YOLO26改进 | 卷积模块 | 利用频域特征加强空间细节与纹理表示能力【CVPR2025】

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FDConv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。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目录1.论文2. FDConv代码实现2.1 将FDConv添加到YOLO26中2.2 更改init.py文件2.3 添加yaml文件2.4 在task.py中进行注册2.5 执行程序3. 完整代码分享4. GFLOPs5. 进阶6.总结1.论文论文地址Frequency Dynamic Convolution for Dense Image Prediction官方代码官方代码仓库点击即可跳转2. FDConv代码实现2.1 将FDConv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FDConv.py粘贴下面代码import torch import torch.nn as nn import torch.nn.functional as F from ultralytics.nn.modules.conv import autopad class StarReLU(nn.Module): StarReLU: s * relu(x) ** 2 b def __init__(self, scale_value1.0, bias_value0.0, scale_learnableTrue, bias_learnableTrue, inplaceFalse): super().__init__() self.relu nn.ReLU(inplaceinplace) self.scale nn.Parameter(torch.tensor(scale_value), requires_gradscale_learnable) self.bias nn.Parameter(torch.tensor(bias_value), requires_gradbias_learnable) def forward(self, x): return self.scale * self.relu(x).pow(2) self.bias def get_fft2freq(d1, d2, use_rfftFalse): freq_h torch.fft.fftfreq(d1) freq_w torch.fft.rfftfreq(d2) if use_rfft else torch.fft.fftfreq(d2) grid torch.stack(torch.meshgrid(freq_h, freq_w), dim-1) dist torch.norm(grid, dim-1) flat dist.view(-1) _, idx torch.sort(flat) if use_rfft: d2_ d2 // 2 1 else: d2_ d2 coords torch.stack([idx // d2_, idx % d2_], dim1) return coords.t(), grid class KernelSpatialModulation_Global(nn.Module): def __init__(self, in_planes, out_planes, kernel_size, reduction0.0625, kernel_num4, min_channel16, temp1.0, kernel_tempNone, att_multi2.0, ksm_only_kernel_attFalse, stride1, spatial_freq_decomposeFalse, act_typesigmoid): super().__init__() attention_channel max(int(in_planes * reduction), min_channel) self.kernel_size kernel_size self.temperature temp self.kernel_temp kernel_temp self.att_multi att_multi self.ksm_only_kernel_att ksm_only_kernel_att self.act_type act_type self.fc nn.Conv2d(in_planes, attention_channel, 1, biasFalse) self.bn nn.BatchNorm2d(attention_channel) self.relu StarReLU() # Channel if ksm_only_kernel_att: self.func_channel lambda x: 1.0 else: out_c in_planes * 2 if spatial_freq_decompose and kernel_size1 else in_planes self.channel_fc nn.Conv2d(attention_channel, out_c, 1) self.func_channel self._get_channel # Filter if in_planesout_planes or ksm_only_kernel_att: self.func_filter lambda x: 1.0 else: out_f out_planes * 2 if spatial_freq_decompose and stride1 else out_planes self.filter_fc nn.Conv2d(attention_channel, out_f, 1, stridestride) self.func_filter self._get_filter # Spatial if kernel_size1 or ksm_only_kernel_att: self.func_spatial lambda x: 1.0 else: self.spatial_fc nn.Conv2d(attention_channel, kernel_size*kernel_size,1) self.func_spatial self._get_spatial # Kernel mixing if kernel_num1: self.func_kernel lambda x: 1.0 else: self.kernel_fc nn.Conv2d(attention_channel, kernel_num,1) self.func_kernel self._get_kernel def _get_channel(self, x): att self.channel_fc(x).view(x.size(0),1,1,-1,1,1) if self.act_typesigmoid: return torch.sigmoid(att/self.temperature)*self.att_multi elif self.act_typetanh: return 1torch.tanh(att/self.temperature) raise def _get_filter(self,x): att self.filter_fc(x).view(x.size(0),1,-1,1,1,1) if self.act_typesigmoid: return torch.sigmoid(att/self.temperature)*self.att_multi elif self.act_typetanh: return 1torch.tanh(att/self.temperature) raise def _get_spatial(self,x): att self.spatial_fc(x).view(x.size(0),1,1,1,self.kernel_size,self.kernel_size) if self.act_typesigmoid: return torch.sigmoid(att/self.temperature)*self.att_multi elif self.act_typetanh: return 1torch.tanh(att/self.temperature) raise def _get_kernel(self,x): att self.kernel_fc(x).view(x.size(0),-1,1,1,1,1) return F.softmax(att/self.kernel_temp,dim1) def forward(self,x): g self.relu(self.bn(self.fc(x))) return self.func_channel(g), self.func_filter(g), self.func_spatial(g), self.func_kernel(g) class KernelSpatialModulation_Local(nn.Module): def __init__(self, channel, kernel_num1, out_n1, k_size3, use_globalFalse): super().