一、引言

赛昉科技提供了许多参考性文档，其中不乏使用昉·星光 2测试在AI图像处理的应用场景。上期咱分享了二维码检测与解码效果，只要摄像头采集图像够清晰，扫描二维码与解析效率还是挺好的。今儿来分享一下，基于上期环境，再对摄像头应用进行展开，评估其在人脸识别、物体识别以及图像边缘检测上的应用。

二、人脸识别

（1）硬件搭建

准备一本封面印有人脸头像的杂志，以便后续测试检测效果。

固定uvc摄像机，然后调好焦距，使其图像采集清晰。

（2）代码实现

与上期的二维码检测实验一样，基于OpenCV模型去实现人脸图像检测、轮廓标注、置信度和识别。在人脸图像检测的同时，显示检测的置信度，置信度小于50%过滤掉。，因此需用到深度学习的检测模型，github上有提供开源的OpenCV DNN或FaceNet模型，可满足置信度显示需求。
准备预训练模型文件
通过github上获取“deploy.prototxt.txt”与“res10_300x300_ssd_iter_140000.caffemodel”文件。
*附件：deploy.prototxt.zip
*附件：res10_300x300_ssd_iter_140000.zip
编写“face_detection.py”源代码，内容如下：

import cv2
import numpy as np

net = cv2.dnn.readNetFromCaffe("deploy.prototxt.txt", "res10_300x300_ssd_iter_140000.caffemodel")

cap = cv2.VideoCapture(4, cv2.CAP_V4L2)

while True:
    ret, frame = cap.read()
    if not ret:
        break

    (h, w) = frame.shape[:2]
    blob = cv2.dnn.blobFromImage(cv2.resize(frame, (300, 300)), 1.0,
                                (300, 300), (104.0, 177.0, 123.0))
    net.setInput(blob)
    detections = net.forward()

    for i in range(detections.shape[2]):
        confidence = detections[0, 0, i, 2]
      
        if confidence < 0.5:
            continue
        box = detections[0, 0, i, 3:7] * np.array([w, h, w, h])
        (startX, startY, endX, endY) = box.astype("int")

        cv2.rectangle(frame, (startX, startY), (endX, endY), (0, 255, 0), 2)
        text = f"Face: {confidence * 100:.2f}%"
        cv2.putText(frame, text, (startX, startY - 10),
        cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)

    cv2.imshow("Face Detection with Confidence", frame)
    key = cv2.waitKey(1) & 0xFF

    # if the `q` key was pressed, break from the loop
    if key == ord("q"):
        break

cap.release()
cv2.destroyAllWindows()

将编写好的“face_detection.py”源文件与上述下载好的预训练模型文件放在同一文件夹下。

桌面方式登录后，然后终端中进到该文件夹下，普通用户身份执行“ python3 face_detection.py”，呈现的检测效果见底部视频。

三、物体识别

因为之前的二维码检测功能基本已实现，因此环境这块，相应的依赖软件包都已安装好了，这里调用YOLO-V3模型进行通用物体的识别。
基于yolo-v3模型的应用“object_detection.cpp”源代码如下：

#include <fstream>
#include <sstream>

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

#if defined(CV_CXX11) && defined(HAVE_THREADS)
#define USE_THREADS 1
#endif

#ifdef USE_THREADS
#include <mutex>
#include <thread>
#include <queue>
#endif

#include "common.hpp"

std::string keys =
    "{ help  h     | | Print help message. }"
    "{ [url=home.php?mod=space&uid=204129]@Alias[/url]      | | An alias name of model to extract preprocessing parameters from models.yml file. }"
    "{ zoo         | models.yml | An optional path to file with preprocessing parameters }"
    "{ device      |  0 | camera device number. }"
    "{ input i     | | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ framework f | | Optional name of an origin framework of the model. Detect it automatically if it does not set. }"
    "{ classes     | | Optional path to a text file with names of classes to label detected objects. }"
    "{ thr         | .5 | Confidence threshold. }"
    "{ nms         | .4 | Non-maximum suppression threshold. }"
    "{ backend     |  0 | Choose one of computation backends: "
                         "0: automatically (by default), "
                         "1: Halide language (http://halide-lang.org/), "
                         "2: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), "
                         "3: OpenCV implementation, "
                         "4: VKCOM, "
                         "5: CUDA }"
    "{ target      | 0 | Choose one of target computation devices: "
                         "0: CPU target (by default), "
                         "1: OpenCL, "
                         "2: OpenCL fp16 (half-float precision), "
                         "3: VPU, "
                         "4: Vulkan, "
                         "6: CUDA, "
                         "7: CUDA fp16 (half-float preprocess) }"
    "{ async       | 0 | Number of asynchronous forwards at the same time. "
                        "Choose 0 for synchronous mode }";

