NCNN人脸检测和关键点检测

参考以下连接获取原始 retinaface模型：

mirrors / wzj5133329 / retinaface_caffe · GitCode

https://github.com/deepinsight/insightface/tree/master/RetinaFace

retinaface(mxnet)的CAFFE模型

https://github.com/Charrin/RetinaFace-Cpp/tree/master/convert_models/mnet

使用caffe2ncnn将模型转换为NCNN格式，需要注意的是，环境需要事先安装protobuf的开发环境，否则CMAKE阶段会警告无法生成caffe2ncnn和onnx2ncnn的问题，因为这两种格式用的都是protobuf来描述的。

sudo apt-get install libprotobuf-dev protobuf-compiler
sudo apt-get autoremove --purge protobuf-compiler
sudo apt-get autoremove --purge libprotobuf-dev

编译成功后，执行命令./caffe2ncnn mnet.prototxt mnet.caffemodel进行模型转换即可。

之后重命名ncnn.bin,ncnn.param为应用识别的文件名，即可以推理验证：

执行命令：./retinaface ./beauty.jpg

lena大妈

百度人脸的结果： 

对结果进行检测:

以上是用NCNN自带的模型测试的效果，下面用自己转的模型验证：

按照符合网络输入要求的图像尺寸做scale后输入：

效果很好！

下面是用NCNN YOLOV5用的模型文件，对比一下人脸检测和人形目标检测算法的区别：

小鲜肉和老腊肉：

网络分析：

一输入九输出

基于NCNN修改的后处理程序：

// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include "net.h"

#if defined(USE_NCNN_SIMPLEOCV)
#include "simpleocv.h"
#else
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#endif
#include <stdio.h>
#include <vector>
#include <vector>
#include <iostream>
using namespace std;

struct FaceObject
{
    cv::Rect_<float> rect;
    cv::Point2f landmark[5];
    float prob;
};

static inline float intersection_area(const FaceObject& a, const FaceObject& b)
{
    cv::Rect_<float> inter = a.rect & b.rect;
    return inter.area();
}

static void qsort_descent_inplace(std::vector<FaceObject>& faceobjects, int left, int right)
{
    int i = left;
    int j = right;
    float p = faceobjects[(left + right) / 2].prob;

    while (i <= j)
    {
        while (faceobjects[i].prob > p)
            i++;

        while (faceobjects[j].prob < p)
            j--;

        if (i <= j)
        {
            // swap
            std::swap(faceobjects[i], faceobjects[j]);

            i++;
            j--;
        }
    }

    #pragma omp parallel sections
    {
        #pragma omp section
        {
            if (left < j) qsort_descent_inplace(faceobjects, left, j);
        }
        #pragma omp section
        {
            if (i < right) qsort_descent_inplace(faceobjects, i, right);
        }
    }
}

static void qsort_descent_inplace(std::vector<FaceObject>& faceobjects)
{
    if (faceobjects.empty())
        return;

    qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);
}

static void nms_sorted_bboxes(const std::vector<FaceObject>& faceobjects, std::vector<int>& picked, float nms_threshold)
{
    picked.clear();

    const int n = faceobjects.size();

    std::vector<float> areas(n);
    for (int i = 0; i < n; i++)
    {
        areas[i] = faceobjects[i].rect.area();
    }

    for (int i = 0; i < n; i++)
    {
        const FaceObject& a = faceobjects[i];

        int keep = 1;
        for (int j = 0; j < (int)picked.size(); j++)
        {
            const FaceObject& b = faceobjects[picked[j]];

            // intersection over union
            float inter_area = intersection_area(a, b);
            float union_area = areas[i] + areas[picked[j]] - inter_area;
            //             float IoU = inter_area / union_area
            if (inter_area / union_area > nms_threshold)
                keep = 0;
        }

        if (keep)
            picked.push_back(i);
    }
}

// copy from src/layer/proposal.cpp
static ncnn::Mat generate_anchors(int base_size, const ncnn::Mat& ratios, const ncnn::Mat& scales)
{
    int num_ratio = ratios.w;
    int num_scale = scales.w;

    ncnn::Mat anchors;
    anchors.create(4, num_ratio * num_scale);

    const float cx = base_size * 0.5f;
    const float cy = base_size * 0.5f;

    for (int i = 0; i < num_ratio; i++)
    {
        float ar = ratios[i];

        int r_w = round(base_size / sqrt(ar));
        int r_h = round(r_w * ar); //round(base_size * sqrt(ar));

        for (int j = 0; j < num_scale; j++)
        {
            float scale = scales[j];

            float rs_w = r_w * scale;
            float rs_h = r_h * scale;

            float* anchor = anchors.row(i * num_scale + j);

            anchor[0] = cx - rs_w * 0.5f;
            anchor[1] = cy - rs_h * 0.5f;
            anchor[2] = cx + rs_w * 0.5f;
            anchor[3] = cy + rs_h * 0.5f;
        }
    }

