Image Classification¶

Classify images using ResNet or EfficientNet models. Pass an image path or BufferedImage, get back labeled predictions with confidence scores.

Quick example¶

try (var classifier = ResNetClassifier.builder().build()) {
    List<Classification> results = classifier.classify(Path.of("cat.jpg"));
    // [Classification[label=tabby cat, confidence=0.87], ...]
}

Full example¶

import io.github.inference4j.vision.ResNetClassifier;
import io.github.inference4j.vision.Classification;
import java.nio.file.Path;

public class ImageClassification {
    public static void main(String[] args) {
        try (var classifier = ResNetClassifier.builder().build()) {
            List<Classification> results = classifier.classify(Path.of("cat.jpg"), 5);

            for (Classification c : results) {
                System.out.printf("%-30s %.2f%%%n", c.label(), c.confidence() * 100);
            }
        }
    }
}

Available models¶

ResNet¶

The default choice for image classification. Uses ImageNet normalization and softmax output.

try (var classifier = ResNetClassifier.builder().build()) {
    classifier.classify(Path.of("image.jpg"));
}

EfficientNet¶

An alternative architecture with built-in softmax. Good for models exported from TensorFlow.

try (var classifier = EfficientNetClassifier.builder().build()) {
    classifier.classify(Path.of("image.jpg"));
}

Note

EfficientNet-Lite4 (the default) has softmax built into the model. If you use a PyTorch-exported EfficientNet that outputs raw logits, override with .outputOperator(OutputOperator.softmax()).

ResNet builder options¶

Method	Type	Default	Description
`.modelId(String)`	`String`	`inference4j/resnet50-v1-7`	HuggingFace model ID
`.modelSource(ModelSource)`	`ModelSource`	`HuggingFaceModelSource`	Model resolution strategy
`.sessionOptions(SessionConfigurer)`	`SessionConfigurer`	default	ONNX Runtime session config
`.labels(Labels)`	`Labels`	auto-loaded from `labels.txt`	Classification labels
`.outputOperator(OutputOperator)`	`OutputOperator`	`softmax()`	Output activation function
`.defaultTopK(int)`	`int`	`5`	Default number of top predictions

EfficientNet builder options¶

Method	Type	Default	Description
`.modelId(String)`	`String`	`inference4j/efficientnet-lite4`	HuggingFace model ID
`.modelSource(ModelSource)`	`ModelSource`	`HuggingFaceModelSource`	Model resolution strategy
`.sessionOptions(SessionConfigurer)`	`SessionConfigurer`	default	ONNX Runtime session config
`.labels(Labels)`	`Labels`	auto-loaded from `labels.txt`	Classification labels
`.outputOperator(OutputOperator)`	`OutputOperator`	`identity()`	Output activation (softmax built-in)
`.defaultTopK(int)`	`int`	`5`	Default number of top predictions

Result type¶

Classification is a record with:

Field	Type	Description
`label()`	`String`	ImageNet class label (e.g., `tabby cat`)
`index()`	`int`	Class index (0–999 for ImageNet)
`confidence()`	`float`	Confidence score (0.0 to 1.0)

Hardware acceleration¶

Image classification benefits significantly from hardware acceleration:

Model	CPU	CoreML	Speedup
ResNet-50	37 ms	10 ms	3.7x

try (var classifier = ResNetClassifier.builder()
        .sessionOptions(opts -> opts.addCoreML())
        .build()) {
    classifier.classify(Path.of("cat.jpg"));
}

See the Hardware Acceleration guide for details.

Tips¶

Use classify(image, topK) to control how many predictions are returned.
Both classifiers accept Path or BufferedImage as input.
Input images are automatically resized and normalized — no manual preprocessing needed.
Input size is auto-detected from the model (ResNet: 224x224, EfficientNet-Lite4: 280x280).