using namespace cv;
using namespace dnn;

float confThreshold, nmsThreshold;
std::vector<std::string> classes;

inline void preprocess(const Mat& frame, Net& net, Size inpSize, float scale,
                       const Scalar& mean, bool swapRB);

void postprocess(Mat& frame, const std::vector<Mat>& out, Net& net, int backend);

void drawPred(int classId, float conf, int left, int top, int right, int bottom, Mat& frame);

void callback(int pos, void* userdata);

#ifdef USE_THREADS
template <typename T>
class QueueFPS : public std::queue<T>
{
public:
    QueueFPS() : counter(0) {}

    void push(const T& entry)
    {
        std::lock_guard<std::mutex> lock(mutex);

        std::queue<T>::push(entry);
        counter += 1;
        if (counter == 1)
        {
            // Start counting from a second frame (warmup).
            tm.reset();
            tm.start();
        }
    }

    T get()
    {
        std::lock_guard<std::mutex> lock(mutex);
        T entry = this->front();
        this->pop();
        return entry;
    }

    float getFPS()
    {
        tm.stop();
        double fps = counter / tm.getTimeSec();
        tm.start();
        return static_cast<float>(fps);
    }

    void clear()
    {
        std::lock_guard<std::mutex> lock(mutex);
        while (!this->empty())
            this->pop();
    }

    unsigned int counter;

private:
    TickMeter tm;
    std::mutex mutex;
};
#endif  // USE_THREADS

int main(int argc, char** argv)
{
    CommandLineParser parser(argc, argv, keys);

    const std::string modelName = parser.get<String>("@alias");
    const std::string zooFile = parser.get<String>("zoo");

    keys += genPreprocArguments(modelName, zooFile);

    parser = CommandLineParser(argc, argv, keys);
    parser.about("Use this script to run object detection deep learning networks using OpenCV.");
    if (argc == 1 || parser.has("help"))
    {
        parser.printMessage();
        return 0;
    }

    confThreshold = parser.get<float>("thr");
    nmsThreshold = parser.get<float>("nms");
    float scale = parser.get<float>("scale");
    Scalar mean = parser.get<Scalar>("mean");
    bool swapRB = parser.get<bool>("rgb");
    int inpWidth = parser.get<int>("width");
    int inpHeight = parser.get<int>("height");
    size_t asyncNumReq = parser.get<int>("async");
    CV_Assert(parser.has("model"));
    std::string modelPath = findFile(parser.get<String>("model"));
    std::string configPath = findFile(parser.get<String>("config"));

    // Open file with classes names.
    if (parser.has("classes"))
    {
        std::string file = parser.get<String>("classes");
        std::ifstream ifs(file.c_str());
        if (!ifs.is_open())
            CV_Error(Error::StsError, "File " + file + " not found");
        std::string line;
        while (std::getline(ifs, line))
        {
            classes.push_back(line);
        }
    }

    // Load a model.
    Net net = readNet(modelPath, configPath, parser.get<String>("framework"));
    int backend = parser.get<int>("backend");
    net.setPreferableBackend(backend);
    net.setPreferableTarget(parser.get<int>("target"));
    std::vector<String> outNames = net.getUnconnectedOutLayersNames();

    // Create a window
    static const std::string kWinName = "Deep learning object detection in OpenCV";
    namedWindow(kWinName, WINDOW_NORMAL);
    int initialConf = (int)(confThreshold * 100);
    createTrackbar("Confidence threshold, %", kWinName, &initialConf, 99, callback);

    // Open a video file or an image file or a camera stream.
    VideoCapture cap;
    if (parser.has("input"))
        cap.open(parser.get<String>("input"));
    else
        cap.open(parser.get<int>("device"));