__init__() self.kn, self.out_n, self.use_global kernel_num, out_n, use_global self.conv1d nn.Conv1d(1, kernel_num*out_n, k_size, padding(k_size-1)//2) self.norm nn.LayerNorm(channel) if use_global: self.complex_w nn.Parameter(torch.randn(1,channel//21,2)*1e-6) def forward(self,x): y x.mean(-1,keepdimTrue).transpose(-1,-2) if self.use_global: fft torch.fft.rfft(y.float(),dim-1) real fft.real*self.complex_w[...,0] imag fft.imag*self.complex_w[...,1] y ytorch.fft.irfft(torch.complex(real,imag),dim-1) y self.norm(y) att self.conv1d(y).reshape(y.size(0),self.kn,self.out_n,-1) return att.permute(0,1,3,2) class FrequencyBandModulation(nn.Module): def __init__(self,in_channels,k_list[2,4,8],lowfreq_attFalse,fs_featfeat,actsigmoid,spatialconv,spatial_group1,spatial_kernel3,initzero): super().__init__() self.k_list,self.lowfreq_att,self.act k_list,lowfreq_att,act # n_freq_groups used here self.spatial_group spatial_group self.convs nn.ModuleList() for _ in range(len(k_list)int(lowfreq_att)): convnn.Conv2d(in_channels, self.spatial_group, spatial_kernel, paddingspatial_kernel//2, groupsself.spatial_group) if initzero: nn.init.normal_(conv.weight, std1e-6); nn.init.zeros_(conv.bias) self.convs.append(conv) def forward(self,x): b,c,h,wx.shape x_ffttorch.fft.rfft2(x,normortho) out0 for i,f in enumerate(self.k_list): masktorch.zeros_like(x_fft[:,:1]) coords,_get_fft2freq(h,w,use_rfftTrue) mask[:,:,coords.max(-1)[0]0.5/f]1 lowtorch.fft.irfft2(x_fft*mask,s(h,w),normortho) highx-low; xlow w_attself.convs[i](x) w_att (w_att.sigmoid()*2) if self.actsigmoid else (1w_att.tanh()) out (w_att.reshape(b,self.spatial_group,-1,h,w)*high.reshape(b,self.spatial_group,-1,h,w)).reshape(b,-1,h,w) if self.lowfreq_att: w_attself.convs[-1](x) w_att(w_att.sigmoid()*2) if self.actsigmoid else (1w_att.tanh()) out (w_att.reshape(b,self.spatial_group,-1,h,w)*x.reshape(b,self.spatial_group,-1,h,w)).reshape(b,-1,h,w) else: outx return out class FDConv(nn.Conv2d): def __init__(self, in_channels, out_channels, kernel_size, stride1, paddingNone, dilation1, groups1, biasTrue, reduction0.0625, kernel_num4, n_freq_groups1, kernel_temp1.0, tempNone, att_multi2.0, param_ratio1, param_reduction1.0, ksm_only_kernel_attFalse, spatial_freq_decomposeFalse, use_ksm_localTrue, ksm_local_actsigmoid, ksm_global_actsigmoid, fbm_cfgNone): padding (kernel_size-1)//2 if padding is None else padding super().__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias) self.kernel_num kernel_num self.att_multi att_multi self.param_ratio param_ratio self.param_reduction param_reduction self.n_freq_groups n_freq_groups # 全域調製模組 if temp is None: temp kernel_temp self.KSMG KernelSpatialModulation_Global( in_planesin_channels, out_planesout_channels, kernel_sizekernel_size, reductionreduction, kernel_numkernel_num, temptemp, kernel_tempkernel_temp, att_multiatt_multi, ksm_only_kernel_attksm_only_kernel_att, spatial_freq_decomposespatial_freq_decompose, act_typeksm_global_act ) # 局部調製模組 self.use_local use_ksm_local if use_ksm_local: self.KSML KernelSpatialModulation_Local( channelin_channels, kernel_num1, out_nout_channels * kernel_size * kernel_size ) # 頻帶調製模組 if fbm_cfg is None: fbm_cfg {} fbm_cfg[spatial_group] n_freq_groups self.FBM FrequencyBandModulation(in_channels, **fbm_cfg) # 準備 DFT 權重 self._prepare_dft() def _prepare_dft(self): # 1) 計算 out×in 大矩陣的 DFT d1, d2 self.out_channels, self.in_channels k self.kernel_size[0] flat self.weight.permute(0,2,1,3).reshape(d1*k, d2*k) freq torch.fft.rfft2(flat, dim(0,1)) w torch.stack([freq.real, freq.imag], dim-1) # shape (d1*k, d2*k, 2) # 2) 將它做成可訓練參數 # shape → (param_ratio, d1*k, d2*k, 2) self.dft_weight nn.Parameter( w[None].repeat(self.param_ratio, 1, 1, 1) // (min(d1, d2)//2) ) # **不要再刪除 self.weight** # if self.bias is not None: # self.bias nn.Parameter(torch.zeros_like(self.bias)) def forward(self, x): # fallback to plain conv if小channel 或 非1/3kernel if min(self.in_channels, self.out_channels) 16 or self.kernel_size[0] not in [1,3]: # 