Color Filter Array Interpolation（CFA插值） 是图像信号处理（ISP）中的核心步骤之一。它的目标是：
将原始 Bayer 图像（只有每个像素一个颜色分量）还原成完整的 RGB 图像，即为每个像素补全缺失的两个颜色通道 —— 这个过程称为 Demosaicing。
什么是 Color Filter Array（CFA）？
传感器每个像素只能采集一个颜色通道（R、G、B）；
为了同时获取三种颜色信息，我们使用了 Bayer Pattern（最常见的一种 CFA）：
Bayer图的像素排序：
R G R G R G...
G B G B...
R G R G...
G B G B...
...Bayer 图中的每个像素只有 R、G 或 B 中的一个值，其余两个颜色是空的。
插值的目标
对每个像素位置，估算出缺失的两个颜色值，使其成为完整的 (R, G, B) 三通道像素。
插值方法种类
1. 最近邻插值（Nearest Neighbor）
简单快速：从邻近像素直接复制缺失值；
缺点：容易出现伪影或马赛克纹理。
2. 双线性插值（Bilinear Interpolation）
使用上下左右的加权平均估算；
比最近邻更平滑，但边缘保留能力差。
3. 双立方插值、方向感知插值（Edge-Aware）
更复杂，考虑图像边缘信息；
可以减少色彩混叠和锯齿，质量更高。
代码实现：
def CFA(awb_img):
"""
inputs:
awb_img = bayer domain image after auto white balance gain control
outputs:
cfa_img = 8 bit RGB image
"""
awb_img = awb_img.astype(np.uint32) # change dtype to allow for larger numbers during demosaicing
max_val = np.max(awb_img)
r = np.empty(awb_img.shape) # create empty arrays for R, G, and B channels
g = np.empty(awb_img.shape)
b = np.empty(awb_img.shape)
padded_img = np.pad(awb_img, (1,1), 'reflect') # pad image to extract shifted channels
# R-channel
r_on_r = padded_img[1:-2:2, 1:-2:2] # red intensity values on red pixels
r_right = padded_img[1:-2:2, 3::2] # shifted matrix of red intensity values in specified direction, same size as r_on_r
r_down = padded_img[3::2, 1:-2:2]
r_down_right = padded_img[3::2, 3::2]
r_on_gr = np.right_shift((r_on_r + r_right), 1) # calculate red intensity values from green and blue pixels
r_on_gb = np.right_shift((r_on_r + r_down), 1)
r_on_b = np.right_shift((r_on_r + r_right + r_down + r_down_right), 2)
r[::2, ::2] = r_on_r # map red intensity values to the entire red channel array
r[::2, 1::2] = r_on_gr
r[1::2, ::2] = r_on_gb
r[1::2, 1::2] = r_on_b
#G-channel
g_on_gr = padded_img[1:-2:2, 2:-1:2] # green intensity values on green pixels
g_on_gb = padded_img[2:-1:2, 1:-2:2]
g_up = padded_img[:-3:2, 1:-2:2] # shifted matrix of green intensity values in specified direction, same size as g_on_gr
g_right = padded_img[2:-1:2, 3::2]
g_left = padded_img[1:-2:2, :-3:2]
g_down = padded_img[3::2, 2:-1:2]
g_on_r = np.right_shift((g_left + g_up + g_on_gr + g_on_gb), 2) # calculate green intensity values from red and blue pixels
g_on_b = np.right_shift((g_right + g_down + g_on_gr + g_on_gb), 2)
g[::2, ::2] = g_on_r # map green intensity values to the entire green channel array
g[::2, 1::2] = g_on_gr
g[1::2, ::2] = g_on_gb
g[1::2, 1::2] = g_on_b
# B-channel
b_on_b = padded_img[2:-1:2, 2:-1:2] # blue intensity values on green pixels
b_left = padded_img[2:-1:2, :-3:2] # shifted matrix of blue intensity values in specified direction, same size as b_on_b
b_up = padded_img[:-3:2, 2:-1:2]
b_up_left = padded_img[:-3:2, :-3:2]
b_on_gr = np.right_shift((b_on_b + b_up), 1) # calculate blue intensity values from red and green pixels
b_on_gb = np.right_shift((b_on_b + b_left), 1)
b_on_r = np.right_shift((b_on_b + b_left + b_up + b_up_left), 2)
b[::2, ::2] = b_on_r # map blue intensity values to the entire blue channel array
b[::2, 1::2] = b_on_gr
b[1::2, ::2] = b_on_gb
b[1::2, 1::2] = b_on_b
r = np.clip(r, 0, max_val) # ensure all values are of the correct and same dtype before stacking
g = np.clip(g, 0, max_val)
b = np.clip(b, 0, max_val)
cfa_img = np.dstack((r, g, b)) # Resize and stack the channels to RGB
cfa_img = ((cfa_img / max_val) * 255).astype(np.uint8) # rescale the image to 8 bit to skip color correction matrix
return cfa_img| 位置 | 插值策略 |
|---|---|
| R | 使用右、下、右下邻域平均 |
| G on R/B | 使用上下左右平均 |
| B | 使用上、左、左上邻域平均 |