基于CNN卷积神经网络的目标识别算法matlab仿真,包含仿真操作录像,代码注释

function cnn=bpcnn(cnn,yy) if cnn.layers{cnn.no_of_layers}.type ~= 'f' zz=[]; for k=1:cnn.layers{cnn.no_of_layers}.no_featuremaps ss =size(cnn.layers{cnn.no_of_layers}.featuremaps{k}); zz =[zz; reshape(cnn.layers{cnn.no_of_layers}.featuremaps{k}, ss(1)*ss(2), ss(3))]; end cnn.layers{cnn.no_of_layers}.outputs = zz; end er = ( cnn.layers{cnn.no_of_layers}.outputs - yy); if cnn.loss_func == 'cros' %cross_entropy' if cnn.layers{cnn.no_of_layers}.act_func == 'sigm' er1 = -1.*sum((yy.*log(cnn.layers{cnn.no_of_layers}.outputs) + (1-yy).*log(1-cnn.layers{cnn.no_of_layers}.outputs)), 1); else % er1 = -1.*sum((yy.*log(cnn.layers{cnn.no_of_layers}.outputs)), 1); error('cross entropy is implemented only when last layer is sigmoid'); end cnn.loss = sum(er1(:))/size(er1,2); %loss over all examples else er1 = er.^2; cnn.loss = sum(er1(:))/(2*size(er1,2)); %loss over all examples end if cnn.CalcLastLayerActDerivative ==1 er =applyactfunccnn(cnn.layers{cnn.no_of_layers}.outputs,cnn.layers{cnn.no_of_layers}.act_func, 1, er); end %%%%%%%%%%%%%%first only last layer calculation if cnn.layers{cnn.no_of_layers}.type == 'f' cnn.layers{cnn.no_of_layers}.er{1} =er; cnn.layers{cnn.no_of_layers}.dW = cnn.layers{cnn.no_of_layers}.er{1} * ( cnn.layers{cnn.no_of_layers-1}.outputs)' / size(cnn.layers{cnn.no_of_layers}.er{1}, 2); cnn.layers{cnn.no_of_layers}.db = mean( cnn.layers{cnn.no_of_layers}.er{1}, 2); elseif cnn.layers{cnn.no_of_layers}.type == 'p' sz2=0; for i=1:cnn.layers{cnn.no_of_layers}.no_featuremaps sz = size(cnn.layers{cnn.no_of_layers}.featuremaps{i}); sz1 = sz(1)*sz(2); cnn.layers{cnn.no_of_layers}.er{i} = reshape(er(sz2+1 : sz2+sz1, : ), sz(1), sz(2), sz(3)); sz2 = sz2+sz1; end %%upsample er for previous layer zz=cnn.layers{cnn.no_of_layers}.subsample_rate; if cnn.layers{cnn.no_of_layers}.subsample_method == 'mean' for i=1:cnn.layers{cnn.no_of_layers}.no_featuremaps sz = size(cnn.layers{cnn.no_of_layers}.featuremaps{i}); ss1 = 1:sz(1); ss1 =kron(ss1, ones([1 zz])); ss2 = 1:sz(2); ss2 =kron(ss2, ones([1 zz])); sf{1}=ss1; sf{2}=ss2; er =cnn.layers{cnn.no_of_layers}.er{i}; new_er = er(sf{:},:); cnn.layers{cnn.no_of_layers}.er{i} = new_er; % kron( cnn.layers{cnn.no_of_layers}.er{i}, ones([zz zz])); end else error 'this subsampling method not implemented'; end elseif cnn.layers{cnn.no_of_layers}.type == 'c' sz2=0; for i=1:cnn.layers{cnn.no_of_layers}.no_featuremaps sz = size(cnn.layers{cnn.no_of_layers}.featuremaps{i}); sz1 = sz(1)*sz(2); cnn.layers{cnn.no_of_layers}.er{i} = reshape(er(sz2+1 : sz2+sz1, : ), sz(1), sz(2), sz(3)); sz2 = sz2+sz1; end kk=0; for i=1:cnn.layers{cnn.no_of_layers}.no_featuremaps for j=1:cnn.layers{cnn.no_of_layers-1}.no_featuremaps zz= convn(cnn.layers{cnn.no_of_layers-1}.featuremaps{j}, rot90(cnn.layers{cnn.no_of_layers}.er{i},2), 'valid'); kk = kk+1; cnn.layers{cnn.no_of_layers}.dK(:,:,kk) = mean(zz,3); end cnn.layers{cnn.no_of_layers}.db(i)= sum(cnn.layers{cnn.no_of_layers}.er{i}(:))/size(cnn.layers{cnn.no_of_layers}.er{i},3); end