After parallelizing the code, time is increasing compared to serial code. How to optimize the code further parallely for time reduction

gaussianPyramid1 = genPyr(normaizedWeightmap1,'gauss',5); gaussianPyramid2 = genPyr(normaizedWeightmap2,'gauss',5);

tic for i = 1 : 5

    tempImg = [];
parpool('local',3)
  parfor j = 1 : size(im,3)
        laplacianPyramid1 = genPyr(firstInput(:,:,j),'laplace',5); %Generating Laplacian Pyramid for derrived inputs
        laplacianPyramid2 = genPyr(secondInput(:,:,j),'laplace',5);
        rowSize = min([size(laplacianPyramid1{i},1),size(laplacianPyramid2{i},1),size(gaussianPyramid1{i},1),size(gaussianPyramid2{i},1)]);
        columnSize = min([size(laplacianPyramid1{i},2),size(laplacianPyramid2{i},2),size(gaussianPyramid1{i},2),size(gaussianPyramid2{i},2)]);
        tempImg(:,:,j) = laplacianPyramid1{i}(1:rowSize , 1:columnSize) .* gaussianPyramid1{i}(1:rowSize, 1:columnSize) + laplacianPyramid2{i}(1:rowSize, 1:columnSize) .* gaussianPyramid2{i}(1:rowSize, 1:columnSize);
   end

    fusedPyramid1{i} = tempImg;
end
toc

[Jan]

What about pre-allocating tempImg?