Fire Detection using Image Processing

Hello!My project is stuck in at the point where my algorithm for detecting fire pixels detects areas that aren't fire.For some images it works perfectly but there are many things to improve.

What should I do to reduce areas that are not fire.I've been thinking removing regular areas but I think the fire pixel algorithm could be improved.What do you think?

Any help is welcomed.

The images resulted are

and

function densitate=calcDensitate(test,densitate,ksize)
 r=(ksize-1)/2;
 [linii,coloane]=size(densitate);
 rowBound=linii-r;
 colBound=coloane-r;
 for j=r+1:colBound
     col1=j-r;
     col2=j+r;
     delta=1;
     for k=1:ksize
         delta=delta+test(k,col2)-test(k,col1);
     end
     densitate(r,j)=densitate(r,j-1)+delta/255;
 end
 for i=r+1:rowBound
     row1=i-r;
     row2=i+r;
     delta=1;
     for k=1:ksize
         delta=delta+test(row2,k)+test(row1,k);
     end
     densitate(i,r)=densitate(i-1,r)+delta/255;
 end
 for i=r+1:rowBound
     for j=r+1:colBound
         delta=(test(r+i,j+r)-test(i-r,j+r)-test(i+r,j-r)+test(i-r,j-r))/255;
         densitate(i,j)=densitate(i-1,j)+densitate(i,j-1)-densitate(i-1,j-1)+delta;
     end
 end 
 

This is the main function

img=imread('cadru230.jpg');
 img=imgaussfilt(img,[3 3]);
 tau=40;
 subplot(2,2,1);
 imshow(img);
 red=img(:,:,1);
green=img(:,:,2);
blue=img(:,:,3);
%img=im2double(img);
hsv=rgb2hsv(img); 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
ycbcrmap=rgb2ycbcr(img);
 Y=ycbcrmap(:,:,1);
% Y=round(Y,2);
 Cb=ycbcrmap(:,:,2);
% Cb=round(Cb,2);
 Cr=ycbcrmap(:,:,3);
% Cr=round(Cr,2);
 %%%%%%%%%%%%%calculare Ymean
 [a,b]=size(Y);
 SumaLinii=sum(Y);
 total=sum(SumaLinii);
 Ymean=total/(a*b);
  [a,b]=size(Cb);
 SumaLinii=sum(Cb);
 total=sum(SumaLinii);
 CbMean=total/(a*b);
% CbMean=round(CbMean,2);
  %%%%%%%%%%%%%calculare CrMean
  [a,b]=size(Cr);
 SumaLinii=sum(Cr);
 total=sum(SumaLinii);
 CrMean=total/(a*b);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
h=hsv(:,:,1);
 h=round(h,2);
 s=hsv(:,:,2);
 s=round(s,2);
 v=hsv(:,:,3);
 v=round(v,2);
[lin,col]=size(s);
test=ones(lin,col)*255;
for i=1:lin
    for j=1:col
        ss=1-3.0*min(red(i,j),min(green(i,j),red(i,j)))/(red(i,j) +green(i,j)+blue(i,j));
        if(~(abs(Cb(i,j)-Cr(i,j)<tau)&&red(i,j)>180 && red(i,j)>=green(i,j) && green(i,j)>blue(i,j) && ss>=((255-red(i,j))*0.2/180) &&(v(i,j)>0.8 && v(i,j)<1)&& (Y(i,j)>=Ymean) && (Cb(i,j)<=CbMean) && (Cr(i,j)>=CrMean)))
       %if(~(red(i,j)>=green(i,j) && green(i,j)>blue(i,j) && (red(i,j)>115 &&red(i,j)<135) && s(i,j)>(255-(red(i,j)*60)/120)))
            test(i,j)=0;
               %contor=contor+1;
        end
    end
end

BW2 = bwareaopen(test,100);

test=imclearborder(test);
stats = struct2table(regionprops(test,{'Area','Solidity','PixelIdxList'}));
idx = stats.Solidity < 0.9 | stats.Area <350;
for kk = find(idx)'
  test(stats.PixelIdxList{kk}) = true;
end
 subplot(2,2,2);
 imshow(test);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%aplic denoise
r=(7-1)/2;
densitate=zeros(lin,col);
densitate=calcDensitate(test,densitate,7);
for i=r:lin-r
    for j=r:col-r
        count=densitate(i,j);
        if(count<5)
            test(i,j)=0;
        end
    end
end
 subplot(2,2,3);
 imshow(test);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%fill


test=medfilt2(test);




sv=fspecial('sobel');
sx=sv';
Gy=imfilter(test,sv,'replicate');
Gx=imfilter(test,sx,'replicate');
TL=0.39; %prag mic si prag mare
TH=0.98;

% kk=edge(test,'sobel',0.39);
% ll=edge(test,'sobel',0.98);
 kk=edge(test,'sobel',0.39);
ll=edge(test,'sobel',0.98);
lll=edge(test,'canny',[TL TH]);
final=imadd(kk,ll);
edgePixels=nnz(final);
area=bwarea(final);
 subplot(2,2,4);
 imshow(final);

[Shubham Channe]

Sir I'm Also trying same project and met with same issue if u got the coorect code can u please help me with correct code.

[KALYAN ACHARJYA]

Because here you are manually adjusting the thresholding values based on specific single image

[G Gion]

When we run this programme we are getting The error as not enough arguments

[Walter Roberson]

You have to run the code that is labeled "This is the main function".

Answer 1

[Alex Che]

That's the solution i found 8 months ago.I used the fluctuation of area and perimeter for every fire-like object and I put it into an algorithm to figure out which is fire. I also put a graphic which show the variation of area and number of pixels and for almost every video input works. In the frame 150 you observe a great variation of area for an object which represent the spreading fire.

Answer 2

[Image Analyst]

Since there will be lamps and other things that could also be bright in the scene I think you'll have to analyze a video stream and see if the area of the bright things varies over time. See attached demo and adapt it as needed.

[Alex Che]

This is the solution that seemed the best to me as well if you don't have enough data to try a Convolutional Neural Network