我目前正在R中实现filter2
MATLAB函数,这是一种2D卷积方法。我已经为2D卷积工作做了准备,但是有效的' filter2中的选项对我来说不太清楚。
这里描述MATLAB函数: http://se.mathworks.com/help/matlab/ref/filter2.html
我的实施:
filter2D <- function(img, window) {
# Algoritm for 2D Convolution
filter_center_index_y <- median(1:dim(window)[1])
filter_max_index_y <- dim(window)[1]
filter_center_index_x <- median(1:dim(window)[2])
filter_max_index_x <- dim(window)[2]
# For each position in the picture, 2D convolution is done by
# calculating a score for all overlapping values within the two matrices
x_min <- 1
x_max <- dim(img)[2]
y_min <- 1
y_max <- dim(img)[1]
df <- NULL
for (x_val in c(x_min:x_max)){
for (y_val in c(y_min:y_max)){
# Distanced from cell
img_dist_left <- x_val-1
img_dist_right <- x_max-x_val
img_dist_up <- y_val-1
img_dist_down <- y_max-y_val
# Overlapping filter cells
filter_x_start <- filter_center_index_x-img_dist_left
if (filter_x_start < 1) {
filter_x_start <- 1
}
filter_x_end <- filter_center_index_x+img_dist_right
if (filter_x_end > filter_max_index_x) {
filter_x_end <- filter_max_index_x
}
filter_y_start <- filter_center_index_y-img_dist_up
if (filter_y_start < 1) {
filter_y_start <- 1
}
filter_y_end <- filter_center_index_y+img_dist_down
if (filter_y_end > filter_max_index_y) {
filter_y_end <- filter_max_index_y
}
# Part of filter that overlaps
filter_overlap_matrix <- filter[filter_y_start:filter_y_end, filter_x_start:filter_x_end]
# Overlapped image cells
image_x_start <- x_val-filter_center_index_x+1
if (image_x_start < 1) {
image_x_start <- 1
}
image_x_end <- x_val+filter_max_index_x-filter_center_index_x
if (image_x_end > x_max) {
image_x_end <- x_max
}
image_y_start <- y_val-filter_center_index_y+1
if (image_y_start < 1) {
image_y_start <- 1
}
image_y_end <- y_val+filter_max_index_y-filter_center_index_y
if (image_y_end > y_max) {
image_y_end <- y_max
}
# Part of image that is overlapped
image_overlap_matrix <- img[image_y_start:image_y_end, image_x_start:image_x_end]
# Calculating the cell value
cell_value <- sum(filter_overlap_matrix*image_overlap_matrix)
df = rbind(df,data.frame(x_val,y_val, cell_value))
}
}
# Axis labels
x_axis <- c(x_min:x_max)
y_axis <- c(y_min:y_max)
# Populating matrix
filter_matrix <- matrix(df[,3], nrow = x_max, ncol = y_max, dimnames = list(x_axis, y_axis))
return(filter_matrix)
}
运行方法:
> image
[,1] [,2] [,3] [,4] [,5] [,6]
[1,] 1 2 3 4 5 6
[2,] 7 8 9 10 11 12
[3,] 13 14 15 16 17 18
[4,] 19 20 21 22 23 24
[5,] 25 26 27 28 29 30
[6,] 31 32 33 34 35 36
> filter
[,1] [,2] [,3]
[1,] 1 2 1
[2,] 0 0 0
[3,] -1 -2 -1
> filter2D(image, filter)
1 2 3 4 5 6
1 -22 -32 -36 -40 -44 -35
2 -36 -48 -48 -48 -48 -36
3 -36 -48 -48 -48 -48 -36
4 -36 -48 -48 -48 -48 -36
5 -36 -48 -48 -48 -48 -36
6 76 104 108 112 116 89
这与filter2(图像,过滤器)在matlab中生成的输出相同,但是,当选项“有效”时,添加以下输出:
-48 -48 -48 -48
-48 -48 -48 -48
-48 -48 -48 -48
-48 -48 -48 -48
如果filter2具有&#39;有效&#39;则不完全明显。选项生成此。它只是使用中心值吗?或者它正在做一些更复杂的事情?
答案 0 :(得分:2)
在开始之前,您的代码实际上正在执行2D 关联。 2D卷积要求在进行加权求和之前在内核上执行180度旋转。如果内核是对称的(即内核的转置等于它自己),则相关和卷积实际上是相同的操作。我只是想在开始之前说清楚。此外,filter2
的文档确实说明正在执行关联。
MATLAB中的'valid'
选项仅意味着它应该仅返回内核在执行过滤时与2D信号完全重叠的输出。因为你有一个3 x 3内核,这意味着在2D信号中的位置(2,2)
,例如,内核不会超出信号边界。因此,返回的是滤波的2D信号的范围,其中内核完全在原始2D信号内。举个例子,如果你把内核放在位置(1,1)
,那么某些内核会超出范围。在过滤时处理越界条件可以通过多种方式完成,这可能会影响结果并在结果归结时解释这些结果。因此,当您使用形成最终结果的真实信息时,需要'valid'
选项。您没有对超出2D信号边界的数据进行插值或执行任何估计。
简单地说,你返回一个删除边框元素的简化矩阵。过滤器形状奇特,这很容易。您只需删除第一个和最后一个floor(M/2)
行以及floor(N/2)
列,其中M x N
是内核的大小。因此,因为你的内核是3 x 3,这意味着我们需要从顶部删除1行,从底部删除1行,从左边删除1列,从右边删除1列。从MATLAB的输出中可以看出,-48
在4 x 4网格内。
因此,如果您想在代码中使用'valid
'选项,只需确保删除结果中的边框元素即可。您可以在返回矩阵之前完成此操作:
# Place your code here...
# ...
# ...
# Now we're at the end of your code
# Populating matrix
filter_matrix <- matrix(df[,3], nrow = x_max, ncol = y_max, dimnames = list(x_axis, y_axis))
# New - Determine rows and columns of matrix as well as the filter kernel
nrow_window <- nrow(window)
ncol_window <- ncol(window)
nrows <- nrow(filter_matrix)
ncols <- ncol(filter_matrix)
# New - Figure out where to cut off
row_cutoff <- floor(nrow_window/2)
col_cutoff <- floor(ncol_window/2)
# New - Remove out borders
filter_matrix <- filter_matrix[((1+row_cutoff):(nrows-row_cutoff)), ((1+col_cutoff):(ncols-col_cutoff))]
# Finally return matrix
return(filter_matrix)
使用您的数据:
> image <- t(matrix(c(1:36), nrow=6, ncol=6))
> image
[,1] [,2] [,3] [,4] [,5] [,6]
[1,] 1 2 3 4 5 6
[2,] 7 8 9 10 11 12
[3,] 13 14 15 16 17 18
[4,] 19 20 21 22 23 24
[5,] 25 26 27 28 29 30
[6,] 31 32 33 34 35 36
> filter <- matrix(c(1,0,-1,2,0,-2,1,0,-1), nrow=3, ncol=3)
> filter
[,1] [,2] [,3]
[1,] 1 2 1
[2,] 0 0 0
[3,] -1 -2 -1
我运行了这个功能,现在我得到了:
> filter2D(image,filter)
2 3 4 5
2 -48 -48 -48 -48
3 -48 -48 -48 -48
4 -48 -48 -48 -48
5 -48 -48 -48 -48
我认为让水平和垂直标签保持原样可能很重要,这样你就可以明确地看到并没有返回所有信号,这就是代码当前正在做的事情......那是尽管如此。我会留给你决定。