YOLO26改进 | 卷积模块 | 空域抓全局 + 频域保细节，小目标检测精度显著提升
·
💡💡💡本专栏所有程序均经过测试，可成功执行💡💡💡
本文给大家带来的教程是将YOLO26的Conv替换为FSConv来提取特征。文章在介绍主要的原理后，将手把手教学如何进行模块的代码添加和修改，并将修改后的完整代码放在文章的最后，方便大家一键运行，小白也可轻松上手实践。以帮助您更好地学习深度学习目标检测YOLO系列的挑战。
目录
1.论文
官方代码：官方代码仓库点击即可跳转
2. FSConv代码实现
2.1 将FSConv添加到YOLO26中
关键步骤一：在ultralytics\ultralytics\nn\modules下面新建文件夹models，在文件夹下新建FSConv.py，粘贴下面代码
import torch, math
import torch.nn as nn
from pytorch_wavelets import DWTForward
from ultralytics.nn.modules.conv import Conv
class SEBlock(nn.Module):
def __init__(self, channel, reduction=16):
super(SEBlock, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
nn.Linear(channel, channel // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channel // reduction, channel, bias=False),
nn.Sigmoid()
)
def forward(self, x):
b, c, _, _ = x.size()
y = self.avg_pool(x).view(b, c)
y = self.fc(y).view(b, c, 1, 1)
return x * y
class FSConv(nn.Module):
def __init__(self, c1, c2, k=3):
super().__init__()
self.c1 = c1
self.conv1 = Conv(c1, 2 * c1, 1, g=c1)
self.wt = DWTForward(J=1, mode='zero', wave='haar')
self.conv2 = Conv(c1, c2, k, 2, g=math.gcd(c1, c2))
self.conv3 = Conv(c1 * 3, c2, 3, d=1, g=math.gcd(c1 * 3, c2))
self.se = SEBlock(c2)
self.conv4 = Conv(c1, c2, 3, g=math.gcd(c1, c2))
self.conv5 = Conv(2 * c2, c2, 1)
def _wavelet_forward_fp32(self, x):
# pytorch_wavelets keeps FP32 filter buffers and its backward path breaks under CUDA AMP
with torch.autocast(device_type=x.device.type, enabled=False):
self.wt = self.wt.float()
return self.wt(x.float())
def forward(self, x):
x0 = self.conv1(x)
x1, x2 = torch.split(x0, self.c1, dim=1)
conv_spatial = self.conv2(x1)
yL, yH = self._wavelet_forward_fp32(x2)
yL = yL.to(dtype=x.dtype)
# Extract the high-frequency subbands
y_HL = yH[0][:, :, 0, :].to(dtype=x.dtype)
y_LH = yH[0][:, :, 1, :].to(dtype=x.dtype)
y_HH = yH[0][:, :, 2, :].to(dtype=x.dtype)
high_frequency_fused = torch.cat([y_HL, y_LH, y_HH], dim=1)
high_frequency_fused_output = self.conv3(high_frequency_fused)
# Apply SE attention
high_frequency_fused_output = self.se(high_frequency_fused_output)
low_frequency_fused_output = self.conv4(yL)
spatial_output = conv_spatial * high_frequency_fused_output
fused = torch.cat([spatial_output, low_frequency_fused_output], dim=1)
out = self.conv5(fused)
return out 2.2 更改init.py文件
关键步骤二：在文件ultralytics\ultralytics\nn\modules\models文件夹下新建__init__.py文件，先导入函数
然后在下面的__all__中声明函数
2.3 添加yaml文件
关键步骤三：在/ultralytics/ultralytics/cfg/models/26下面新建文件yolo26_FSConv.yaml文件，粘贴下面的内容
- 目标检测
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
# Model docs: https://docs.ultralytics.com/models/yolo26
# Task docs: https://docs.ultralytics.com/tasks/detect
# Parameters
nc: 80 # number of classes
end2end: True # whether to use end-to-end mode
reg_max: 1 # DFL bins
scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs
# YOLO26n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, FSConv, [64]] # 0-P1/2
- [-1, 1, FSConv, [128]] # 1-P2/4
