DNN Module

Deep Neural Networks inference with OpenCV’s DNN module

Overview

The DNN (Deep Neural Networks) module provides:

• Loading models from popular frameworks (TensorFlow, PyTorch, ONNX, Caffe, Darknet)
• Forward inference (no training)
• Multiple backend support (CPU, OpenCL, CUDA)
• Pre-trained model zoo

The DNN module is for inference only. For training, use frameworks like TensorFlow or PyTorch.

Quick Start

#include <opencv2/dnn.hpp>

using namespace cv::dnn;

// Load model
Net net = readNet("model.onnx");

// Prepare input
Mat blob = blobFromImage(img, 1.0/255, Size(224, 224), 
                         Scalar(), true, false);

// Set input
net.setInput(blob);

// Forward pass
Mat output = net.forward();

Net Class

Loading Models

// From ONNX
Net net = readNetFromONNX("model.onnx");

// From TensorFlow
Net net = readNetFromTensorflow("model.pb", "config.pbtxt");

// From Caffe  
Net net = readNetFromCaffe("deploy.prototxt", "model.caffemodel");

// From Darknet (YOLO)
Net net = readNetFromDarknet("yolov4.cfg", "yolov4.weights");

// From PyTorch (via ONNX)
Net net = readNetFromONNX("model.onnx");

// Auto-detect format
Net net = readNet("model.onnx");  // Detects format automatically

Setting Backend and Target

// CPU backend
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(DNN_TARGET_CPU);

// OpenCL (GPU)
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(DNN_TARGET_OPENCL);

// CUDA (NVIDIA GPU)
net.setPreferableBackend(DNN_BACKEND_CUDA);
net.setPreferableTarget(DNN_TARGET_CUDA);

// Intel OpenVINO
net.setPreferableBackend(DNN_BACKEND_INFERENCE_ENGINE);
net.setPreferableTarget(DNN_TARGET_CPU);

Inference

// Single output
net.setInput(blob, "input_name");
Mat output = net.forward("output_name");

// Multiple outputs
std::vector<String> outNames = {"output1", "output2"};
std::vector<Mat> outputs;
net.forward(outputs, outNames);

// All outputs
std::vector<String> outNames = net.getUnconnectedOutLayersNames();
std::vector<Mat> outputs;
net.forward(outputs, outNames);

Blob Preparation

blobFromImage

Mat blob = blobFromImage(
    image,              // Input image
    1.0/255.0,         // Scale factor
    Size(224, 224),    // Target size
    Scalar(0, 0, 0),   // Mean subtraction
    true,              // swapRB (BGR to RGB)
    false,             // crop
    CV_32F             // Output type
);

blobFromImages (Batch)

std::vector<Mat> images = {img1, img2, img3};
Mat blob = blobFromImages(images, 1.0/255.0, 
                          Size(224, 224), Scalar(), true);

Blob Format

Blobs use NCHW format:

• N: Batch size
• C: Channels
• H: Height
• W: Width

// Blob shape: [1, 3, 224, 224]
// 1 image, 3 channels (RGB), 224x224 pixels

Common Tasks

Image Classification

// Load model
Net net = readNet("mobilenet_v2.onnx");
net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(DNN_TARGET_CPU);

// Prepare input
Mat img = imread("image.jpg");
Mat blob = blobFromImage(img, 1.0/255, Size(224, 224), 
                         Scalar(), true, false);

// Inference
net.setInput(blob);
Mat prob = net.forward();

// Get top class
Point classIdPoint;
minMaxLoc(prob.reshape(1, 1), 0, 0, 0, &classIdPoint);
int classId = classIdPoint.x;

Object Detection (YOLO)

// Load YOLO
Net net = readNetFromDarknet("yolov4.cfg", "yolov4.weights");

// Prepare input
Mat blob = blobFromImage(img, 1/255.0, Size(416, 416), 
                         Scalar(), true, false);

// Forward
net.setInput(blob);
std::vector<Mat> outputs;
net.forward(outputs, net.getUnconnectedOutLayersNames());

// Process detections
for(Mat& output : outputs) {
    for(int i = 0; i < output.rows; i++) {
        float* data = output.ptr<float>(i);
        float confidence = data[4];
        
        if(confidence > 0.5) {
            int classId = max_element(data+5, data+output.cols) - (data+5);
            float x = data[0] * img.cols;
            float y = data[1] * img.rows;
            float w = data[2] * img.cols;
            float h = data[3] * img.rows;
            // Draw bounding box
        }
    }
}

Semantic Segmentation

Net net = readNet("fcn-resnet50.onnx");

Mat blob = blobFromImage(img, 1.0, Size(500, 500));
net.setInput(blob);
Mat score = net.forward();

// score shape: [1, num_classes, H, W]
// Get class per pixel
Mat classMap;
for(int h = 0; h < score.size[2]; h++) {
    for(int w = 0; w < score.size[3]; w++) {
        // Get class with max score
        // ...
    }
}

Model Zoo

OpenCV provides pre-trained models:

// Face detection
Net faceNet = readNet("opencv_face_detector.caffemodel",
                      "opencv_face_detector.prototxt");

// Age/Gender
Net ageNet = readNet("age_net.caffemodel", "age_deploy.prototxt");
Net genderNet = readNet("gender_net.caffemodel", "gender_deploy.prototxt");

// OpenPose (pose estimation)
Net poseNet = readNet("pose_iter_440000.caffemodel",
                      "pose_deploy_linevec.prototxt");

Performance Optimization

Backend Selection

CPU

net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(DNN_TARGET_CPU);

net.setPreferableBackend(DNN_BACKEND_CUDA);
net.setPreferableTarget(DNN_TARGET_CUDA);

net.setPreferableBackend(DNN_BACKEND_OPENCV);
net.setPreferableTarget(DNN_TARGET_OPENCL);

Input Size

// Smaller input = faster inference
Mat blob = blobFromImage(img, 1.0/255, 
                         Size(320, 320),  // Reduce from 640x640
                         Scalar(), true);

Batch Processing

// Process multiple images at once
std::vector<Mat> images;
for(int i = 0; i < 4; i++) {
    images.push_back(imread(files[i]));
}

Mat blob = blobFromImages(images, 1.0/255, Size(224, 224));
net.setInput(blob);
Mat output = net.forward();  // Batch inference

Best Practices

Use ONNX Format

ONNX provides best compatibility across frameworks

Enable GPU

Use CUDA backend for 5-10x speedup on NVIDIA GPUs

Optimize Input Size

Smaller inputs trade accuracy for speed

Batch When Possible

Batch processing improves GPU utilization

Troubleshooting

Model Loading Issues

if(net.empty()) {
    std::cerr << "Failed to load model\n";
    return -1;
}

Check Backend Support

auto backends = getAvailableBackends();
for(auto& backend : backends) {
    std::cout << "Backend: " << backend.first 
              << ", Target: " << backend.second << "\n";
}

Enable Diagnostic Mode

enableModelDiagnostics(true);
Net net = readNet("model.onnx");  // Verbose loading

See Also

• DNN Network API - Detailed Net class reference
• DNN Layers - Layer types and custom layers
• DNN Inference - Advanced inference techniques
• Object Detection Module - Classic detection methods