#ifdef USE_THREADS
    bool process = true;

    // Frames capturing thread
    QueueFPS<Mat> framesQueue;
    std::thread framesThread([&](){
        Mat frame;
        while (process)
        {
            cap >> frame;
            if (!frame.empty())
                framesQueue.push(frame.clone());
            else
                break;
        }
    });

    // Frames processing thread
    QueueFPS<Mat> processedFramesQueue;
    QueueFPS<std::vector<Mat> > predictionsQueue;
    std::thread processingThread([&](){
        std::queue<AsyncArray> futureOutputs;
        Mat blob;
        while (process)
        {
            // Get a next frame
            Mat frame;
            {
                if (!framesQueue.empty())
                {
                    frame = framesQueue.get();
                    if (asyncNumReq)
                    {
                        if (futureOutputs.size() == asyncNumReq)
                            frame = Mat();
                    }
                    else
                        framesQueue.clear();  // Skip the rest of frames
                }
            }

            // Process the frame
            if (!frame.empty())
            {
                preprocess(frame, net, Size(inpWidth, inpHeight), scale, mean, swapRB);
                processedFramesQueue.push(frame);

                if (asyncNumReq)
                {
                    futureOutputs.push(net.forwardAsync());
                }
                else
                {
                    std::vector<Mat> outs;
                    net.forward(outs, outNames);
                    predictionsQueue.push(outs);
                }
            }

            while (!futureOutputs.empty() &&
                   futureOutputs.front().wait_for(std::chrono::seconds(0)))
            {
                AsyncArray async_out = futureOutputs.front();
                futureOutputs.pop();
                Mat out;
                async_out.get(out);
                predictionsQueue.push({out});
            }
        }
    });

    // Postprocessing and rendering loop
    while (waitKey(1) < 0)
    {
        if (predictionsQueue.empty())
            continue;

        std::vector<Mat> outs = predictionsQueue.get();
        Mat frame = processedFramesQueue.get();

        postprocess(frame, outs, net, backend);

        if (predictionsQueue.counter > 1)
        {
            std::string label = format("Camera: %.2f FPS", framesQueue.getFPS());
            putText(frame, label, Point(0, 15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));

            label = format("Network: %.2f FPS", predictionsQueue.getFPS());
            putText(frame, label, Point(0, 30), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));

            label = format("Skipped frames: %d", framesQueue.counter - predictionsQueue.counter);
            putText(frame, label, Point(0, 45), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));
        }
        imshow(kWinName, frame);
    }

    process = false;
    framesThread.join();
    processingThread.join();

#else  // USE_THREADS
    if (asyncNumReq)
        CV_Error(Error::StsNotImplemented, "Asynchronous forward is supported only with Inference Engine backend.");

    // Process frames.
    Mat frame, blob;
    while (waitKey(1) < 0)
    {
        cap >> frame;
        if (frame.empty())
        {
            waitKey();
            break;
        }

        preprocess(frame, net, Size(inpWidth, inpHeight), scale, mean, swapRB);

        std::vector<Mat> outs;
        net.forward(outs, outNames);

        postprocess(frame, outs, net, backend);

        // Put efficiency information.
        std::vector<double> layersTimes;
        double freq = getTickFrequency() / 1000;
        double t = net.getPerfProfile(layersTimes) / freq;
        std::string label = format("Inference time: %.2f ms", t);
        putText(frame, label, Point(0, 15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));

        imshow(kWinName, frame);
    }
#endif  // USE_THREADS
    return 0;
}

inline void preprocess(const Mat& frame, Net& net, Size inpSize, float scale,
                       const Scalar& mean, bool swapRB)
{
    static Mat blob;
    // Create a 4D blob from a frame.
    if (inpSize.width <= 0) inpSize.width = frame.cols;
    if (inpSize.height <= 0) inpSize.height = frame.rows;
    blobFromImage(frame, blob, 1.0, inpSize, Scalar(), swapRB, false, CV_8U);

