论文信息
标题: Frequency-aware Feature Fusion for Dense Image Prediction
作者: Linwei Chen, Ying Fu, Lin Gu, Chenggang Yan, Tatsuya Harada, Gao Huang
论文链接：https://arxiv.org/pdf/2408.12879
GitHub链接：https://github.com/Linwei-Chen/FreqFusion
创新点
本论文提出了一种新的特征融合方法，称为频率感知特征融合（FreqFusion），旨在解决现有密集图像预测模型中存在的两个主要问题：
- 类别内不一致性：由于高频特征的干扰，导致同一类别内的特征值快速变化。
- 边界模糊：融合特征的边界缺乏准确的高频信息，造成边界位移。
方法
FreqFusion方法集成了三个关键组件：
-
自适应低通滤波器（ALPF）生成器：用于预测空间变化的低通滤波器，旨在减少上采样过程中对象内部的高频成分，从而降低类别内的不一致性。
-
偏移生成器：通过重采样，将不一致的特征替换为更一致的特征，以提高同一类别目标特征的一致性。
-
自适应高通滤波器（AHPF）生成器：用于增强在下采样过程中丢失的高频细节边界信息，从而提高边界的清晰度。
这三个组件协同工作，旨在恢复具有一致类别信息和明确边界的融合特征。
效果
通过综合可视化和定量分析，FreqFusion显著提高了特征的一致性，并锐化了对象的边界。该方法在多个密集预测任务中表现出色，能够有效改善现有模型的性能。
实验结果
在各种密集图像预测任务中，FreqFusion的实验结果如下：
-
语义分割：在轻量化语义分割模型SegNeXt上，提升了2.4 mIoU；在强大的Mask2Former上，提升了1.4 mIoU。
-
目标检测：在Faster RCNN上，提升了1.9 AP。
-
实例分割：在Mask R-CNN上，提升了1.7 box AP和1.3 mask AP。
-
全景分割：在PanopticFCN上，提升了2.5 PQ。
这些结果表明，FreqFusion在提高特征质量和模型性能方面具有显著的优势，尤其是在处理复杂的图像密集预测任务时。
代码
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
import warnings
import numpy as npdef xavier_init(module: nn.Module,gain: float = 1,bias: float = 0,distribution: str = 'normal') -> None:assert distribution in ['uniform', 'normal']if hasattr(module, 'weight') and module.weight is not None:if distribution == 'uniform':nn.init.xavier_uniform_(module.weight, gain=gain)else:nn.init.xavier_normal_(module.weight, gain=gain)if hasattr(module, 'bias') and module.bias is not None:nn.init.constant_(module.bias, bias)def carafe(x, normed_mask, kernel_size, group=1, up=1):b, c, h, w = x.shape_, m_c, m_h, m_w = normed_mask.shapeprint('x', x.shape)print('normed_mask', normed_mask.shape)# assert m_c == kernel_size ** 2 * up ** 2assert m_h == up * hassert m_w == up * wpad = kernel_size // 2# print(pad)pad_x = F.pad(x, pad=[pad] * 4, mode='reflect')# print(pad_x.shape)unfold_x = F.unfold(pad_x, kernel_size=(kernel_size, kernel_size), stride=1, padding=0)# unfold_x = unfold_x.reshape(b, c, 1, kernel_size, kernel_size, h, w).repeat(1, 1, up ** 2, 1, 1, 1, 1)unfold_x = unfold_x.reshape(b, c * kernel_size * kernel_size, h, w)unfold_x = F.interpolate(unfold_x, scale_factor=up, mode='nearest')# normed_mask = normed_mask.reshape(b, 1, up ** 2, kernel_size, kernel_size, h, w)unfold_x = unfold_x.reshape(b, c, kernel_size * kernel_size, m_h, m_w)normed_mask = normed_mask.reshape(b, 1, kernel_size * kernel_size, m_h, m_w)res = unfold_x * normed_mask# test# res[:, :, 0] = 1# res[:, :, 1] = 2# res[:, :, 2] = 3# res[:, :, 3] = 4res = res.sum(dim=2).reshape(b, c, m_h, m_w)# res = F.pixel_shuffle(res, up)# print(res.shape)# print(res)return resdef normal_init(module, mean=0, std=1, bias=0):if hasattr(module, 'weight') and module.weight is not None:nn.init.normal_(module.weight, mean, std)if hasattr(module, 'bias') and module.bias is not None:nn.init.constant_(module.bias, bias)def constant_init(module, val, bias=0):if hasattr(module, 'weight') and module.weight is not None:nn.init.constant_(module.weight, val)if hasattr(module, 'bias') and module.bias is not None:nn.init.constant_(module.bias, bias)def resize(input,size=None,scale_factor=None,mode='nearest',align_corners=None,warning=True):if warning:if size is not None and align_corners:input_h, input_w = tuple(int(x) for x in input.shape[2:])output_h, output_w = tuple(int(x) for x in size)if output_h > input_h or output_w > input_w:if ((output_h > 1 and output_w > 1 and input_h > 1and input_w > 1) and (output_h - 1) % (input_h - 1)and (output_w - 1) % (input_w - 1)):warnings.warn(f'When align_corners={align_corners}, ''the output would more aligned if 'f'input size {(input_h, input_w)} is `x+1` and 'f'out size {(output_h, output_w)} is `nx+1`')return F.interpolate(input, size, scale_factor, mode, align_corners)def hamming2D(M, N):"""生成二维Hamming窗参数：- M：窗口的行数- N：窗口的列数返回：- 二维Hamming窗"""hamming_x = np.hamming(M)hamming_y = np.hamming(N)# 通过外积生成二维Hamming窗hamming_2d = np.outer(hamming_x, hamming_y)return hamming_2dclass FreqFusion(nn.Module):def __init__(self,hr_channels,lr_channels,scale_factor=1,lowpass_kernel=5,highpass_kernel=3,up_group=1,encoder_kernel=3,encoder_dilation=1,compressed_channels=64,align_corners=False,upsample_mode='nearest',feature_resample=False, # use offset generator or notfeature_resample_group=4,comp_feat_upsample=True, # use ALPF & AHPF for init upsamplinguse_high_pass=True,use_low_pass=True,hr_residual=True,semi_conv=True,hamming_window=True, # for regularization, do not matter reallyfeature_resample_norm=True,**kwargs):super().__init__()self.scale_factor = scale_factorself.lowpass_kernel = lowpass_kernelself.highpass_kernel = highpass_kernelself.up_group = up_groupself.encoder_kernel = encoder_kernelself.encoder_dilation = encoder_dilationself.compressed_channels = compressed_channelsself.hr_channel_compressor = nn.Conv2d(hr_channels, self.compressed_channels, 1)self.lr_channel_compressor = nn.Conv2d(lr_channels, self.compressed_channels, 1)self.content_encoder = nn.Conv2d( # ALPF generatorself.compressed_channels,lowpass_kernel ** 2 * self.up_group * self.scale_factor * self.scale_factor,self.encoder_kernel,padding=int((self.encoder_kernel - 1) * self.encoder_dilation / 2),dilation=self.encoder_dilation,groups=1)self.align_corners = align_cornersself.upsample_mode = upsample_modeself.hr_residual = hr_residualself.use_high_pass = use_high_passself.use_low_pass = use_low_passself.semi_conv = semi_convself.feature_resample = feature_resampleself.comp_feat_upsample = comp_feat_upsampleif self.feature_resample:self.dysampler = LocalSimGuidedSampler(in_channels=compressed_channels, scale=2, style='lp',groups=feature_resample_group, use_direct_scale=True,kernel_size=encoder_kernel, norm=feature_resample_norm)if self.use_high_pass:self.content_encoder2 = nn.Conv2d( # AHPF generatorself.compressed_channels,highpass_kernel ** 2 * self.up_group * self.scale_factor * self.scale_factor,self.encoder_kernel,padding=int((self.encoder_kernel - 1) * self.encoder_dilation / 2),dilation=self.encoder_dilation,groups=1)self.hamming_window = hamming_windowlowpass_pad = 0highpass_pad = 0if self.hamming_window:self.register_buffer('hamming_lowpass', torch.FloatTensor(hamming2D(lowpass_kernel + 2 * lowpass_pad, lowpass_kernel + 2 * lowpass_pad))[None, None,])self.register_buffer('hamming_highpass', torch.FloatTensor(hamming2D(highpass_kernel + 2 * highpass_pad, highpass_kernel + 2 * highpass_pad))[None, None,])else:self.register_buffer('hamming_lowpass', torch.FloatTensor([1.0]))self.register_buffer('hamming_highpass', torch.FloatTensor([1.0]))self.init_weights()def init_weights(self):for m in self.modules():# print(m)if isinstance(m, nn.Conv2d):xavier_init(m, distribution='uniform')normal_init(self.content_encoder, std=0.001)if self.use_high_pass:normal_init(self.content_encoder2, std=0.001)def kernel_normalizer(self, mask, kernel, scale_factor=None, hamming=1):if scale_factor is not None:mask = F.pixel_shuffle(mask, self.scale_factor)n, mask_c, h, w = mask.size()mask_channel = int(mask_c / float(kernel ** 2)) # group# mask = mask.view(n, mask_channel, -1, h, w)# mask = F.softmax(mask, dim=2, dtype=mask.dtype)# mask = mask.view(n, mask_c, h, w).contiguous()mask = mask.view(n, mask_channel, -1, h, w)mask = F.softmax(mask, dim=2, dtype=mask.dtype)mask = mask.view(n, mask_channel, kernel, kernel, h, w)mask = mask.permute(0, 1, 4, 5, 2, 3).view(n, -1, kernel, kernel)# mask = F.pad(mask, pad=[padding] * 4, mode=self.padding_mode) # kernel + 2 * paddingmask = mask * hammingmask /= mask.sum(dim=(-1, -2), keepdims=True)# print(hamming)# print(mask.shape)mask = mask.view(n, mask_channel, h, w, -1)mask = mask.permute(0, 1, 4, 2, 3).view(n, -1, h, w).contiguous()return maskdef forward(self, hr_feat, lr_feat, use_checkpoint=False): # use check_point to save GPU memoryif use_checkpoint:return checkpoint(self._forward, hr_feat, lr_feat)else:return self._forward(hr_feat, lr_feat)def _forward(self, hr_feat, lr_feat):compressed_hr_feat = self.hr_channel_compressor(hr_feat)compressed_lr_feat = self.lr_channel_compressor(lr_feat)if self.semi_conv:if self.comp_feat_upsample:if self.use_high_pass:mask_hr_hr_feat = self.content_encoder2(compressed_hr_feat) # 从hr_feat得到初始高通滤波特征mask_hr_init = self.kernel_normalizer(mask_hr_hr_feat, self.highpass_kernel,hamming=self.hamming_highpass) # kernel归一化得到初始高通滤波compressed_hr_feat = compressed_hr_feat + compressed_hr_feat - carafe(compressed_hr_feat,mask_hr_init,self.highpass_kernel,self.up_group,1) # 利用初始高通滤波对压缩hr_feat的高频增强 （x-x的低通结果=x的高通结果）mask_lr_hr_feat = self.content_encoder(compressed_hr_feat) # 从hr_feat得到初始低通滤波特征mask_lr_init = self.kernel_normalizer(mask_lr_hr_feat, self.lowpass_kernel,hamming=self.hamming_lowpass) # kernel归一化得到初始低通滤波mask_lr_lr_feat_lr = self.content_encoder(compressed_lr_feat) # 从hr_feat得到另一部分初始低通滤波特征mask_lr_lr_feat = F.interpolate( # 利用初始低通滤波对另一部分初始低通滤波特征上采样carafe(mask_lr_lr_feat_lr, mask_lr_init, self.lowpass_kernel, self.up_group, 2),size=compressed_hr_feat.shape[-2:], mode='nearest')mask_lr = mask_lr_hr_feat + mask_lr_lr_feat # 将两部分初始低通滤波特征合在一起mask_lr_init = self.kernel_normalizer(mask_lr, self.lowpass_kernel,hamming=self.hamming_lowpass) # 得到初步融合的初始低通滤波mask_hr_lr_feat = F.interpolate( # 使用初始低通滤波对lr_feat处理，分辨率得到提高carafe(self.content_encoder2(compressed_lr_feat), mask_lr_init, self.lowpass_kernel,self.up_group, 2), size=compressed_hr_feat.shape[-2:], mode='nearest')mask_hr = mask_hr_hr_feat + mask_hr_lr_feat # 最终高通滤波特征else:raise NotImplementedErrorelse:mask_lr = self.content_encoder(compressed_hr_feat) + F.interpolate(self.content_encoder(compressed_lr_feat), size=compressed_hr_feat.shape[-2:], mode='nearest')if self.use_high_pass:mask_hr = self.content_encoder2(compressed_hr_feat) + F.interpolate(self.content_encoder2(compressed_lr_feat), size=compressed_hr_feat.shape[-2:], mode='nearest')else:compressed_x = F.interpolate(compressed_lr_feat, size=compressed_hr_feat.shape[-2:],mode='nearest') + compressed_hr_featmask_lr = self.content_encoder(compressed_x)if self.use_high_pass:mask_hr = self.content_encoder2(compressed_x)mask_lr = self.kernel_normalizer(mask_lr, self.lowpass_kernel, hamming=self.hamming_lowpass)if self.semi_conv:lr_feat = carafe(lr_feat, mask_lr, self.lowpass_kernel, self.up_group, 2)else:lr_feat = resize(input=lr_feat,size=hr_feat.shape[2:],mode=self.upsample_mode,align_corners=None if self.upsample_mode == 'nearest' else self.align_corners)lr_feat = carafe(lr_feat, mask_lr, self.lowpass_kernel, self.up_group, 1)if self.use_high_pass:mask_hr = self.kernel_normalizer(mask_hr, self.highpass_kernel, hamming=self.hamming_highpass)hr_feat_hf = hr_feat - carafe(hr_feat, mask_hr, self.highpass_kernel, self.up_group, 1)if self.hr_residual:# print('using hr_residual')hr_feat = hr_feat_hf + hr_featelse:hr_feat = hr_feat_hfif self.feature_resample:# print(lr_feat.shape)lr_feat = self.dysampler(hr_x=compressed_hr_feat,lr_x=compressed_lr_feat, feat2sample=lr_feat)return mask_lr, hr_feat, lr_featclass LocalSimGuidedSampler(nn.Module):"""offset generator in FreqFusion"""def __init__(self, in_channels, scale=2, style='lp', groups=4, use_direct_scale=True, kernel_size=1, local_window=3,sim_type='cos', norm=True, direction_feat='sim_concat'):super().__init__()assert scale == 2assert style == 'lp'self.scale = scaleself.style = styleself.groups = groupsself.local_window = local_windowself.sim_type = sim_typeself.direction_feat = direction_featif style == 'pl':assert in_channels >= scale ** 2 and in_channels % scale ** 2 == 0assert in_channels >= groups and in_channels % groups == 0if style == 'pl':in_channels = in_channels // scale ** 2out_channels = 2 * groupselse:out_channels = 2 * groups * scale ** 2if self.direction_feat == 'sim':self.offset = nn.Conv2d(local_window ** 2 - 1, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)elif self.direction_feat == 'sim_concat':self.offset = nn.Conv2d(in_channels + local_window ** 2 - 1, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)else:raise NotImplementedErrornormal_init(self.offset, std=0.001)if use_direct_scale:if self.direction_feat == 'sim':self.direct_scale = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)elif self.direction_feat == 'sim_concat':self.direct_scale = nn.Conv2d(in_channels + local_window ** 2 - 1, out_channels,kernel_size=kernel_size, padding=kernel_size // 2)else:raise NotImplementedErrorconstant_init(self.direct_scale, val=0.)out_channels = 2 * groupsif self.direction_feat == 'sim':self.hr_offset = nn.Conv2d(local_window ** 2 - 1, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)elif self.direction_feat == 'sim_concat':self.hr_offset = nn.Conv2d(in_channels + local_window ** 2 - 1, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)else:raise NotImplementedErrornormal_init(self.hr_offset, std=0.001)if use_direct_scale:if self.direction_feat == 'sim':self.hr_direct_scale = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size,padding=kernel_size // 2)elif self.direction_feat == 'sim_concat':self.hr_direct_scale = nn.Conv2d(in_channels + local_window ** 2 - 1, out_channels,kernel_size=kernel_size, padding=kernel_size // 2)else:raise NotImplementedErrorconstant_init(self.hr_direct_scale, val=0.)self.norm = normif self.norm:self.norm_hr = nn.GroupNorm(in_channels // 8, in_channels)self.norm_lr = nn.GroupNorm(in_channels // 8, in_channels)else:self.norm_hr = nn.Identity()self.norm_lr = nn.Identity()self.register_buffer('init_pos', self._init_pos())def _init_pos(self):h = torch.arange((-self.scale + 1) / 2, (self.scale - 1) / 2 + 1) / self.scalereturn torch.stack(torch.meshgrid([h, h])).transpose(1, 2).repeat(1, self.groups, 1).reshape(1, -1, 1, 1)def sample(self, x, offset, scale=None):if scale is None: scale = self.scaleB, _, H, W = offset.shapeoffset = offset.view(B, 2, -1, H, W)coords_h = torch.arange(H) + 0.5coords_w = torch.arange(W) + 0.5coords = torch.stack(torch.meshgrid([coords_w, coords_h])).transpose(1, 