Deeplearning4J (DL4J) offers a comprehensive Java framework for deep learning, while Spring Boot streamlines the development of production-ready applications. By combining these two technologies, you unlock a flexible platform for building intelligent services that can handle various types of data. In this guide, we’ll explore how to integrate DL4J into your Spring Boot project and demonstrate its usage with both structured numerical data and images.


  • Basic knowledge of Spring Boot
  • Familiarity with Maven or Gradle
  • A pre-trained DL4J model (or the intention to train one)


  1. Project Setup

    • Start by creating a new Spring Boot project.
    • Add the DL4J and ND4J dependencies to your project’s build configuration (e.g., in your pom.xml for Maven or build.gradle for Gradle):
  • For image processing, include the JavaCV library and DL4J’s image processing tools:

    <artifactId>deeplearning4j-zoo</artifactId> <version>1.0.0-M2</version>

    <artifactId>deeplearning4j-datavec-image</artifactId> <version>1.0.0-M2</version>
  1. Model Loading

    • Place your pre-trained model(s) (usually with a .zip extension) in a suitable location within your Spring Boot project, such as the resources directory.
    • Create a Spring component (ModelLoader) to load and manage your models:
public class ModelLoader {

    private MultiLayerNetwork numericalModel;
    private MultiLayerNetwork imageModel; // If you have a separate image model

    public void init() throws IOException {
        numericalModel = ModelSerializer.restoreMultiLayerNetwork(
            new ClassPathResource("").getFile());
        // If using a separate model for image processing:
        imageModel = ModelSerializer.restoreMultiLayerNetwork(
            new ClassPathResource("").getFile()); 

    // ... (Getters for models)
  1. Handling Structured Data
public class PredictionController {

    private ModelLoader modelLoader;

    public Map<String, Double> predict(@RequestBody Map<String, Double> inputData) {
        INDArray input = Nd4j.create(inputData.values().stream().mapToDouble(d -> d).toArray());
        INDArray output = modelLoader.getNumericalModel().output(input);
        // ... (Process output and return results)
  1. Handling Images
public class ImageController {

    private ModelLoader modelLoader;
    private static final NativeImageLoader LOADER = new NativeImageLoader(224, 224, 3); // Example dimensions

    public Map<String, Double> classifyImage(@RequestParam("image") MultipartFile file) throws IOException {
        // Load the image
        Image image =;

        // Convert to INDArray 
        INDArray input = LOADER.asMatrix(image);

        // Preprocess (normalize, etc.) - Adjust as needed for your model
        // DataNormalization scaler = new ImagePreProcessingScaler(0, 1);
        // scaler.transform(input);

        // Perform prediction using the model
        INDArray output = modelLoader.getImageModel().output(input); 
        // ... Convert output to probabilities (e.g., using Softmax)
        // ... Return probabilities

DL4J Model Formats: A Deep Dive

Deeplearning4j (DL4J) is a versatile deep learning library for Java, supporting various model formats:

  • Keras: Import models trained with Keras (TensorFlow backend) seamlessly.
    ComputationGraph model = KerasModelImport.importKerasModelAndWeights("path/to/model.h5");
  • TensorFlow: Directly load TensorFlow models.
    ComputationGraph model = SavedModelBundle.load("path/to/model/directory").getComputationGraph();
  • ONNX: Utilize the Open Neural Network Exchange format for interoperability.
    ComputationGraph model = ModelImporter.importModel("path/to/model.onnx");
  • SavedModel: Work with TensorFlow’s SavedModel format. (Same as TensorFlow instructions above)
  • DL4J Model: Load a pre-trained or saved DL4J model.
    MultiLayerNetwork model = ModelSerializer.restoreMultiLayerNetwork("path/to/");


DL4J is designed for Java environments, often showing comparable or better performance than other Java-based libraries. Benchmarking deep learning models is complex and depends heavily on hardware, model architecture, and dataset.

Key Points & Best Practices

  • Model Optimization: Consider techniques like model quantization or pruning for production deployment to improve performance.
  • Model Type: Ensure your DL4J model is designed for the type of image analysis you want to perform (e.g., classification, object detection, segmentation).
  • Image Preprocessing: Pay careful attention to preprocessing steps. Image resizing, normalization, and other transformations are often crucial for good model performance.
  • Error Handling: Implement robust error handling for cases where the image upload fails, the file format is incorrect, or the model encounters issues.
  • Deployment: Package your application into an executable JAR and leverage containerization (e.g., Docker) for scalability.
  • Security: Always prioritize security when exposing AI models through APIs. If your API allows image uploads from untrusted sources, consider adding security measures to prevent malicious uploads.
  • Testing: Thoroughly test your model integration with various input types and edge cases.

By Jeffery Miller

I am known for being able to quickly decipher difficult problems to assist development teams in producing a solution. I have been called upon to be the Team Lead for multiple large-scale projects. I have a keen interest in learning new technologies, always ready for a new challenge.