灰度图像中的 MATLAB 鼠标分割，对阴影是不变的

Question

搜索了 2 或 3 天后，我仍然没有找到解决问题的方法。

我想创建一个没有阴影的鼠标分割。问题是如果我设法去除阴影，我也会去除尾巴和脚，这是一个问题。阴影来自鼠标所在的竞技场的墙壁。

我想从灰度图像中移除阴影，但我不知道该怎么做。首先我去掉了图片的背景，得到了下面的图片。

edit1：感谢您的回答，当阴影不接触鼠标时效果很好。这就是我得到的：

来自这个原始图像：

我正在从 tif 文件中提取每一帧，并为每一帧应用您的代码。这是我使用的代码：

for k=1:1000

    %reads image
    I = imread('souris3.tif',k);

    %first stage: perform thesholding and fill holes

    seg = I >20000;
    seg = imfill(seg,'holes');

    %fixes the missing tail problem
    %extract edges, and add them to the segmentation.
    edges =  edge(I);
    seg = seg | edges;

    %fill holes (again)
    seg = imfill(seg,'holes'); 

    %find all the connected components
    CC = bwconncomp(seg,8);

    %keeps only the biggest CC
    numPixels = cellfun(@numel,CC.PixelIdxList);
    [biggest,idx] = max(numPixels);
    seg = zeros(size(edges));
    seg(CC.PixelIdxList{idx}) = 1;

    imshow(seg);

end

我使用命令impixelinfo 选择20000 作为步长，因为图像在uint16 中，它是鼠标的平均值。

如果你想拥有 tif 文件，这是链接：

souris3.tif

感谢您的帮助。

Answer 1

我建议以下方法：

1. 对图像进行阈值处理，得到一个包含老鼠大部分身体的蒙版，没有尾巴和腿。
2. 使用 MATLAB 的 imfill 函数执行孔填充。在这个阶段，分割几乎是完美的，除了尾巴的一部分缺失。
3. 使用边缘图来找到尾部的边界。这可以通过将边缘图添加到分割并再次执行孔填充来完成。在这个阶段只保留最大的连接组件。

代码：

%reads image
I = rgb2gray(imread('mSWm4.png'));

%defines thersholds (you may want to tweak these thresholds, or find
%a way to calculate it automatically).
FIRST_STAGE_THRESHOLD = 70;
IM_BOUNDARY_RELEVANCE_THRESHOLD = 10;

%perform thesholding and fill holes, the tail is still missing
seg = I > FIRST_STAGE_THRESHOLD;
seg = imfill(seg,'holes');

%second stage fix the missing tail problem:
%extract edges from relevant areas (in which the matter is not too dark), and add them to the segmentation.
%the boundries of the image which are close enough to edges are also considered as edges
edges =  edge(I);
imageBoundries = ones(size(I));
imageBoundries(2:end-1,2:end-1) = 0;
relevantDistFromEdges = bwdist(edges) > IM_BOUNDARY_RELEVANCE_THRESHOLD;
imageBoundries(bwdist(edges) > IM_BOUNDARY_RELEVANCE_THRESHOLD) = 0;
seg = seg | (edges | imageBoundries);

%fill holes (again) and perform noise cleaning
seg = imfill(seg,'holes');
seg = getBiggestCC(imopen(seg,strel('disk',1)));

getBiggestCC 函数：

function [ res ] = getBiggestCC(mask)
CC = bwconncomp(mask,8);
numPixels = cellfun(@numel,CC.PixelIdxList);
[~,idx] = max(numPixels);
res = zeros(size(mask));
res(CC.PixelIdxList{idx}) = 1;
end

结果

每个阶段的结果：

结果 图片 1 结果：

图片 2 结果：

另一个视图（分割为红色）：