    // Run a model.
    net.setInput(blob, "", scale, mean);
    if (net.getLayer(0)->outputNameToIndex("im_info") != -1)  // Faster-RCNN or R-FCN
    {
        resize(frame, frame, inpSize);
        Mat imInfo = (Mat_<float>(1, 3) << inpSize.height, inpSize.width, 1.6f);
        net.setInput(imInfo, "im_info");
    }
}

void postprocess(Mat& frame, const std::vector<Mat>& outs, Net& net, int backend)
{
    static std::vector<int> outLayers = net.getUnconnectedOutLayers();
    static std::string outLayerType = net.getLayer(outLayers[0])->type;

    std::vector<int> classIds;
    std::vector<float> confidences;
    std::vector<Rect> boxes;
    if (outLayerType == "DetectionOutput")
    {
        // Network produces output blob with a shape 1x1xNx7 where N is a number of
        // detections and an every detection is a vector of values
        // [batchId, classId, confidence, left, top, right, bottom]
        CV_Assert(outs.size() > 0);
        for (size_t k = 0; k < outs.size(); k++)
        {
            float* data = (float*)outs[k].data;
            for (size_t i = 0; i < outs[k].total(); i += 7)
            {
                float confidence = data[i + 2];
                if (confidence > confThreshold)
                {
                    int left   = (int)data[i + 3];
                    int top    = (int)data[i + 4];
                    int right  = (int)data[i + 5];
                    int bottom = (int)data[i + 6];
                    int width  = right - left + 1;
                    int height = bottom - top + 1;
                    if (width <= 2 || height <= 2)
                    {
                        left   = (int)(data[i + 3] * frame.cols);
                        top    = (int)(data[i + 4] * frame.rows);
                        right  = (int)(data[i + 5] * frame.cols);
                        bottom = (int)(data[i + 6] * frame.rows);
                        width  = right - left + 1;
                        height = bottom - top + 1;
                    }
                    classIds.push_back((int)(data[i + 1]) - 1);  // Skip 0th background class id.
                    boxes.push_back(Rect(left, top, width, height));
                    confidences.push_back(confidence);
                }
            }
        }
    }
    else if (outLayerType == "Region")
    {
        for (size_t i = 0; i < outs.size(); ++i)
        {
            // Network produces output blob with a shape NxC where N is a number of
            // detected objects and C is a number of classes + 4 where the first 4
            // numbers are [center_x, center_y, width, height]
            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 > confThreshold)
                {
                    int centerX = (int)(data[0] * frame.cols);
                    int centerY = (int)(data[1] * frame.rows);
                    int width = (int)(data[2] * frame.cols);
                    int height = (int)(data[3] * frame.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));
                }
            }
        }
    }
    else
        CV_Error(Error::StsNotImplemented, "Unknown output layer type: " + outLayerType);

    // NMS is used inside Region layer only on DNN_BACKEND_OPENCV for another backends we need NMS in sample
    // or NMS is required if number of outputs > 1
    if (outLayers.size() > 1 || (outLayerType == "Region" && backend != DNN_BACKEND_OPENCV))
    {
        std::map<int, std::vector<size_t> > class2indices;
        for (size_t i = 0; i < classIds.size(); i++)
        {
            if (confidences[i] >= confThreshold)
            {
                class2indices[classIds[i]].push_back(i);
            }
        }
        std::vector<Rect> nmsBoxes;
        std::vector<float> nmsConfidences;
        std::vector<int> nmsClassIds;
        for (std::map<int, std::vector<size_t> >::iterator it = class2indices.begin(); it != class2indices.end(); ++it)
        {
            std::vector<Rect> localBoxes;
            std::vector<float> localConfidences;
            std::vector<size_t> classIndices = it->second;
            for (size_t i = 0; i < classIndices.size(); i++)
            {
                localBoxes.push_back(boxes[classIndices[i]]);
                localConfidences.push_back(confidences[classIndices[i]]);
            }
            std::vector<int> nmsIndices;
            NMSBoxes(localBoxes, localConfidences, confThreshold, nmsThreshold, nmsIndices);
            for (size_t i = 0; i < nmsIndices.size(); i++)
            {
                size_t idx = nmsIndices[i];
                nmsBoxes.push_back(localBoxes[idx]);
                nmsConfidences.push_back(localConfidences[idx]);
                nmsClassIds.push_back(it->first);
            }
        }
        boxes = nmsBoxes;
        classIds = nmsClassIds;
        confidences = nmsConfidences;
    }

