It is a coding function net1.net to detect images

clc;clear;close all;

image_folder = 'daftar uang';

filenames = dir(fullfile(image_folder, '*.jpg'));

total_images = numel(filenames);

for n = 1:total_images

full_name= fullfile(image_folder, filenames(n).name);

I = imread(full_name);

J = I(:,:,1);

K = imbinarize(J,.6);

L = imcomplement(K);

str = strel('disk',5);

M = imclose(L,str);

N = imfill(M,'holes');

O = bwareaopen(N,1000);

stats = regionprops(O,'Area','Perimeter','Eccentricity');

area(n) = stats.Area;

perimeter(n) = stats.Perimeter;

metric(n) = 4*pi*area(n)/(perimeter(n)^2);

eccentricity(n) = stats.Eccentricity;

end

input = [metric;eccentricity];

target = zeros(1,48);

target(:,1:12) = 1;

target(:,13:24) = 2;

target(:,25:36) = 3;

target(:,37:48) = 4;

net = newff(input,target,[10 5],{'logsig','logsig'},'trainlm');

net.trainParam.epochs = 1000;

net.trainParam.goal = 1e-6;

net = train(net,input,target);

output = round(sim(net,input));

save net1.mat net

[m,n] = find(output==target);

akurasi = sum(m)/total_images*100

"Why is the accuracy I get only 25%? help me please?"

Answer 1

[Harsh]

Hi @Adi,

From what I can gather, the neural network mentioned above has the following characteristics:

• Two hidden layers with 10 and 5 neurons, respectively.
• The activation function being used is the logistic sigmoid.
• The Levenberg-Marquardt backpropagation algorithm is used for training the network.

Additionally, it looks like there are four categories for classifying the images, and the number of images used for training is only 48.

To improve accuracy and create a more robust network, consider the following suggestions:

• Increase the depth of the neural network, as the current one is too simple, as an example I have created 3 hidden layers in the following snippet of code with 20, 10 and 5 neurons respectively.

net = newff(input, target, [20 10 5], {'logsig', 'logsig', 'logsig'}, 'trainlm');

• If the dataset is limited to 48 images, try creating new images using techniques like augmentation. For example, images can be skewed using the "imageDataAugmenter" function:

augmentedImages = imageDataAugmenter('RandRotation', [-20, 20], 'RandXTranslation', [-3, 3], 'RandYTranslation', [-3, 3]);

• A pre-trained network like AlexNet can also be used for better results.

net = alexnet;
layersTransfer = net.Layers(1:end-3);
numClasses = 4;
layers = [
    layersTransfer
    fullyConnectedLayer(numClasses,'WeightLearnRateFactor',10,'BiasLearnRateFactor',10)
    softmaxLayer
    classificationLayer];
options = trainingOptions('sgdm', 'InitialLearnRate',0.001, 'MaxEpochs',20, 'MiniBatchSize',64, 'ValidationData',{valImages,valLabels}, 'ValidationFrequency',30, 'Verbose',false, 'Plots','training-progress');
netTransfer = trainNetwork(trainImages, trainLabels, layers, options);

It was also noticed that the same dataset is being used for both training and validation. Instead, split the dataset into two parts for better validation.

[trainInd, valInd, testInd] = dividerand(size(input, 2), 0.7, 0.15, 0.15);
net.divideFcn = 'divideind';
net.divideParam.trainInd = trainInd;
net.divideParam.valInd = valInd;
net.divideParam.testInd = testInd;

I hope this helps. Thanks!