图像处理中的双边滤波(bilateral filter)数据增强：原理剖析与实战优化

1次阅读

共计 2634 个字符，预计需要花费 7 分钟才能阅读完成。

在计算机视觉任务中，数据增强是提升模型泛化能力的关键手段。然而传统方法如高斯模糊在注入噪声时，往往会破坏图像的边缘特征，导致关键信息丢失。本文将深入探讨双边滤波在数据增强中的应用，它通过独特的权重控制机制，实现在平滑噪声的同时保留边缘信息。

双边滤波的核心思想是同时考虑空间距离和像素值差异，其数学表达式为：

$$I^{filtered}(x) = \frac{1}{W_p} \sum_{x_i \in \Omega} I(x_i)f_r(||I(x_i)-I(x)||)g_s(||x_i-x||)$$

其中：

$f_r$ 是值域核，控制像素值相似性的权重
$g_s$ 是空间核，控制空间距离的权重
$W_p$ 是归一化因子

均值滤波：均匀平滑，边缘完全模糊
高斯滤波：基于距离加权平滑，但边缘仍然模糊
双边滤波：在平滑区域效果类似高斯滤波，但在边缘处保持锐利

$\sigma_d$（空间核标准差）：
控制平滑的空间范围
经验值：通常设为图像尺寸的 1 -2%
$\sigma_r$（值域核标准差）：
控制像素值差异的敏感度
经验值：对于 8 位图像，建议 10-75 之间

import cv2
import numpy as np

def bilateral_augmentation(img, d=9, sigma_color=75, sigma_space=75):
    """
    Apply bilateral filter for data augmentation
    :param d: Diameter of pixel neighborhood
    :param sigma_color: Filter sigma in the color space
    :param sigma_space: Filter sigma in the coordinate space
    """
    # Automatically adjust parameters based on image size
    height, width = img.shape[:2]
    sigma_space = max(5, min(height, width) * 0.01)

    # Apply bilateral filter
    augmented = cv2.bilateralFilter(img, d, sigma_color, sigma_space)

    # Blend with original image
    alpha = 0.7  # Strength of augmentation
    result = cv2.addWeighted(img, 1-alpha, augmented, alpha, 0)

    return result

import torch
import torch.nn.functional as F

def gaussian_kernel(size, sigma):
    """Create 2D Gaussian kernel"""
    coords = torch.arange(size).float() - size//2
    g = torch.exp(-(coords**2)/(2*sigma**2))
    g /= g.sum()
    return g.outer(g)

def bilateral_filter_pytorch(image, spatial_sigma=5, range_sigma=0.1):
    """
    PyTorch implementation of bilateral filter
    :param image: Input tensor (C,H,W)
    :param spatial_sigma: Sigma for spatial kernel
    :param range_sigma: Sigma for range kernel
    """
    C, H, W = image.shape
    kernel_size = int(6*spatial_sigma + 1)

    # Create spatial kernel
    spatial_kernel = gaussian_kernel(kernel_size, spatial_sigma)
    spatial_kernel = spatial_kernel.view(1, 1, kernel_size, kernel_size).to(image.device)

    # Pad image for convolution
    pad = kernel_size // 2
    padded = F.pad(image.unsqueeze(0), (pad, pad, pad, pad), mode='reflect')

    # Compute range weights
    unfolded = F.unfold(padded, kernel_size, dilation=1, padding=0)
    diff = unfolded - image.view(1, C, H*W, 1)
    range_weights = torch.exp(-(diff**2)/(2*range_sigma**2))

    # Combine weights and normalize
    weights = spatial_kernel * range_weights.prod(dim=1, keepdim=True)
    weights = weights / weights.sum(dim=3, keepdim=True).sum(dim=2, keepdim=True)

    # Apply filter
    filtered = (unfolded * weights).sum(dim=2).view_as(image)

    return filtered

我们设计了以下实验来验证双边滤波数据增强的效果：