- [-1, 2, C3k2, [256, False, 0.25]] # 2-P2/4
- [-1, 1, FSConv, [256]] # 3-P3/8
- [-1, 2, C3k2, [512, False, 0.25]] # 4-P3/8
- [-1, 1, FSConv, [512]] # 5-P4/16
- [-1, 2, C3k2, [512, True]] # 6-P4/16
- [-1, 1, FSConv, [1024]] # 7-P5/32
- [-1, 2, C3k2, [1024, True]] # 8-P5/32
- [-1, 1, SPPF, [1024, 5, 3, True]] # 9-P5/32
- [-1, 2, C2PSA, [1024]] # 10-P5/32
# YOLO26n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 11-P4/16
- [[-1, 6], 1, Concat, [1]] # 12-P4/16
- [-1, 2, C3k2, [512, True]] # 13-P4/16
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 14-P3/8
- [[-1, 4], 1, Concat, [1]] # 15-P3/8
- [-1, 2, C3k2, [256, True]] # 16-P3/8
- [-1, 1, FSConv, [256]] # 17-P4/16
- [[-1, 13], 1, Concat, [1]] # 18-P4/16
- [-1, 2, C3k2, [512, True]] # 19-P4/16
- [-1, 1, FSConv, [512]] # 20-P5/32
- [[-1, 10], 1, Concat, [1]] # 21-P5/32
- [-1, 1, C3k2, [1024, True, 0.5, True]] # 22-P5/32
- [[16, 19, 22], 1, Detect, [nc]] # 23-P3/8,P4/16,P5/32- 语义分割
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
# Model docs: https://docs.ultralytics.com/models/yolo26
# Task docs: https://docs.ultralytics.com/tasks/detect
# Parameters
nc: 80 # number of classes
end2end: True # whether to use end-to-end mode
reg_max: 1 # DFL bins
scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs
# YOLO26n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, FSConv, [64]] # 0-P1/2
- [-1, 1, FSConv, [128]] # 1-P2/4
- [-1, 2, C3k2, [256, False, 0.25]] # 2-P2/4
- [-1, 1, FSConv, [256]] # 3-P3/8
- [-1, 2, C3k2, [512, False, 0.25]] # 4-P3/8
- [-1, 1, FSConv, [512]] # 5-P4/16
- [-1, 2, C3k2, [512, True]] # 6-P4/16
- [-1, 1, FSConv, [1024]] # 7-P5/32
- [-1, 2, C3k2, [1024, True]] # 8-P5/32
- [-1, 1, SPPF, [1024, 5, 3, True]] # 9-P5/32
- [-1, 2, C2PSA, [1024]] # 10-P5/32
# YOLO26n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 11-P4/16
- [[-1, 6], 1, Concat, [1]] # 12-P4/16
- [-1, 2, C3k2, [512, True]] # 13-P4/16
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 14-P3/8
- [[-1, 4], 1, Concat, [1]] # 15-P3/8
- [-1, 2, C3k2, [256, True]] # 16-P3/8
- [-1, 1, FSConv, [256]] # 17-P4/16
- [[-1, 13], 1, Concat, [1]] # 18-P4/16
- [-1, 2, C3k2, [512, True]] # 19-P4/16
- [-1, 1, FSConv, [512]] # 20-P5/32
- [[-1, 10], 1, Concat, [1]] # 21-P5/32
- [-1, 1, C3k2, [1024, True, 0.5, True]] # 22-P5/32
- [[16, 19, 22], 1, Segment, [nc, 32, 256]]- 旋转目标检测
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
# Model docs: https://docs.ultralytics.com/models/yolo26
# Task docs: https://docs.ultralytics.com/tasks/detect
# Parameters
nc: 80 # number of classes
end2end: True # whether to use end-to-end mode
reg_max: 1 # DFL bins
scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs
# YOLO26n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, FSConv, [64]] # 0-P1/2
- [-1, 1, FSConv, [128]] # 1-P2/4
- [-1, 2, C3k2, [256, False, 0.25]] # 2-P2/4
- [-1, 1, FSConv, [256]] # 3-P3/8
- [-1, 2, C3k2, [512, False, 0.25]] # 4-P3/8
- [-1, 1, FSConv, [512]] # 5-P4/16
