提高matlab代码的运行时间
我编写了一个matlab代码,用于使用Hough变换检测灰度图像中的圆。我希望尽可能减少运行时间 我使用的边缘检测是自定义实现,但是它的运行时间足够快,可以满足我的需要(大约0.06秒)。但是,瓶颈是代码的其余部分(总运行时间约为6.35秒)。顺便说一句,我使用tic/toc来计算运行时间 这是代码,如果有人能看一下,我将非常感激:提高matlab代码的运行时间,matlab,optimization,image-processing,Matlab,Optimization,Image Processing,我编写了一个matlab代码,用于使用Hough变换检测灰度图像中的圆。我希望尽可能减少运行时间 我使用的边缘检测是自定义实现,但是它的运行时间足够快,可以满足我的需要(大约0.06秒)。但是,瓶颈是代码的其余部分(总运行时间约为6.35秒)。顺便说一句,我使用tic/toc来计算运行时间 这是代码,如果有人能看一下,我将非常感激: function [ circles ] = findCircles(img) % set low and high bounds for radii valu
function [ circles ] = findCircles(img)
% set low and high bounds for radii values
minR = 9;
[imgRows, imgCols] = size(img);
maxR = ceil(min(imgRows, imgCols)/2);
tic
% run edge detection on image
edgeImg = edgeDetect(img);
% get image size
[rows, cols] = size(edgeImg);
% initialize accumulator
houghAcc = zeros(rows, cols, maxR);
% get all edge pixels from image
edges = find(edgeImg);
% find number of edge pixels
edgeNum = size(edges);
% scan each edge
for currEdge = 1 : edgeNum
% get current edge x and y coordinations
[edgeY edgeX] = ind2sub([rows, cols], edges(currEdge));
% scan each all possible radii
for r = minR : maxR
% go over all possible 2*pi*r circle centers
for ang = 0 : 360
t = (ang * pi) / 180;
cX = round(edgeX - r*cos(t));
cY = round(edgeY - r*sin(t));
% check if center found is within image boundaries
if ( cX < cols && cX > 0 && cY < rows && cY > 0 )
% found circle with (cX,cY) as center and r as radius
houghAcc(cY,cX,r)=houghAcc(cY,cX,r)+1; % increment matching counter
end
end
end
end
% initialize circle list
circles = [];
% intialize index for next found circle
nextCircleIndx = 1;
% get counter list dimensions
[sizeX sizeY sizeR] = size(houghAcc);
% get max counter value from hough counter matrix
m = max(max(max(houghAcc)));
% calculate the minimal pixels that circle should have on perimeter
t = m * 0.42;
% scan each found circle
for cX = 1 : sizeX
for cY = 1 : sizeY
for r = 1 : sizeR
% threshold values
if houghAcc(cX, cY, r) > t
% circle is dominant enough, add it
circles(nextCircleIndx,:) = [cY , cX , r ,houghAcc(cX, cY, r)];
% increment index
nextCircleIndx = nextCircleIndx + 1;
end
end
end
end
% sort counters in descending order (according to votes for each
% circle)
circles = flipud(sortrows(circles,4));
% get circle list's size
[rows cols] = size(circles);
% scan circle list and check each pair of found circles
for i = 1 : rows-1
% get first circle's details:
% center
cX1 = circles(i,1);
cY1 = circles(i,2);
% radius
r1 = circles(i,3);
%hough counter
h1 = circles(i,4);
for j = i+1 : rows
%get second circle's details:
% center
cX2 = circles(j,1);
cY2 = circles(j,2);
% radius
r2 = circles(j,3);
%hough counter
h2 = circles(j,4);
% check if circle's actual difference is smaller than minimal
% radius allowed
if (cX1 - cX2)*(cX1 - cX2)+ (cY1 - cY2)*(cY1 - cY2) < (min(r1,r2))*(min(r1,r2)) && abs(r1 - r2) < minR
% both circles are similar, sum their counters and merge
% them to a circle with their avaraged values
circles(i,:)=[(cX1+cX2)/2, (cY1+cY2)/2, (r1+r2)/2, h1+h2];
% remove similar circle
circles(j,:)=[0,0,0,0];
end
end
end
sortParam = 3; % 1: x-center, 2: y-center, 3: radius, 4: hough counter
% sort the circles by the sort parameter, in descending order
circles = flipud(sortrows(circles,sortParam));
% get number of remained circles (= rows with non-zero values)
