Deep Learning with OpenCV DNN Module

Learn how to use OpenCV’s DNN (Deep Neural Networks) module to load and run pre-trained models for object detection, classification, and more.

Introduction to OpenCV DNN

OpenCV’s DNN module allows you to:

Load models from TensorFlow, PyTorch, Caffe, ONNX, and Darknet
Run inference without installing deep learning frameworks
Deploy on CPU, GPU (CUDA), or OpenVINO backends
Use pre-trained models for various tasks

Supported Frameworks

Supported Model Formats

ONNX (.onnx) - Universal format, recommended
TensorFlow (.pb, .pbtxt)
PyTorch (via ONNX export)
Caffe (.caffemodel, .prototxt)
Darknet (.weights, .cfg) - YOLO models
TensorFlow Lite (.tflite)

Loading and Running Models

Basic Model Loading

Python
C++

import cv2 as cv
import numpy as np

# Load a model (example: ONNX format)
net = cv.dnn.readNet('model.onnx')

# Or load specific formats:
# net = cv.dnn.readNetFromTensorflow('model.pb', 'model.pbtxt')
# net = cv.dnn.readNetFromCaffe('deploy.prototxt', 'model.caffemodel')
# net = cv.dnn.readNetFromDarknet('yolov3.cfg', 'yolov3.weights')

# Set computation backend and target
net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CPU)

# For GPU acceleration:
# net.setPreferableBackend(cv.dnn.DNN_BACKEND_CUDA)
# net.setPreferableTarget(cv.dnn.DNN_TARGET_CUDA)

#include <opencv2/dnn.hpp>
#include <opencv2/opencv.hpp>
using namespace cv;
using namespace cv::dnn;

int main() {
    // Load model
    Net net = readNet("model.onnx");
    
    // Or specific formats:
    // Net net = readNetFromTensorflow("model.pb", "model.pbtxt");
    // Net net = readNetFromCaffe("deploy.prototxt", "model.caffemodel");
    // Net net = readNetFromDarknet("yolov3.cfg", "yolov3.weights");
    
    // Set backend and target
    net.setPreferableBackend(DNN_BACKEND_OPENCV);
    net.setPreferableTarget(DNN_TARGET_CPU);
    
    // For GPU:
    // net.setPreferableBackend(DNN_BACKEND_CUDA);
    // net.setPreferableTarget(DNN_TARGET_CUDA);
    
    return 0;
}

YOLO Object Detection

YOLO (You Only Look Once) is a popular real-time object detection system.

YOLOv3/YOLOv4 Detection

Python
C++

import cv2 as cv
import numpy as np

# Load YOLO network
net = cv.dnn.readNetFromDarknet('yolov4.cfg', 'yolov4.weights')
net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CPU)

# Load class names
with open('coco.names', 'r') as f:
    classes = [line.strip() for line in f.readlines()]

# Load image
img = cv.imread('street.jpg')
height, width = img.shape[:2]

# Create blob from image
blob = cv.dnn.blobFromImage(img, 1/255.0, (416, 416), 
                           swapRB=True, crop=False)

# Set input and run forward pass
net.setInput(blob)

# Get output layer names
output_layers = net.getUnconnectedOutLayersNames()

# Forward pass
outputs = net.forward(output_layers)

# Process detections
boxes = []
confidences = []
class_ids = []

for output in outputs:
    for detection in output:
        scores = detection[5:]
        class_id = np.argmax(scores)
        confidence = scores[class_id]
        
        if confidence > 0.5:
            # Scale bounding box back to image size
            center_x = int(detection[0] * width)
            center_y = int(detection[1] * height)
            w = int(detection[2] * width)
            h = int(detection[3] * height)
            
            # Get top-left corner
            x = int(center_x - w / 2)
            y = int(center_y - h / 2)
            
            boxes.append([x, y, w, h])
            confidences.append(float(confidence))
            class_ids.append(class_id)

# Apply Non-Maximum Suppression
indices = cv.dnn.NMSBoxes(boxes, confidences, 0.5, 0.4)

# Draw detections
if len(indices) > 0:
    for i in indices.flatten():
        x, y, w, h = boxes[i]
        label = f"{classes[class_ids[i]]}: {confidences[i]:.2f}"
        color = (0, 255, 0)
        
        cv.rectangle(img, (x, y), (x+w, y+h), color, 2)
        cv.putText(img, label, (x, y-10), 
                  cv.FONT_HERSHEY_SIMPLEX, 0.5, color, 2)

cv.imshow('YOLO Detection', img)
cv.waitKey(0)