- [-1, 2, C3k2, [512, True]] # 6-P4/16
- [-1, 1, FSConv, [1024]] # 7-P5/32
- [-1, 2, C3k2, [1024, True]] # 8-P5/32
- [-1, 1, SPPF, [1024, 5, 3, True]] # 9-P5/32
- [-1, 2, C2PSA, [1024]] # 10-P5/32
# YOLO26n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 11-P4/16
- [[-1, 6], 1, Concat, [1]] # 12-P4/16
- [-1, 2, C3k2, [512, True]] # 13-P4/16
- [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 14-P3/8
- [[-1, 4], 1, Concat, [1]] # 15-P3/8
- [-1, 2, C3k2, [256, True]] # 16-P3/8
- [-1, 1, FSConv, [256]] # 17-P4/16
- [[-1, 13], 1, Concat, [1]] # 18-P4/16
- [-1, 2, C3k2, [512, True]] # 19-P4/16
- [-1, 1, FSConv, [512]] # 20-P5/32
- [[-1, 10], 1, Concat, [1]] # 21-P5/32
- [-1, 1, C3k2, [1024, True, 0.5, True]] # 22-P5/32
- [[16, 19, 22], 1, OBB, [nc, 1]]温馨提示：本文只是对yolo26基础上添加模块，如果要对yolo26 n/l/m/x进行添加则只需要指定对应的depth_multiple 和 width_multiple
end2end: True # whether to use end-to-end mode
reg_max: 1 # DFL bins
scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs2.4 在task.py中进行注册
关键步骤四：在parse_model函数中进行注册，添加FSConv
先在task.py导入函数
然后在task.py文件下找到parse_model这个函数，如下图，添加FSConv
elif m in {FSConv}:
c1, c2 = ch[f], args[0]
if c2 != nc: # if c2 != nc (e.g., Classify() output)
c2 = make_divisible(min(c2, max_channels) * width, 8)
args = [c1, c2, *args[1:]]2.5 执行程序
关键步骤五: 在ultralytics文件中新建train.py，将model的参数路径设置为yolo26_FSConv.yaml的路径即可 【注意是在外边的Ultralytics下新建train.py】
from ultralytics import YOLO
import warnings
warnings.filterwarnings('ignore')
from pathlib import Path
if __name__ == '__main__':
# 加载模型
model = YOLO("ultralytics/cfg/26/yolo26.yaml") # 你要选择的模型yaml文件地址
# Use the model
results = model.train(data=r"你的数据集的yaml文件地址",
epochs=100, batch=16, imgsz=640, workers=4, name=Path(model.cfg).stem) # 训练模型🚀运行程序，如果出现下面的内容则说明添加成功🚀
from n params module arguments
0 -1 1 2850 ultralytics.nn.models.FSConv.FSConv [3, 16]
1 -1 1 3968 ultralytics.nn.models.FSConv.FSConv [16, 32]
2 -1 1 6640 ultralytics.nn.modules.block.C3k2 [32, 64, 1, False, 0.25]
3 -1 1 12480 ultralytics.nn.models.FSConv.FSConv [64, 64]
4 -1 1 26080 ultralytics.nn.modules.block.C3k2 [64, 128, 1, False, 0.25]
5 -1 1 42368 ultralytics.nn.models.FSConv.FSConv [128, 128]
6 -1 1 87040 ultralytics.nn.modules.block.C3k2 [128, 128, 1, True]
7 -1 1 153600 ultralytics.nn.models.FSConv.FSConv [128, 256]
8 -1 1 346112 ultralytics.nn.modules.block.C3k2 [256, 256, 1, True]
9 -1 1 164608 ultralytics.nn.modules.block.SPPF [256, 256, 5, 3, True]
10 -1 1 249728 ultralytics.nn.modules.block.C2PSA [256, 256, 1]
11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
12 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1]
13 -1 1 119808 ultralytics.nn.modules.block.C3k2 [384, 128, 1, True]
14 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
15 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1]
16 -1 1 34304 ultralytics.nn.modules.block.C3k2 [256, 64, 1, True]
17 -1 1 12480 ultralytics.nn.models.FSConv.FSConv [64, 64]
18 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1]