end for i=cnn.no_of_layers-1:-1:1 if cnn.layers{i}.type == 'f' cnn.layers{i}.er{1} = ( (cnn.layers{i+1}.W)' * cnn.layers{i+1}.er{1} ); cnn.layers{i}.er{1} = applyactfunccnn(cnn.layers{i}.outputs,cnn.layers{i}.act_func, 1, cnn.layers{i}.er{1} ); cnn.layers{i}.dW = cnn.layers{i}.er{1} * ( cnn.layers{i-1}.outputs)' / size(cnn.layers{i}.er{1}, 2); cnn.layers{i}.db = mean( cnn.layers{i}.er{1}, 2); elseif cnn.layers{i}.type == 'p' %er =cnn.layers{i+1}.er; cnn.layers{i}.er{1}= 0; if cnn.layers{i+1}.type == 'f' sz2=0; er = ( (cnn.layers{i+1}.W)' * cnn.layers{i+1}.er{1} ); for j=1:cnn.layers{i}.no_featuremaps sz = size(cnn.layers{i}.featuremaps{j}); sz1 = sz(1)*sz(2); cnn.layers{i}.er{j} = reshape(er(sz2+1 : sz2+sz1, : ), sz(1), sz(2), sz(3)); sz2 = sz2+sz1; end elseif cnn.layers{i+1}.type == 'c' er =cnn.layers{i+1}.er; for k=1:cnn.layers{i}.no_featuremaps cnn.layers{i}.er{k}=zeros(size(cnn.layers{i}.featuremaps{k})); end kk=0; for j=1:cnn.layers{i+1}.no_featuremaps for k=1:cnn.layers{i}.no_featuremaps kk = kk+1; cnn.layers{i}.er{k} = cnn.layers{i}.er{k} + convn(er{j}, rot90(cnn.layers{i+1}.K(:,:,kk),2), 'full'); end end else er =cnn.layers{i+1}.er; cnn.layers{i}.er =er; end %upsample er for previous layer zz=cnn.layers{i}.subsample_rate; if cnn.layers{i}.subsample_method == 'mean' for j=1:cnn.layers{i}.no_featuremaps sz = size(cnn.layers{i}.featuremaps{j}); ss1 = 1:sz(1); ss1 =kron(ss1, ones([1 zz])); ss2 = 1:sz(2); ss2 =kron(ss2, ones([1 zz])); sf{1}=ss1; sf{2}=ss2; er =cnn.layers{i}.er{j}; new_er = er(sf{:},:); cnn.layers{i}.er{j} = new_er./(zz*zz); % kron( cnn.layers{cnn.no_of_layers}.er{i}, ones([zz zz])); end else error 'this subsampling method not implemented'; end elseif cnn.layers{i}.type == 'c' er =0; cnn.layers{i}.er{1}= 0; if cnn.layers{i+1}.type == 'f' sz2=0; er1 = ( (cnn.layers{i+1}.W)' * cnn.layers{i+1}.er{1} ); for j=1:cnn.layers{i}.no_featuremaps sz = size(cnn.layers{i}.featuremaps{j}); sz1 = sz(1)*sz(2); cnn.layers{i}.er{j} = reshape(er1(sz2+1 : sz2+sz1, : ), sz(1), sz(2), sz(3)); sz2 = sz2+sz1; end er = cnn.layers{i}.er; elseif cnn.layers{i+1}.type == 'c' error('not implemented yet- convolution layer with convolution layer. Instead, use Pooling layer with subsampling factor 1 between two conv layer'); else er =cnn.layers{i+1}.er; end if cnn.layers{i}.act_func == 'soft' % error('softmax for backpropagation is not implemented yet'); err1 = zeros(size(er{1})); for j=1:cnn.layers{i}.no_featuremaps err1 = err1 + er{j}.*cnn.layers{i}.featuremaps{j}; end for j=1:cnn.layers{i}.no_featuremaps cnn.layers{i}.er{j} = cnn.layers{i}.featuremaps{j}.*(er{j} - err1); end else for j=1:cnn.layers{i}.no_featuremaps cnn.layers{i}.er{j} =applyactfunccnn(cnn.layers{i}.featuremaps{j},cnn.layers{i}.act_func, 1, er{j}); end end %%calculate gradient of filters and bias kk=0; for ii=1:cnn.layers{i}.no_featuremaps for j=1:cnn.layers{i-1}.no_featuremaps %%%%%following commented code is slow, but shows how we can %%%%%calculate dK % zz=0; % for mm=1:size(cnn.layers{