#include <opencv2/dnn.hpp>
#include <opencv2/opencv.hpp>
#include <fstream>
using namespace cv;
using namespace cv::dnn;
using namespace std;

int main() {
    // Load network
    Net net = readNetFromDarknet("yolov4.cfg", "yolov4.weights");
    net.setPreferableBackend(DNN_BACKEND_OPENCV);
    net.setPreferableTarget(DNN_TARGET_CPU);
    
    // Load class names
    vector<string> classes;
    ifstream ifs("coco.names");
    string line;
    while(getline(ifs, line)) classes.push_back(line);
    
    // Load image
    Mat img = imread("street.jpg");
    
    // Create blob
    Mat blob;
    blobFromImage(img, blob, 1/255.0, Size(416, 416), 
                 Scalar(), true, false);
    
    net.setInput(blob);
    
    // Get output layers
    vector<String> outNames = net.getUnconnectedOutLayersNames();
    vector<Mat> outs;
    net.forward(outs, outNames);
    
    // Process detections
    vector<int> classIds;
    vector<float> confidences;
    vector<Rect> boxes;
    
    for(size_t i = 0; i < outs.size(); ++i) {
        float* data = (float*)outs[i].data;
        for(int j = 0; j < outs[i].rows; ++j, data += outs[i].cols) {
            Mat scores = outs[i].row(j).colRange(5, outs[i].cols);
            Point classIdPoint;
            double confidence;
            minMaxLoc(scores, 0, &confidence, 0, &classIdPoint);
            
            if(confidence > 0.5) {
                int centerX = (int)(data[0] * img.cols);
                int centerY = (int)(data[1] * img.rows);
                int width = (int)(data[2] * img.cols);
                int height = (int)(data[3] * img.rows);
                int left = centerX - width / 2;
                int top = centerY - height / 2;
                
                classIds.push_back(classIdPoint.x);
                confidences.push_back((float)confidence);
                boxes.push_back(Rect(left, top, width, height));
            }
        }
    }
    
    // NMS
    vector<int> indices;
    NMSBoxes(boxes, confidences, 0.5, 0.4, indices);
    
    // Draw
    for(size_t i = 0; i < indices.size(); ++i) {
        int idx = indices[i];
        Rect box = boxes[idx];
        rectangle(img, box, Scalar(0, 255, 0), 2);
        
        string label = classes[classIds[idx]] + ": " + 
                      format("%.2f", confidences[idx]);
        putText(img, label, Point(box.x, box.y - 10),
               FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0), 2);
    }
    
    imshow("YOLO Detection", img);
    waitKey(0);
    return 0;
}

YOLOv8 with ONNX

Modern YOLO versions export to ONNX format:

Python

import cv2 as cv
import numpy as np

# Load YOLOv8 model (ONNX format)
net = cv.dnn.readNetFromONNX('yolov8n.onnx')
net.setPreferableBackend(cv.dnn.DNN_BACKEND_OPENCV)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CPU)

# Load image
img = cv.imread('image.jpg')
original_height, original_width = img.shape[:2]

# Preprocess
input_size = 640
blob = cv.dnn.blobFromImage(img, 1/255.0, (input_size, input_size),
                           swapRB=True, crop=False)

# Run inference
net.setInput(blob)
output = net.forward()

# YOLOv8 outputs shape: [1, 84, 8400] for COCO
# Format: [x, y, w, h, class_scores...]
output = output[0].transpose()  # [8400, 84]

# Process detections
boxes = []
confidences = []
class_ids = []

img_height, img_width = img.shape[:2]
x_scale = img_width / input_size
y_scale = img_height / input_size

for detection in output:
    # Extract box coordinates
    x, y, w, h = detection[:4]
    
    # Get class scores and find max
    class_scores = detection[4:]
    class_id = np.argmax(class_scores)
    confidence = class_scores[class_id]
    
    if confidence > 0.5:
        # Scale back to original image
        x = int((x - w/2) * x_scale)
        y = int((y - h/2) * y_scale)
        w = int(w * x_scale)
        h = int(h * y_scale)
        
        boxes.append([x, y, w, h])
        confidences.append(float(confidence))
        class_ids.append(class_id)

# Apply NMS
indices = cv.dnn.NMSBoxes(boxes, confidences, 0.5, 0.4)

# Draw results
for i in indices.flatten():
    x, y, w, h = boxes[i]
    cv.rectangle(img, (x, y), (x+w, y+h), (0, 255, 0), 2)
    label = f"Class {class_ids[i]}: {confidences[i]:.2f}"
    cv.putText(img, label, (x, y-10),
              cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

cv.imshow('YOLOv8 Detection', img)
cv.waitKey(0)