這裡 super().forward 會呼叫 nn.Conv2d.forward必須有 self.weight return super().forward(x) b, _, h, w x.shape # 計算各種 attention map ch, fl, sp, kn self.KSMG(F.adaptive_avg_pool2d(x, 1)) hr 1 if self.use_local: local self.KSML(F.adaptive_avg_pool2d(x, 1)) hr (local.sigmoid() * self.att_multi) if self.KSML.use_global else (1 local.tanh()) # 建 DFT map 並還原到 spatial domain coords, _ get_fft2freq( self.out_channels * self.kernel_size[0], self.in_channels * self.kernel_size[1], use_rfftTrue ) DFTmap torch.zeros( (b, self.out_channels * self.kernel_size[0], self.in_channels * self.kernel_size[1]//21, 2), devicex.device ) for i in range(self.param_ratio): w self.dft_weight[i][coords[0], coords[1]][None] DFTmap[...,0] w[...,0] * kn[:, i] DFTmap[...,1] w[...,1] * kn[:, i] adapt torch.fft.irfft2(torch.view_as_complex(DFTmap), dim(1,2)) adapt adapt.reshape( b, 1, self.out_channels, self.kernel_size[0], self.in_channels, self.kernel_size[1] ).permute(0,1,2,4,3,5) # 頻帶模組 x_fbm self.FBM(x) x_in x_fbm if hasattr(self, FBM) else x # 聚合所有 attention agg sp * ch * fl * adapt * hr agg agg.sum(1).view( -1, self.in_channels // self.groups, self.kernel_size[0], self.kernel_size[1] ) # Depthwise batch conv trick xcat x_in.reshape(1, -1, h, w) out F.conv2d( xcat, agg, None, self.stride, self.padding, self.dilation, self.groups * b ) out out.view(b, self.out_channels, out.size(-2), out.size(-1)) if self.bias is not None: out self.bias.view(1, -1, 1, 1) return out class FDConv_cfg(nn.Module): Standard convolution Block with Batch Norm and Activation, using FDConv. default_act nn.SiLU() # Default activation def __init__(self, c1, c2, k1, s1, freq_groups1, pNone, g1, d1, actTrue): Args: c1 (int): input channels c2 (int): output channels k (int): kernel size s (int): stride p (int or None): padding; if None uses autopad g (int): groups d (int): dilation act (bool or nn.Module): activation to use (True→default_act, False→Identity, or custom) super().__init__() padding autopad(k, p, d) # replace nn.Conv2d with FDConv self.conv FDConv( in_channelsc1, out_channelsc2, kernel_sizek, strides, n_freq_groupsfreq_groups, paddingpadding, dilationd, groupsg, biasFalse ) self.bn nn.BatchNorm2d(c2) self.act ( self.default_act if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) ) def forward(self, x): Conv → BatchNorm → Activation x self.conv(x) x self.bn(x) return self.act(x) def forward_fuse(self, x): Conv only (for fused deployments) → Activation return self.act(self.conv(x)) if __name__ __main__: # Usage example: fd FDConv(in_channels64, out_channels64, kernel_size3, n_freq_groups64) print(fd)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_FDConv.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. modelyolo26n.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, FDConv_cfg, [64, 3, 2, 3]] # 0-P1/2 - [-1, 1, FDConv_cfg, [128, 3, 2, 16]] # 1-P2/4 - [-1, 2, C3k2, [256, False, 0.25]] - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8 - [-1, 2, C3k2, [512, False, 0.25]] - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16 - [-1, 2, C3k2, [512, True]] - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32 - [-1, 2, C3k2, [1024, True]] - [-1, 1, SPPF, [1024, 5]] # 9 - [-1, 2, C2PSA, [1024]] # 10 # YOLO26n head head: - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 6], 1, Concat, [1]] # cat backbone P4 - [-1, 2, C3k2, [512, False]] # 13 - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 4], 1, Concat, [1]] # cat backbone P3 - [-1, 2, C3k2, [256, False]] # 16 (P3/8-small) - [-1, 1, Conv, [256, 3, 2]] - [[-1, 13], 1, Concat, [1]] # cat head P4 - [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium) - [-1, 1, Conv, [512, 3, 2]] - [[-1, 10], 1, Concat, [1]] # cat head P5 - [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large) - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)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. modelyolo26n.