The cv::dnn module in OpenCV enables loading and running deep learning models in C++. It supports frameworks such as TensorFlow, Caffe, ONNX, and PyTorch (via ONNX export).

• Supports inference on CPU, OpenCL, and CUDA backends.
• Used for image classification, object detection, semantic segmentation, and more.

Steps to use DNN module:

1. Load the model using cv::dnn::readNet.
2. Convert input image to a blob using cv::dnn::blobFromImage.
3. Set input to the network via net.setInput().
4. Run inference using net.forward().
5. Post-process the results.

#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>

using namespace cv;
using namespace cv::dnn;

int main() {
    // Load model
    Net net = readNetFromONNX("model.onnx");

    // Read image
    Mat img = imread("example.jpg");

    // Convert image to blob
    Mat blob = blobFromImage(img, 1.0/255.0, Size(224,224), Scalar(), true, false);

    // Set blob as input
    net.setInput(blob);

    // Perform inference
    Mat output = net.forward();

    // Print top class index
    Point classIdPoint;
    double confidence;
    minMaxLoc(output.reshape(1, 1), 0, &confidence, 0, &classIdPoint);
    int classId = classIdPoint.x;

    std::cout << "Class: " << classId << ", Confidence: " << confidence << std::endl;
    return 0;
}
    

Commonly Supported Formats:

• .pb (TensorFlow)
• .onnx (ONNX format from PyTorch, TensorFlow, etc.)
• .caffemodel + .prototxt (Caffe)
• .t7 (Torch)

Note: Use readNetFromONNX, readNetFromTensorflow, readNetFromCaffe, and readNetFromTorch accordingly.

Available Backends:

• DNN_BACKEND_OPENCV
• DNN_BACKEND_INFERENCE_ENGINE (OpenVINO)
• DNN_BACKEND_CUDA

Available Targets:

• DNN_TARGET_CPU
• DNN_TARGET_OPENCL
• DNN_TARGET_CUDA

Net net = readNetFromONNX("model.onnx");
net.setPreferableBackend(DNN_BACKEND_CUDA);
net.setPreferableTarget(DNN_TARGET_CUDA);
    

blobFromImage Parameters:

blobFromImage(image, scaleFactor, size, mean, swapRB, crop, ddepth)
    

Note: Use swapRB=true for RGB models, and set mean subtraction for normalization like Scalar(104,117,123) for ImageNet models.

Postprocessing:

YOLO-style models require output parsing into bounding boxes, confidences, and class IDs, followed by non-maximum suppression.

// postprocess output of net.forward()
for (int i = 0; i < detections.rows; ++i) {
    float conf = detections.at<float>(i, 4);
    if (conf > threshold) {
        // Extract class scores and apply NMS
    }
}
    

Note: Use NMSBoxes() for final selection of bounding boxes.

• Use net.enableWinograd(false) for better accuracy if needed.
• Use net.setPreferableBackend() and setPreferableTarget() to optimize for GPU or CPU.
• Preprocess images in batches for performance using blobFromImages().

• TensorFlow Object Detection with OpenCV
• ONNX Model Zoo
• OpenCV Model Zoo with C++ examples