2).unsqueeze(1).unsqueeze(0).type(x.dtype).to(x.device)normalizer = torch.tensor([W, H], dtype=x.dtype, device=x.device).view(1, 2, 1, 1, 1)coords = 2 * (coords + offset) / normalizer - 1coords = F.pixel_shuffle(coords.view(B, -1, H, W), scale).view(B, 2, -1, scale * H, scale * W).permute(0, 2, 3, 4, 1).contiguous().flatten(0, 1)return F.grid_sample(x.reshape(B * self.groups, -1, x.size(-2), x.size(-1)), coords, mode='bilinear',align_corners=False, padding_mode="border").view(B, -1, scale * H, scale * W)def forward(self, hr_x, lr_x, feat2sample):hr_x = self.norm_hr(hr_x)lr_x = self.norm_lr(lr_x)if self.direction_feat == 'sim':hr_sim = compute_similarity(hr_x, self.local_window, dilation=2, sim='cos')lr_sim = compute_similarity(lr_x, self.local_window, dilation=2, sim='cos')elif self.direction_feat == 'sim_concat':hr_sim = torch.cat([hr_x, compute_similarity(hr_x, self.local_window, dilation=2, sim='cos')], dim=1)lr_sim = torch.cat([lr_x, compute_similarity(lr_x, self.local_window, dilation=2, sim='cos')], dim=1)hr_x, lr_x = hr_sim, lr_sim# offset = self.get_offset(hr_x, lr_x)offset = self.get_offset_lp(hr_x, lr_x, hr_sim, lr_sim)return self.sample(feat2sample, offset)# def get_offset_lp(self, hr_x, lr_x):def get_offset_lp(self, hr_x, lr_x, hr_sim, lr_sim):if hasattr(self, 'direct_scale'):# offset = (self.offset(lr_x) + F.pixel_unshuffle(self.hr_offset(hr_x), self.scale)) * (self.direct_scale(lr_x) + F.pixel_unshuffle(self.hr_direct_scale(hr_x), self.scale)).sigmoid() + self.init_posoffset = (self.offset(lr_sim) + F.pixel_unshuffle(self.hr_offset(hr_sim), self.scale)) * (self.direct_scale(lr_x) + F.pixel_unshuffle(self.hr_direct_scale(hr_x),self.scale)).sigmoid() + self.init_pos# offset = (self.offset(lr_sim) + F.pixel_unshuffle(self.hr_offset(hr_sim), self.scale)) * (self.direct_scale(lr_sim) + F.pixel_unshuffle(self.hr_direct_scale(hr_sim), self.scale)).sigmoid() + self.init_poselse:offset = (self.offset(lr_x) + F.pixel_unshuffle(self.hr_offset(hr_x), self.scale)) * 0.25 + self.init_posreturn offsetdef get_offset(self, hr_x, lr_x):if self.style == 'pl':raise NotImplementedErrorreturn self.get_offset_lp(hr_x, lr_x)def compute_similarity(input_tensor, k=3, dilation=1, sim='cos'):"""计算输入张量中每一点与周围KxK范围内的点的余弦相似度。参数：- input_tensor: 输入张量，形状为[B, C, H, W]- k: 范围大小，表示周围KxK范围内的点返回：- 输出张量，形状为[B, KxK-1, H, W]"""B, C, H, W = input_tensor.shape# 使用零填充来处理边界情况# padded_input = F.pad(input_tensor, (k // 2, k // 2, k // 2, k // 2), mode='constant', value=0)# 展平输入张量中每个点及其周围KxK范围内的点unfold_tensor = F.unfold(input_tensor, k, padding=(k // 2) * dilation, dilation=dilation) # B, CxKxK, HW# print(unfold_tensor.shape)unfold_tensor = unfold_tensor.reshape(B, C, k ** 2, H, W)# 计算余弦相似度if sim == 'cos':similarity = F.cosine_similarity(unfold_tensor[:, :, k * k // 2:k * k // 2 + 1], unfold_tensor[:, :, :], dim=1)elif sim == 'dot':similarity = unfold_tensor[:, :, k * k // 2:k * k // 2 + 1] * unfold_tensor[:, :, :]similarity = similarity.sum(dim=1)else:raise NotImplementedError# 移除中心点的余弦相似度，得到[KxK-1]的结果similarity = torch.cat((similarity[:, :k * k // 2], similarity[:, k * k // 2 + 1:]), dim=1)# 将结果重塑回[B, KxK-1, H, W]的形状similarity = similarity.view(B, k * k - 1, H, W)return similarityif __name__ == '__main__':hr_feat = torch.rand(1, 128, 512, 512)lr_feat = torch.rand(1, 128, 256, 256)model = FreqFusion(hr_channels=128, lr_channels=128)mask_lr, hr_feat, lr_feat = model(hr_feat=hr_feat, lr_feat=lr_feat)print(mask_lr.shape)