    for (size_t idx = 0; idx < boxes.size(); ++idx)
    {
        Rect box = boxes[idx];
        drawPred(classIds[idx], confidences[idx], box.x, box.y,
                 box.x + box.width, box.y + box.height, frame);
    }
}

void drawPred(int classId, float conf, int left, int top, int right, int bottom, Mat& frame)
{
    rectangle(frame, Point(left, top), Point(right, bottom), Scalar(0, 255, 0));

    std::string label = format("%.2f", conf);
    if (!classes.empty())
    {
        CV_Assert(classId < (int)classes.size());
        label = classes[classId] + ": " + label;
    }

    int baseLine;
    Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);

    top = max(top, labelSize.height);
    rectangle(frame, Point(left, top - labelSize.height),
              Point(left + labelSize.width, top + baseLine), Scalar::all(255), FILLED);
    putText(frame, label, Point(left, top), FONT_HERSHEY_SIMPLEX, 0.5, Scalar());
}

void callback(int pos, void*)
{
    confThreshold = pos * 0.01f;
}

桌面方式登录后，终端中user身份直接运行“v4l2-ctl --list-device”指令，确认USB摄像机的ID号。然后再运行“example_dnn_object_detection --config=/usr/share/opencv4/yolo-v3/yolov3-tiny.cfg --model=/usr/share/opencv4/yolo-v3/yolov3-tiny.weights --classes=/usr/share/opencv4/yolo-v3/classes.txt --width=416--height=416--scale=0.00392--rgb --target=1--device=4”

四、图像边缘检测

这里调用OpenCV官方提供的“edge.py”源文件，内容如下：

#!/usr/bin/env python

'''
This sample demonstrates Canny edge detection.

Usage:
  edge.py [<video source>]

  Trackbars control edge thresholds.

'''

# Python 2/3 compatibility
from __future__ import print_function

import cv2 as cv
import numpy as np

# relative module
import video

# built-in module
import sys


def main():
    try:
        fn = sys.argv[1]
        cam = int(fn)
    except:
        fn = 0

    def nothing(*arg):
        pass

    cv.namedWindow('edge', 0)
    cv.resizeWindow('edge', 640, 480)
    cv.createTrackbar('thrs1', 'edge', 2000, 5000, nothing)
    cv.createTrackbar('thrs2', 'edge', 4000, 5000, nothing)

    cap = cv.VideoCapture(cam, cv.CAP_V4L2)
    if cam == 1:
        cap.set(cv.CAP_PROP_CONVERT_RGB, 0)

    while True:
        if cam == 1:
            ret, img_origin = cap.read()
            img = cv.cvtColor(img_origin, cv.COLOR_YUV2BGR_NV12)
        else:
            ret, img = cap.read()
        gray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
        thrs1 = cv.getTrackbarPos('thrs1', 'edge')
        thrs2 = cv.getTrackbarPos('thrs2', 'edge')
        edge = cv.Canny(gray, thrs1, thrs2, apertureSize=5)
        vis = img.copy()
        vis = np.uint8(vis/2.)
        vis[edge != 0] = (0, 255, 0)
        cv.imshow('edge', vis)
        ch = cv.waitKey(5)
        if ch == 27:
            break

    print('Done')


if __name__ == '__main__':
    print(__doc__)
    main()
    cv.destroyAllWindows()

同样的，桌面方式登录后，终端中user身份直接运行“v4l2-ctl --list-device”指令，确认USB摄像机的ID号。然后再运行：

cd /usr/share/doc/opencv-doc/examples/python/
python3 edge.py 4

默认参数下运行效果如下：

调节thrs1、thrs2参数后，检测的图像边缘区分更明显。

五、小结

综上评测效果，昉·星光2连接一个通用的USB摄像机，即可高效实现基于OpenCV的人脸检测识别、通用物体识别、二维码检测与解析，图像边缘检测等功能，证实该开发板可以运用在AI智能终端产品设计上，赛昉科技也提供了诸多指导性文档，一个应用可通过多种模型来实现，有提供C++源代码方式的接口，也有python源代码。提供的参考例程很有代表性，足见该开发板应用的场景广泛，尤其是在边缘AI设计中，具有很好的市场竞争力。