View Example on GitHub

Lego Figure Recognition

This walkthrough trains a CNN to decide whether an image contains a Lego figure.

Problem: Classify an image as legoman or negative.

Dataset: The original indexed Deep Netts LegoPeople dataset, including repeated negative entries used for balancing.

At a glance

1. Configure the runtime

int width = 96, height = 96;
int epochs = Integer.getInteger("deepnetts.epochs", 15);
String dataset = "src/main/resources/datasets/lego-people";
System.setProperty("java.awt.headless", "true");
DeepNetts.getInstance().setUseCuda(false);

2. Load and preprocess the indexed dataset

ImageSet images = ImageData.load(dataset, "train.txt", width, height, false, true);
images.zeroMeanPerChannel();

Unlike the other examples, this project preserves the professor's original train.txt. It deliberately repeats negative entries to reduce class imbalance. The loader resizes RGB images, scales their values, and zero-centers each channel.

3. Inspect the images

System.out.println("Images: " + images.size());
System.out.println("Images by class: " + images.countByClasses());

Output:

Images: 634
Number of images by label/class
negativ : 6
negative : 172
legoman : 456
Images by class: {negativ=6, negative=172, legoman=456}

Counts describe indexed entries, including intentional repeats. The output also reveals six historical negativ labels in train.txt. They are counted separately from negative, so the current loaded dataset contains three label spellings even though the prediction task is intended to be binary.

4. Shuffle and split

images.shuffle(42);
ImageSet[] split = images.split(0.8, 0.2);
ImageSet trainingSet = split[0];
ImageSet testSet = split[1];

Output:

Splitting data set: [0.8, 0.2]

The fixed seed makes the 80/20 split reproducible.

5. Build the CNN

ConvolutionalNetwork network = ConvolutionalNetwork.builder()
        .addInputLayer(width, height, 3)
        .addConvolutionalLayer(12, Filters.ofSize(5), ActivationType.TANH)
        .addMaxPoolingLayer(Filters.ofSize(2).stride(2))
        .addFullyConnectedLayer(30, ActivationType.TANH)
        .addFullyConnectedLayer(10, ActivationType.TANH)
        .addOutputLayer(1, ActivationType.SIGMOID)
        .lossFunction(LossType.CROSS_ENTROPY)
        .randomSeed(42).build();

This architecture retains the character of the original example while using the current Deep Netts 4 API.

6. Configure and train

network.getTrainer().setStopEpochs(epochs).setLearningRate(0.001f)
        .setOptimizer(OptimizerType.ADAM).setBatchMode(true)
        .setBatchSize(16).setShuffle(true);
network.train(trainingSet);

Output:

TRAINING NEURAL NETWORK
--------------------------------------------------------------

Initial Train Error:0.6297051
Epoch:1, Time:53197ms, TrainError:0.45970678, TrainErrorChange:-0.16999829, TrainAccuracy:0.7968442

TRAINING COMPLETED
Total Training Time: 59585ms
--------------------------------------------------------------

This one-epoch smoke test reduced training error and reached about 79.7% training accuracy. A longer run is still needed to assess final model quality.

7. Evaluate on held-out images

System.out.println(network.test(testSet));

Output:

Class: legoman
Total items: 126
True positive:86.0 Number of examples correctly classified as positive
True negative:17.0 Number of examples correctly classified as negative
False positive:18.0 Number of examples incorrectly classified as positive
False negative:5.0 Number of examples incorrectly classified as negative
Accuracy (ACC): 0.8174603 How often is a classifier correct in total (percent of correct classifications)
Precision (PPV): 0.8269231 How often is a classifier correct when it gives positive prediction
Recall: 0.94505495 When it is actually positive class, how often does it give positive prediction
F1 Score: 0.8820513 Harmonic average (balance) of precision and recall
Specificity (TNR): 0.4857143 When it is actually negative class, how often does it give negative prediction
Fall-out (FPR): 0.51428574 How often it gives false positive prediction in total (percent of false positive predictions)
False negative rate (FNR): 0.054945055 How often it gives false negative prediction in total (percent of false negative predictions)

The model finds most Lego figures, as shown by high recall, but its specificity is below 0.5 because 18 negative images were incorrectly predicted as legoman. The held-out metrics are more informative than training accuracy.

8. Save the model

Files.createDirectories(Path.of("models"));
network.save("models/lego-figure-classifier.dnet");

The generated .dnet file contains the trained network.

9. Make a prediction

ExampleImage image = testSet.get(0);
float probability = network.predict(image.getInput()).getValues()[0];
String predicted = probability >= 0.5f ? "legoman" : "negative";
System.out.println("Actual: " + image.getLabel());
System.out.println("Predicted: " + predicted);
System.out.printf("Lego probability: %.4f%n", probability);

Output:

Actual: negative
Predicted: legoman
Lego probability: 0.7656

The Sigmoid output is above the 0.5 threshold, so the model predicts legoman. The actual label is negative, making this one of the false positives reflected in the evaluation report.

10. Release resources

network.getThreadPool().shutdown();

The final call stops the network's background worker threads.

Run the example

Pass these options to the Java process through your IDE's application run configuration or the command line:

--add-modules=jdk.incubator.vector -Ddeepnetts.epochs=1

Output:

WARNING: Using incubator modules: jdk.incubator.vector

This one-epoch run completed in about 62 seconds. Timings, metrics, and probabilities can differ between machines and runs. Remove the epoch override to use the default 15 epochs.

← Back to Image Classification Examples