
  s: [0.33, 0.50, 1024]  # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
  m: [0.67, 0.75, 768]   # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
  l: [1.00, 1.00, 512]   # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
  x: [1.00, 1.25, 512]   # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs

# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]]  # 9

# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  - [-1, 3, C2f, [512]]  # 12
  - [-1, 1, ShuffleAttention, [512]]

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  - [-1, 3, C2f, [256]]  # 16 (P3/8-small)
  - [-1, 1, ShuffleAttention, [256]]

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]]  # cat head P4
  - [-1, 3, C2f, [512]]  # 20 (P4/16-medium)
  - [-1, 1, ShuffleAttention, [512]]

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]]  # cat head P5
  - [-1, 3, C2f, [1024]]  # 24 (P5/32-large)
  - [-1, 1, ShuffleAttention, [1024]]

  - [[17, 21, 25], 1, Detect, [nc]]  # Detect(P3, P4, P5)
Global Attention Mechanism

Global Attention Mechanism

GAM采用了CBAM提出的由通道注意力和空间注意力组成的主要架构，并重新设计了子模块，如图所示。此外，我在GAM内的各层之间添加了快捷连接，这使得输入能够更快地向前传播。

# GAM.yaml
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect

# Parameters
nc: 9  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]  # YOLOv8n summary: 225 layers,  3157200 parameters,  3157184 gradients,   8.9 GFLOPs
  s: [0.33, 0.50, 1024]  # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients,  28.8 GFLOPs
  m: [0.67, 0.75, 768]   # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients,  79.3 GFLOPs
  l: [1.00, 1.00, 512]   # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPs
  x: [1.00, 1.25, 512]   # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOPs
 
# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]]  # 9
 
# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
  - [-1, 3, C2f, [512]]  # 12
  - [-1, 1, GAM_Attention, [512,512]]
 
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
  - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
  - [-1, 3, C2f, [256]]  # 16 (P3/8-small)
  - [-1, 1, GAM_Attention, [256,256]]
 
  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]]  # cat head P4
  - [-1, 3, C2f, [512]]  # 20 (P4/16-medium)
  - [-1, 1, GAM_Attention, [512,512]]
 
  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]]  # cat head P5
  - [-1, 3, C2f, [1024]]  # 24 (P5/32-large)
  - [-1, 1, GAM_Attention, [1024,1024]]
 
  - [[17, 21, 25], 1, Detect, [nc]]  # Detect(P3, P4, P5)

ResBlock + Global Attention Mechanism

原理和resnet一样

实验结果（供参考）

ResBlock + Convolutional Block Attention Module
Shuffle Attention

Efficient Channel Attention
Global Attention Mechanism

ResBlock + Global Attention Mechanism
定量比较（Precision/Recall/F1/mAP）

可论文指导--------->v jiabei-545
改进代码(失效+ v 👆)

链接: https://pan.baidu.com/s/1Fi7ghwJ6XiXrDDnoCvlvrQ?pwd=zk88 提取码: zk88

欢迎大家在评论区进行讨论

相关推荐
love_and_hope1 分钟前
Pytorch学习--神经网络--搭建小实战（手撕CIFAR 10 model structure）和 Sequential 的使用
人工智能·pytorch·python·深度学习·学习
2403_8757368717 分钟前
道品科技智慧农业中的自动气象检测站
网络·人工智能·智慧城市
学术头条41 分钟前
AI 的「phone use」竟是这样练成的，清华、智谱团队发布 AutoGLM 技术报告
人工智能·科技·深度学习·语言模型
准橙考典42 分钟前
怎么能更好的通过驾考呢？
人工智能·笔记·自动驾驶·汽车·学习方法
ai_xiaogui1 小时前
AIStarter教程：快速学会卸载AI项目【AI项目管理平台】
人工智能·ai作画·语音识别·ai写作·ai软件
孙同学要努力1 小时前
《深度学习》——深度学习基础知识（全连接神经网络）
人工智能·深度学习·神经网络
喵~来学编程啦2 小时前
【论文精读】LPT: Long-tailed prompt tuning for image classification
人工智能·深度学习·机器学习·计算机视觉·论文笔记
深圳市青牛科技实业有限公司2 小时前
【青牛科技】应用方案｜D2587A高压大电流DC-DC
人工智能·科技·单片机·嵌入式硬件·机器人·安防监控
水豚AI课代表2 小时前
分析报告、调研报告、工作方案等的提示词
大数据·人工智能·学习·chatgpt·aigc
几两春秋梦_2 小时前
符号回归概念
人工智能·数据挖掘·回归
热门推荐
01【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总02(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明03组基轨迹建模 GBTM的介绍与实现（Stata 或 R）04【AIGC】重塑未来的科技巨轮05全面解析：构建基于深度学习的安全帽检测系统（UI界面+YOLO代码+数据集）06【经验分享】Ubuntu22.04安装微信(linux官方版)07基于YOLOv10深度学习的CT扫描图像肾结石智能检测系统【python源码+Pyqt5界面+数据集+训练代码】深度学习实战、目标检测08Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO09RAG 实践- Ollama+RagFlow 部署本地知识库10【TC3xx芯片】TC3xx芯片电源管理系统PMS详解