SSD Object Detection

SSD (Single Shot MultiBox Detector) for faster detection:

Python
C++

import cv2 as cv
import numpy as np

# Load MobileNet-SSD model
net = cv.dnn.readNetFromCaffe(
    'MobileNetSSD_deploy.prototxt',
    'MobileNetSSD_deploy.caffemodel'
)

# COCO class names
classes = ["background", "aeroplane", "bicycle", "bird", "boat",
           "bottle", "bus", "car", "cat", "chair", "cow",
           "diningtable", "dog", "horse", "motorbike", "person",
           "pottedplant", "sheep", "sofa", "train", "tvmonitor"]

# Load image
img = cv.imread('image.jpg')
height, width = img.shape[:2]

# Prepare input
blob = cv.dnn.blobFromImage(img, 0.007843, (300, 300), 127.5)

# Run detection
net.setInput(blob)
detections = net.forward()

# Process detections
for i in range(detections.shape[2]):
    confidence = detections[0, 0, i, 2]
    
    if confidence > 0.5:
        class_id = int(detections[0, 0, i, 1])
        
        # Get box coordinates
        box = detections[0, 0, i, 3:7] * np.array([width, height, width, height])
        (x1, y1, x2, y2) = box.astype("int")
        
        # Draw detection
        label = f"{classes[class_id]}: {confidence:.2f}"
        cv.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv.putText(img, label, (x1, y1-10),
                  cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

cv.imshow('SSD Detection', img)
cv.waitKey(0)

#include <opencv2/dnn.hpp>
#include <opencv2/opencv.hpp>
using namespace cv;
using namespace cv::dnn;

int main() {
    Net net = readNetFromCaffe(
        "MobileNetSSD_deploy.prototxt",
        "MobileNetSSD_deploy.caffemodel"
    );
    
    Mat img = imread("image.jpg");
    
    Mat blob;
    blobFromImage(img, blob, 0.007843, Size(300, 300), 
                 Scalar(127.5, 127.5, 127.5));
    
    net.setInput(blob);
    Mat detections = net.forward();
    
    Mat detectionMat(detections.size[2], detections.size[3], 
                    CV_32F, detections.ptr<float>());
    
    for(int i = 0; i < detectionMat.rows; i++) {
        float confidence = detectionMat.at<float>(i, 2);
        
        if(confidence > 0.5) {
            int x1 = detectionMat.at<float>(i, 3) * img.cols;
            int y1 = detectionMat.at<float>(i, 4) * img.rows;
            int x2 = detectionMat.at<float>(i, 5) * img.cols;
            int y2 = detectionMat.at<float>(i, 6) * img.rows;
            
            rectangle(img, Point(x1, y1), Point(x2, y2),
                     Scalar(0, 255, 0), 2);
        }
    }
    
    imshow("SSD Detection", img);
    waitKey(0);
    return 0;
}

Image Classification

Python

import cv2 as cv
import numpy as np

# Load ResNet model
net = cv.dnn.readNetFromCaffe(
    'ResNet-50-deploy.prototxt',
    'ResNet-50-model.caffemodel'
)

# Load ImageNet class labels
with open('imagenet_classes.txt', 'r') as f:
    classes = [line.strip() for line in f.readlines()]

# Load and preprocess image
img = cv.imread('dog.jpg')

# Create blob (ResNet expects 224x224 input)
blob = cv.dnn.blobFromImage(img, 1.0, (224, 224),
                           (104, 117, 123), swapRB=False, crop=False)

# Run inference
net.setInput(blob)
predictions = net.forward()

# Get top 5 predictions
top5_indices = np.argsort(predictions[0])[::-1][:5]

print("Top 5 predictions:")
for i, idx in enumerate(top5_indices):
    label = classes[idx]
    confidence = predictions[0][idx]
    print(f"{i+1}. {label}: {confidence*100:.2f}%")

# Display result
top_label = classes[top5_indices[0]]
cv.putText(img, top_label, (10, 30),
          cv.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
cv.imshow('Classification', img)
cv.waitKey(0)

Face Detection with DNN

Deep learning-based face detection (more accurate than Haar cascades):

Python

import cv2 as cv

# Load face detection model
net = cv.dnn.readNetFromCaffe(
    'deploy.prototxt',
    'res10_300x300_ssd_iter_140000.caffemodel'
)