len = length(find(circles~=0))/4;
% remove duplicate similar circles from previus step
circles(circles == 0) = [];
% reshape circle list back to matrix form (previous step converted it
% to a vector)
circles = reshape(circles,len,4);
% get max value according to sort parameter
m = max(circles(:,sortParam));
%get size of new circle list (with no duplicate circles)
[newH newW] = size(circles);
% thresholding: remove hough counters that are less than 30% from sort
% parameter
for i= 1 : newH
% check if current circle's sorting parameter's value is smaller
% than threshold
if m - circles(i,sortParam) < m * 0.3
% plot(circles(i,1),circles(i,2),'xr'); % DEBUG - show centers
else
% remove current circle
circles(i,:)=[0,0,0,0];
end
end
% find number of remaining circles after thresholding
len = length(find(circles~=0))/4;
% delete rows that match circles removed in thresholding
circles(circles==0)=[];
% reshape circle list back to matrix form
circles=reshape(circles,len,4);
% convert circle list's values to integers (hough counters are already
% integers)
circles = uint8(circles(:,1:3));
toc
end
function[circles]=findCircles(img)
%设置半径值的下限和上限
minR=9;
[imgRows,imgCols]=大小(img);
maxR=ceil(最小值(imgrowth,imgCols)/2);
抽搐
%对图像进行边缘检测
edgeImg=edgeDetect(img);
%获取图像大小
[行,列]=大小(edgeImg);
%初始化累加器
houghAcc=零(行、列、最大值);
%从图像中获取所有边缘像素
边=查找(边);
%查找边缘像素数
edgeNum=尺寸(边);
%扫描每条边
对于currEdge=1:edgeNum
%获取当前边x和y坐标
[edgeY edgeX]=ind2sub([rows,cols],edges(currendge));
%扫描每个可能的半径
对于r=minR:maxR
%检查所有可能的2*pi*r圆心
对于ang=0:360
t=(ang*pi)/180;
cX=圆形(edgeX-r*cos(t));
cY=圆形(edgeY-r*sin(t));
%检查找到的中心是否在图像边界内
if(cX0&&cY0)
%以(cX,cY)为中心,以r为半径的圆
houghAcc(cY,cX,r)=houghAcc(cY,cX,r)+1;%增量匹配计数器
结束
结束
结束
结束
%初始化循环列表
圆圈=[];
%初始化下一个找到的圆的索引
nextCircleIndx=1;
%获取计数器列表维度
[sizeX sizeY sizeR]=尺寸(houghAcc);
%从hough计数器矩阵获取最大计数器值
m=最大值(最大值(最大值(houghAcc));
%计算圆周长上应有的最小像素
t=m*0.42;
%扫描找到的每个圆圈
对于cX=1:sizeX
对于cY=1:sizeY
对于r=1:sizeR
%阈值
如果houghAcc(cX,cY,r)>t
%圆圈足够占主导地位,添加它
圆(nextCircleIndx,:)=[cY,cX,r,houghAcc(cX,cY,r)];
%增量指数
nextCircleIndx=nextCircleIndx+1;
结束
结束
结束
结束
%按降序排序计数器(根据每个计数器的投票)
%圆圈)
圆圈=flipud(sortrows(圆圈,4));
%获取圆列表的大小
[行cols]=大小(圆);
%扫描圆列表并检查找到的每对圆
对于i=1:rows-1
%获取第一个圆圈的详细信息:
%居中
cX1=圆(i,1);
cY1=圆(i,2);
%半径
r1=圆(i,3);
%霍夫计数器
h1=圆(i,4);
对于j=i+1:行
%获取第二个圆圈的详细信息:
%居中
cX2=圆(j,1);
cY2=圆(j,2);
%半径
r2=圆(j,3);
%霍夫计数器
h2=圆(j,4);
%检查圆的实际差异是否小于最小值
%允许半径
如果(cX1-cX2)*(cX1-cX2)+(cY1-cY2)*(cY1-cY2)<(min(r1,r2))*(min(r1,r2))&abs(r1-r2)
在哪里可以改进此代码?谢谢你的帮助 对于填充
houghAcc
矩阵的第一个块,我建议进行以下替换:
r = minR : maxR;
t = ( 0 : 359 ) * pi / 180; % following HighPerformaceMark suggestion
rsin = bsxfun( @times, r', sin(t) ); %'
rcos = bsxfun( @times, r', cos(t) ); %'
[edgeY edgeX] = find( edgeImg );
cX = round( bsxfun( @minus, edgeX, permute( rcos, [3 1 2] ) ) );
cY = round( bsxfun( @minus, edgeY, permute( rsin, [3 1 2] ) ) );
R = permute( repmat( r', [ 1 size(cX,1) size(cX,3) ] ), [2 1 3] ); %' to index accHough
% select valid indices
sel = ( cX > 0 & cY > 0 & cY < rows & cX < cols );
houghAcc = accumarray( {cY(sel(:)), cX(sel(:)), R(sel(:))}, 1, [rows, cols, maxR] );
ind = find( houghAcc > t );
% sort the scores
sc = houghAcc(ind);
[sc si] = sort( sc , 'descend' );
% convert linear indices to x,y,r
[cX cY r] = ind2sub( size( houghAcc ), ind(si) );
circles = [ cX(:) cY(:) r(:) sc(:) ];
for-for-for-if在算法中出现两次。如果其他人没有帮助我使用profile
运行它,我将尝试稍后再讨论这个问题,这样您就可以更好地了解代码的大部分时间都花在了什么地方。利润