一、将 Ultralytics YOLO26 模型导出为 ONNX 格式
是通过ultralytics-yolo的源码将yolo26n.pt的模型转换成onnx格式，将pt格式的模型转换成onnx，ONNX Runtime 支持各种平台，例如 Windows、macOS 和 Linux，可以使用 ONNX Runtime 运行，java貌似支持。ONNX Runtime 可以提供高达 3 倍的 CPU 加速。
python
from ultralytics import YOLO
def export_onnx():#将 Ultralytics YOLO26 模型导出为 ONNX 格式
# Load the YOLO26 model
model = YOLO("yolo26n.pt")
# Export the model to ONNX format
model.export(format="onnx") # creates 'yolo26n.onnx'
# Load the exported ONNX model
onnx_model = YOLO("yolo26n.onnx")
# Run inference
# results = onnx_model(source='ultralytics/assets/bus.jpg', save=True)
if __name__ == '__main__':
export_onnx()
二、准备yolo26x.pt模型的分类数据集
官网上关于
yolo26n.pt的数据集有80个，整理成coco.name
三、准备一张yolo官网的测试图片
四、java环境下引入依赖
XML
<!-- ONNX Runtime Java API -->
<dependency>
<groupId>com.microsoft.onnxruntime</groupId>
<artifactId>onnxruntime</artifactId>
<version>1.17.3</version>
</dependency>
<!-- OpenCV Java 绑定，用于图像处理和绘制 -->
<dependency>
<groupId>org.openpnp</groupId>
<artifactId>opencv</artifactId>
<version>4.5.5-1</version>
</dependency>五、直接上代码
java
import ai.onnxruntime.*;
import nu.pattern.OpenCV;
import org.opencv.core.*;
import org.opencv.imgproc.Imgproc;
import org.opencv.imgcodecs.Imgcodecs;
import java.io.IOException;
import java.nio.FloatBuffer;
import java.util.*;
public class YOLOv26Inference {
static {
OpenCV.loadShared();
System.out.println("OpenCV loaded successfully");
}
private OrtSession session;
private OrtEnvironment env;
private long[] inputShape = new long[]{1, 3, 640, 640};
private float confThreshold = 0.45f;
private List<String> classes;
public YOLOv26Inference(String modelPath, String classesPath) throws OrtException, IOException {
env = OrtEnvironment.getEnvironment();
OrtSession.SessionOptions opts = new OrtSession.SessionOptions();
opts.setOptimizationLevel(OrtSession.SessionOptions.OptLevel.ALL_OPT);
this.session = env.createSession(modelPath, opts);
classes = loadClasses(classesPath);
System.out.println("Loaded " + classes.size() + " classes");
}
public void detectAndDraw(String imagePath, String outputPath) throws Exception {
// 1. 读取图片
Mat img = Imgcodecs.imread(imagePath);
if (img.empty()) {
throw new Exception("图片加载失败: " + imagePath);
}
// 获取原始图像尺寸
int originalWidth = img.width();
int originalHeight = img.height();
System.out.println("原始图像尺寸: " + originalWidth + " x " + originalHeight);
// 2. 预处理（保持宽高比，使用 Letterbox）
float[] scale = new float[1];
int[] pad = new int[2];
float[][][] inputTensor = preprocessWithLetterbox(img, scale, pad);
// 3. 创建输入张量
OnnxTensor tensor = OnnxTensor.createTensor(env, FloatBuffer.wrap(flatten(inputTensor)), inputShape);
// 4. 推理
Map<String, OnnxTensor> inputs = Collections.singletonMap("images", tensor);
OrtSession.Result results = session.run(inputs);
// 5. 解析输出
String outputName = results.iterator().next().getKey();
OnnxTensor outputTensor = (OnnxTensor) results.get(outputName).get();
long[] shape = outputTensor.getInfo().getShape();
System.out.println("Output shape: " + Arrays.toString(shape));
// 获取输出数据 [1, 300, 6]
float[][][] output = (float[][][]) outputTensor.getValue();
List<Detection> detections = parseDetectionsWithScale(output[0], originalWidth, originalHeight, scale[0], pad[0], pad[1]);
System.out.println("检测到 " + detections.size() + " 个物体");
// 6. 绘制检测框
for (Detection det : detections) {
// 绘制矩形框
Imgproc.rectangle(img,
new Point(det.x1, det.y1),
new Point(det.x2, det.y2),
new Scalar(0, 255, 0), 2);
// 绘制标签背景
String label = String.format("%s: %.2f", classes.get(det.classId), det.confidence);
int[] baseLine = new int[1];
Size labelSize = Imgproc.getTextSize(label, Imgproc.FONT_HERSHEY_SIMPLEX, 0.5, 1, baseLine);
// 确保标签不会超出图像顶部
double labelY = det.y1 - 5;
if (labelY - labelSize.height < 0) {
labelY = det.y1 + labelSize.height + 5;
Imgproc.rectangle(img,
new Point(det.x1, det.y1),
new Point(det.x1 + labelSize.width, det.y1 + labelSize.height + 5),
new Scalar(0, 255, 0), -1);
Imgproc.putText(img, label,
new Point(det.x1, det.y1 + labelSize.height),
Imgproc.FONT_HERSHEY_SIMPLEX, 0.5,
new Scalar(0, 0, 0), 1);
} else {
Imgproc.rectangle(img,