19 -1 1 95232 ultralytics.nn.modules.block.C3k2 [192, 128, 1, True]
20 -1 1 42368 ultralytics.nn.models.FSConv.FSConv [128, 128]
21 [-1, 10] 1 0 ultralytics.nn.modules.conv.Concat [1]
22 -1 1 463104 ultralytics.nn.modules.block.C3k2 [384, 256, 1, True, 0.5, True]
23 [16, 19, 22] 1 309656 ultralytics.nn.modules.head.Detect [80, 1, True, [64, 128, 256]]
YOLO26_FSConv summary: 358 layers, 2,172,426 parameters, 2,172,426 gradients, 5.8 GFLOPs3. 完整代码分享
主页侧边
4. GFLOPs
关于GFLOPs的计算方式可以查看：百面算法工程师 | 卷积基础知识——Convolution
未改进的YOLO26n GFLOPs
改进后的GFLOPs
from n params module arguments
0 -1 1 2850 ultralytics.nn.models.FSConv.FSConv [3, 16]
1 -1 1 3968 ultralytics.nn.models.FSConv.FSConv [16, 32]
2 -1 1 6640 ultralytics.nn.modules.block.C3k2 [32, 64, 1, False, 0.25]
3 -1 1 12480 ultralytics.nn.models.FSConv.FSConv [64, 64]
4 -1 1 26080 ultralytics.nn.modules.block.C3k2 [64, 128, 1, False, 0.25]
5 -1 1 42368 ultralytics.nn.models.FSConv.FSConv [128, 128]
6 -1 1 87040 ultralytics.nn.modules.block.C3k2 [128, 128, 1, True]
7 -1 1 153600 ultralytics.nn.models.FSConv.FSConv [128, 256]
8 -1 1 346112 ultralytics.nn.modules.block.C3k2 [256, 256, 1, True]
9 -1 1 164608 ultralytics.nn.modules.block.SPPF [256, 256, 5, 3, True]
10 -1 1 249728 ultralytics.nn.modules.block.C2PSA [256, 256, 1]
11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
12 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1]
13 -1 1 119808 ultralytics.nn.modules.block.C3k2 [384, 128, 1, True]
14 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
15 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1]
16 -1 1 34304 ultralytics.nn.modules.block.C3k2 [256, 64, 1, True]
17 -1 1 12480 ultralytics.nn.models.FSConv.FSConv [64, 64]
18 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1]
19 -1 1 95232 ultralytics.nn.modules.block.C3k2 [192, 128, 1, True]
20 -1 1 42368 ultralytics.nn.models.FSConv.FSConv [128, 128]
21 [-1, 10] 1 0 ultralytics.nn.modules.conv.Concat [1]
22 -1 1 463104 ultralytics.nn.modules.block.C3k2 [384, 256, 1, True, 0.5, True]
23 [16, 19, 22] 1 309656 ultralytics.nn.modules.head.Detect [80, 1, True, [64, 128, 256]]
YOLO26_FSConv summary: 358 layers, 2,172,426 parameters, 2,172,426 gradients, 5.8 GFLOPs5. 进阶
可以与其他的注意力机制或者损失函数等结合，进一步提升检测效果
6.总结
通过以上的改进方法，我们成功提升了模型的表现。这只是一个开始，未来还有更多优化和技术深挖的空间。在这里，我想隆重向大家推荐我的专栏——<专栏地址：YOLO26改进-论文涨点——点击跳转看所有内容，关注不迷路！>。这个专栏专注于前沿的深度学习技术，特别是目标检测领域的最新进展，不仅包含对YOLO26的深入解析和改进策略，还会定期更新来自各大顶会（如CVPR、NeurIPS等）的论文复现和实战分享。
为什么订阅我的专栏？ ——专栏地址：YOLO26改进-论文涨点——点击跳转看所有内容，关注不迷路！
-
前沿技术解读：专栏不仅限于YOLO系列的改进，还会涵盖各类主流与新兴网络的最新研究成果，帮助你紧跟技术潮流。
-
详尽的实践分享：所有内容实践性也极强。每次更新都会附带代码和具体的改进步骤，保证每位读者都能迅速上手。
-
问题互动与答疑：订阅我的专栏后，你将可以随时向我提问，获取及时的答疑。
-
实时更新，紧跟行业动态：不定期发布来自全球顶会的最新研究方向和复现实验报告，让你时刻走在技术前沿。
专栏适合人群：
-
对目标检测、YOLO系列网络有深厚兴趣的同学
-
希望在用YOLO算法写论文的同学
-
对YOLO算法感兴趣的同学等
AtomGit 是由开放原子开源基金会联合 CSDN 等生态伙伴共同推出的新一代开源与人工智能协作平台。平台坚持“开放、中立、公益”的理念，把代码托管、模型共享、数据集托管、智能体开发体验和算力服务整合在一起，为开发者提供从开发、训练到部署的一站式体验。
更多推荐
7
0- 0
已为社区贡献20条内容
所有评论(0)