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, FDConv_cfg, [64, 3, 2, 3]] # 0-P1/2 - [-1, 1, FDConv_cfg, [128, 3, 2, 16]] # 1-P2/4 - [-1, 2, C3k2, [256, False, 0.25]] - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8 - [-1, 2, C3k2, [512, False, 0.25]] - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16 - [-1, 2, C3k2, [512, True]] - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32 - [-1, 2, C3k2, [1024, True]] - [-1, 1, SPPF, [1024, 5]] # 9 - [-1, 2, C2PSA, [1024]] # 10 # YOLO26n head head: - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 6], 1, Concat, [1]] # cat backbone P4 - [-1, 2, C3k2, [512, False]] # 13 - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 4], 1, Concat, [1]] # cat backbone P3 - [-1, 2, C3k2, [256, False]] # 16 (P3/8-small) - [-1, 1, Conv, [256, 3, 2]] - [[-1, 13], 1, Concat, [1]] # cat head P4 - [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium) - [-1, 1, Conv, [512, 3, 2]] - [[-1, 10], 1, Concat, [1]] # cat head P5 - [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large) - [[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. modelyolo26n.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, FDConv_cfg, [64, 3, 2, 3]] # 0-P1/2 - [-1, 1, FDConv_cfg, [128, 3, 2, 16]] # 1-P2/4 - [-1, 2, C3k2, [256, False, 0.25]] - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8 - [-1, 2, C3k2, [512, False, 0.25]] - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16 - [-1, 2, C3k2, [512, True]] - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32 - [-1, 2, C3k2, [1024, True]] - [-1, 1, SPPF, [1024, 5]] # 9 - [-1, 2, C2PSA, [1024]] # 10 # YOLO26n head head: - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 6], 1, Concat, [1]] # cat backbone P4 - [-1, 2, C3k2, [512, False]] # 13 - [-1, 1, nn.Upsample, [None, 2, nearest]] - [[-1, 4], 1, Concat, [1]] # cat backbone P3 - [-1, 2, C3k2, [256, False]] # 16 (P3/8-small) - [-1, 1, Conv, [256, 3, 2]] - [[-1, 13], 1, Concat, [1]] # cat head P4 - [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium) - [-1, 1, Conv, [512, 3, 2]] - [[-1, 10], 1, Concat, [1]] # cat head P5 - [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large) - [[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_multipleend2end: True # whether to use end-to-end mode reg_max: 1 # DFL bins scales: # model compound scaling constants, i.e. modelyolo26n.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FDConv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FDConv2.5 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_FDConv.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(datar你的数据集的yaml文件地址, epochs100, batch16, imgsz640, workers4, namePath(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 2242 ultralytics.nn.models.FDConv.FDConv_cfg [3, 16, 3, 2, 3] 1 -1 1 12463 ultralytics.nn.models.FDConv.FDConv_cfg [16, 32, 3, 2, 16] 2 -1 1 6640 ultralytics.nn.modules.block.C3k2 [32, 64, 1, False, 0.25] 3 -1 1 36992 ultralytics.nn.modules.conv.Conv [64, 64, 3, 2] 4 -1 1 26080 ultralytics.nn.modules.block.C3k2 [64, 128, 1, False, 0.25] 5 -1 1 147712 ultralytics.nn.modules.conv.Conv [128, 128, 3, 2] 6 -1 1 87040 ultralytics.nn.modules.block.C3k2 [128, 128, 1, True] 7 -1 1 295424 ultralytics.nn.modules.conv.Conv [128, 256, 3, 2] 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] 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 111296 ultralytics.nn.modules.block.C3k2 [384, 128, 1, False] 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 32096 ultralytics.nn.modules.block.C3k2 [256, 64, 1, False] 17 -1 1 36992 ultralytics.nn.modules.conv.Conv [64, 64, 3, 2] 18 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1] 19 -1 1 86720 ultralytics.nn.modules.block.C3k2 [192, 128, 1, False] 20 -1 1 147712 ultralytics.nn.modules.conv.Conv [128, 128, 3, 2] 21 [-1, 10] 1 0 ultralytics.nn.modules.conv.Concat [1] 22 -1 1 378880 ultralytics.nn.modules.block.C3k2 [384, 256, 1, True] 23 [16, 19, 22] 1 309656 ultralytics.nn.modules.head.Detect [80, 1, True, [64, 128, 256]] YOLO26_FDConv summary: 249 layers, 2,478,393 parameters, 2,478,393 gradients, 5.6 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