# Load image
img = cv.imread('faces.jpg')
height, width = img.shape[:2]

# Preprocess
blob = cv.dnn.blobFromImage(img, 1.0, (300, 300),
                           (104.0, 177.0, 123.0))

# Detect faces
net.setInput(blob)
detections = net.forward()

# Draw detections
for i in range(detections.shape[2]):
    confidence = detections[0, 0, i, 2]
    
    if confidence > 0.5:
        box = detections[0, 0, i, 3:7] * np.array([width, height, width, height])
        (x1, y1, x2, y2) = box.astype("int")
        
        # Draw box and confidence
        text = f"{confidence*100:.2f}%"
        cv.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
        cv.putText(img, text, (x1, y1-10),
                  cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

cv.imshow('Face Detection', img)
cv.waitKey(0)

Video Processing with DNN

Python

import cv2 as cv
import time

# Load model
net = cv.dnn.readNetFromCaffe(
    'MobileNetSSD_deploy.prototxt',
    'MobileNetSSD_deploy.caffemodel'
)

# Open video
cap = cv.VideoCapture('video.mp4')

while True:
    ret, frame = cap.read()
    if not ret:
        break
    
    height, width = frame.shape[:2]
    
    # Prepare input
    blob = cv.dnn.blobFromImage(frame, 0.007843, (300, 300), 127.5)
    
    # Measure inference time
    start = time.time()
    net.setInput(blob)
    detections = net.forward()
    end = time.time()
    
    # Process detections
    for i in range(detections.shape[2]):
        confidence = detections[0, 0, i, 2]
        
        if confidence > 0.5:
            box = detections[0, 0, i, 3:7] * np.array([width, height, width, height])
            (x1, y1, x2, y2) = box.astype("int")
            cv.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
    
    # Display FPS
    fps = 1 / (end - start)
    cv.putText(frame, f'FPS: {fps:.1f}', (10, 30),
              cv.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
    
    cv.imshow('Detection', frame)
    
    if cv.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv.destroyAllWindows()

Performance Optimization

Using GPU Acceleration

import cv2 as cv

net = cv.dnn.readNet('model.onnx')

# CUDA backend (requires OpenCV built with CUDA)
net.setPreferableBackend(cv.dnn.DNN_BACKEND_CUDA)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CUDA)

# Or CUDA with FP16 (faster, slightly less accurate)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CUDA_FP16)

OpenVINO Backend

# Intel OpenVINO for optimized inference on Intel hardware
net.setPreferableBackend(cv.dnn.DNN_BACKEND_INFERENCE_ENGINE)
net.setPreferableTarget(cv.dnn.DNN_TARGET_CPU)

# Or use Intel GPU
net.setPreferableTarget(cv.dnn.DNN_TARGET_OPENCL)

Batch Processing

# Process multiple images at once
images = [img1, img2, img3, img4]

# Create batch blob
blob = cv.dnn.blobFromImages(images, 1/255.0, (640, 640))

net.setInput(blob)
outputs = net.forward()

Backend and target options:

DNN_BACKEND_OPENCV + DNN_TARGET_CPU: Default, works everywhere
DNN_BACKEND_CUDA + DNN_TARGET_CUDA: NVIDIA GPU acceleration
DNN_BACKEND_INFERENCE_ENGINE + DNN_TARGET_CPU: Intel OpenVINO
DNN_TARGET_OPENCL: OpenCL acceleration
DNN_TARGET_CUDA_FP16: Half-precision for faster inference

Common issues:

Model input size must match the size used during training
Check if the model expects RGB or BGR input (use swapRB parameter)
Normalize input values correctly (typically 0-1 or mean subtraction)
Ensure OpenCV is built with the desired backend support

Downloading Pre-trained Models

OpenCV provides scripts to download common models:

# Download YOLOv3
python opencv/samples/dnn/download_models.py --name yolo

# Download all models
python opencv/samples/dnn/download_models.py --all

Common model sources:

Next Steps

Explore the OpenCV Model Zoo for more pre-trained models
Learn about Camera Calibration for 3D vision tasks
Combine with Video Processing for real-time applications

Documentation Index

​Deep Learning with OpenCV DNN Module

​Introduction to OpenCV DNN

​Supported Frameworks

​Loading and Running Models

​Basic Model Loading

​YOLO Object Detection

​YOLOv3/YOLOv4 Detection

​YOLOv8 with ONNX

​SSD Object Detection

​Image Classification

​Face Detection with DNN

​Video Processing with DNN

​Performance Optimization

​Downloading Pre-trained Models

​Next Steps