new Point(det.x1, det.y1 - labelSize.height - 5),
new Point(det.x1 + labelSize.width, det.y1),
new Scalar(0, 255, 0), -1);
Imgproc.putText(img, label,
new Point(det.x1, det.y1 - 5),
Imgproc.FONT_HERSHEY_SIMPLEX, 0.5,
new Scalar(0, 0, 0), 1);
}
}
// 7. 保存结果
Imgcodecs.imwrite(outputPath, img);
System.out.println("检测完成，结果保存至：" + outputPath);
results.close();
tensor.close();
}
/**
* Letterbox 预处理：保持宽高比，填充到 640x640
*/
private float[][][] preprocessWithLetterbox(Mat img, float[] scale, int[] pad) {
int originalWidth = img.width();
int originalHeight = img.height();
// 计算缩放比例（保持宽高比）
float scaleW = 640.0f / originalWidth;
float scaleH = 640.0f / originalHeight;
float scaleFactor = Math.min(scaleW, scaleH);
scale[0] = scaleFactor;
int newWidth = Math.round(originalWidth * scaleFactor);
int newHeight = Math.round(originalHeight * scaleFactor);
// 计算填充
int padW = (640 - newWidth) / 2;
int padH = (640 - newHeight) / 2;
pad[0] = padW;
pad[1] = padH;
System.out.println("缩放比例: " + scaleFactor);
System.out.println("填充: padW=" + padW + ", padH=" + padH);
// 缩放图像
Mat resized = new Mat();
Imgproc.resize(img, resized, new Size(newWidth, newHeight));
// 创建 640x640 画布并填充灰色 (114,114,114) - YOLO 默认填充色
Mat canvas = new Mat(640, 640, resized.type(), new Scalar(114, 114, 114));
// 将缩放后的图像放到画布中央
resized.copyTo(canvas.submat(padH, padH + newHeight, padW, padW + newWidth));
// BGR 转 RGB
Mat rgbImg = new Mat();
Imgproc.cvtColor(canvas, rgbImg, Imgproc.COLOR_BGR2RGB);
// 归一化并转换为 NCHW 格式
float[][][] tensor = new float[3][640][640];
for (int i = 0; i < 640; i++) {
for (int j = 0; j < 640; j++) {
double[] pixel = rgbImg.get(i, j);
tensor[0][i][j] = (float) (pixel[0] / 255.0);
tensor[1][i][j] = (float) (pixel[1] / 255.0);
tensor[2][i][j] = (float) (pixel[2] / 255.0);
}
}
return tensor;
}
/**
* 解析检测结果，并缩放坐标回原始图像尺寸
*/
private List<Detection> parseDetectionsWithScale(float[][] output, int originalWidth, int originalHeight,
float scale, int padW, int padH) {
List<Detection> detections = new ArrayList<>();
for (int i = 0; i < output.length; i++) {
float[] pred = output[i];
float confidence = pred[4];
// 过滤低置信度
if (confidence < confThreshold) continue;
int classId = (int) pred[5];
// 模型输出的坐标是在 640x640（包含填充）上的
float modelX1 = pred[0];
float modelY1 = pred[1];
float modelX2 = pred[2];
float modelY2 = pred[3];
// 移除填充，得到在有效图像区域的坐标
float x1_noPad = modelX1 - padW;
float y1_noPad = modelY1 - padH;
float x2_noPad = modelX2 - padW;
float y2_noPad = modelY2 - padH;
// 缩放回原始图像尺寸
float x1 = x1_noPad / scale;
float y1 = y1_noPad / scale;
float x2 = x2_noPad / scale;
float y2 = y2_noPad / scale;
// 确保坐标在图像范围内
x1 = Math.max(0, Math.min(originalWidth, x1));
y1 = Math.max(0, Math.min(originalHeight, y1));
x2 = Math.max(0, Math.min(originalWidth, x2));
y2 = Math.max(0, Math.min(originalHeight, y2));
detections.add(new Detection(x1, y1, x2, y2, confidence, classId));
}
return detections;
}
private float[] flatten(float[][][] tensor) {
float[] flat = new float[3 * 640 * 640];
int idx = 0;
for (int c = 0; c < 3; c++) {
for (int h = 0; h < 640; h++) {
for (int w = 0; w < 640; w++) {
flat[idx++] = tensor[c][h][w];
}
}
}
return flat;
}
private List<String> loadClasses(String path) throws IOException {
return java.nio.file.Files.readAllLines(java.nio.file.Paths.get(path));
}
static class Detection {
float x1, y1, x2, y2, confidence;
int classId;
Detection(float x1, float y1, float x2, float y2, float conf, int id) {
this.x1 = x1;
this.y1 = y1;
this.x2 = x2;
this.y2 = y2;
this.confidence = conf;
this.classId = id;
}
}
public static void main(String[] args) throws Exception {
YOLOv26Inference detector = new YOLOv26Inference(
"/Users/longjun/GeoServerProjects/greatmapserver-ai/gmserver-ai-mcp-client/src/test/java/models/yolo26n.onnx",
"/Users/longjun/GeoServerProjects/greatmapserver-ai/gmserver-ai-mcp-client/src/test/java/models/coco.names"
);
detector.detectAndDraw(
"/Users/longjun/GeoServerProjects/greatmapserver-ai/gmserver-ai-mcp-client/src/test/java/models/bus.jpg",
"/Users/longjun/GeoServerProjects/greatmapserver-ai/gmserver-ai-mcp-client/src/test/java/models/output.jpg"
);
}
}六、看运行结果output.jpg
