Efficient-Net基于Efficient-Net效滤网的目标识别算法的MATLAB仿真

Posted 2022-03-23 fpga&matlab

%定义efficientnet的结构
layers = [
    imageInputLayer([128 128 3]);%注意，128,128，3是训练样本的大小，这个和参考文献不一样，要根据实际输入设置
    %stage1
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 1
    batchNormalizationLayer;
    reluLayer;
    maxPooling2dLayer(floor(resl)+1,'Stride',2);
    
    %stage2
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 1
    batchNormalizationLayer;
    reluLayer;
    
    %stage3
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 2
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 2
    batchNormalizationLayer;
    reluLayer; 
    
    %stage4
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 2
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 2
    batchNormalizationLayer;
    reluLayer;
    maxPooling2dLayer(floor(resl)+1,'Stride',2);
    
    %stage5
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    reluLayer;
    maxPooling2dLayer(floor(resl)+1,'Stride',2);
    
    %stage6
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 3
    batchNormalizationLayer;
    reluLayer;
    
    %stage7
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 4
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 4
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 4
    batchNormalizationLayer;
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 4
    batchNormalizationLayer;
    reluLayer;
    maxPooling2dLayer(floor(resl)+1,'Stride',2);
    
    %stage8
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 1
    batchNormalizationLayer;
    reluLayer;
    maxPooling2dLayer(floor(resl)+1,'Stride',2);
    
    %stage9
    convolution2dLayer(floor(depth)+1,floor(width)+1,'Padding','same');%layer 1
    batchNormalizationLayer;
    reluLayer;
    %FC
    fullyConnectedLayer(CLASSNUM);
    %softmax
    softmaxLayer;
    %输出分类结果
    classificationLayer;];





options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.01, ...
    'MaxEpochs', 200, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', imdsValidation, ...
    'ValidationFrequency', 5, ...
    'Verbose', false, ...
    'Plots', 'training-progress');
rng(1);
%使用训练集训练网络
net         = trainNetwork(imdsTrain, layers, options);

训练过程如下：

训练精度为94.17%。

平均损失过程如下：

不同训练样本数量对应的训练性能（注意，每次训练会有一定的波动和偏差）

训练样本比例	改进前的训练性能	改进后的训练性能
5%	85.46%	92.23%
10%	89.20%	90.08%
20%	94.65%	92.94%
40%	93.53%	94.82%
60%	94.66%	98.06%
80%	94.67%	98.08%
90%	98.08%	100%

 A05-79