    return anchors;
}

static void generate_proposals(const ncnn::Mat& anchors, int feat_stride, const ncnn::Mat& score_blob, const ncnn::Mat& bbox_blob, const ncnn::Mat& landmark_blob, float prob_threshold, std::vector<FaceObject>& faceobjects)
{
    int w = score_blob.w;
    int h = score_blob.h;

	printf("%s line %d, w %d, h %d.\n", __func__, __LINE__, w, h);
	printf("%s line %d, d %d, c %d %d  %d step %ld  %ld \n", __func__, __LINE__, score_blob.d, score_blob.c, score_blob.dims, score_blob.elempack, score_blob.cstep, score_blob.elemsize);
	printf("%s line %d, d %d, c %d %d  %d step %ld  %ld \n", __func__, __LINE__, bbox_blob.d, bbox_blob.c, bbox_blob.dims, bbox_blob.elempack, bbox_blob.cstep, bbox_blob.elemsize);
	printf("%s line %d, d %d, c %d %d  %d step %ld  %ld \n", __func__, __LINE__, landmark_blob.d, landmark_blob.c, landmark_blob.dims, landmark_blob.elempack, landmark_blob.cstep, landmark_blob.elemsize);

    // generate face proposal from bbox deltas and shifted anchors
    const int num_anchors = anchors.h;

    for (int q = 0; q < num_anchors; q++)
    {
        const float* anchor = anchors.row(q);

        const ncnn::Mat score = score_blob.channel(q + num_anchors);
        const ncnn::Mat bbox = bbox_blob.channel_range(q * 4, 4);
        const ncnn::Mat landmark = landmark_blob.channel_range(q * 10, 10);

        // shifted anchor
        float anchor_y = anchor[1];

        float anchor_w = anchor[2] - anchor[0];
        float anchor_h = anchor[3] - anchor[1];

        for (int i = 0; i < h; i++)
        {
            float anchor_x = anchor[0];

            for (int j = 0; j < w; j++)
            {
                int index = i * w + j;

                float prob = score[index];

                if (prob >= prob_threshold)
                {
                    // apply center size
                    float dx = bbox.channel(0)[index];
                    float dy = bbox.channel(1)[index];
                    float dw = bbox.channel(2)[index];
                    float dh = bbox.channel(3)[index];

                    float cx = anchor_x + anchor_w * 0.5f;
                    float cy = anchor_y + anchor_h * 0.5f;

                    float pb_cx = cx + anchor_w * dx;
                    float pb_cy = cy + anchor_h * dy;

                    float pb_w = anchor_w * exp(dw);
                    float pb_h = anchor_h * exp(dh);

                    float x0 = pb_cx - pb_w * 0.5f;
                    float y0 = pb_cy - pb_h * 0.5f;
                    float x1 = pb_cx + pb_w * 0.5f;
                    float y1 = pb_cy + pb_h * 0.5f;

                    FaceObject obj;
                    obj.rect.x = x0;
                    obj.rect.y = y0;
                    obj.rect.width = x1 - x0 + 1;
                    obj.rect.height = y1 - y0 + 1;
                    obj.landmark[0].x = cx + (anchor_w + 1) * landmark.channel(0)[index];
                    obj.landmark[0].y = cy + (anchor_h + 1) * landmark.channel(1)[index];
                    obj.landmark[1].x = cx + (anchor_w + 1) * landmark.channel(2)[index];
                    obj.landmark[1].y = cy + (anchor_h + 1) * landmark.channel(3)[index];
                    obj.landmark[2].x = cx + (anchor_w + 1) * landmark.channel(4)[index];
                    obj.landmark[2].y = cy + (anchor_h + 1) * landmark.channel(5)[index];
                    obj.landmark[3].x = cx + (anchor_w + 1) * landmark.channel(6)[index];
                    obj.landmark[3].y = cy + (anchor_h + 1) * landmark.channel(7)[index];
                    obj.landmark[4].x = cx + (anchor_w + 1) * landmark.channel(8)[index];
                    obj.landmark[4].y = cy + (anchor_h + 1) * landmark.channel(9)[index];
                    obj.prob = prob;

                    faceobjects.push_back(obj);
                }

                anchor_x += feat_stride;
            }

            anchor_y += feat_stride;
        }
    }
}

int get_tensor_data(string file_path, float **data)
{
	int len = 0;
	static float *memory = NULL;
	static int max_len = 10*1024*1024;

	if(memory == NULL)
	{
		memory = (float *)malloc(max_len * sizeof(float));
	}

	FILE *fp = NULL;
	if((fp = fopen(file_path.c_str(), "r")) == NULL)
	{
		cout << "open tensor file error!" << endl;
		exit(-1);
	}

	int file_len = 0;
	while(!feof(fp))
	{
		fscanf(fp, "%f ", &memory[len ++]);
	}

	*data = (float *)malloc(len * sizeof(float));
	memcpy(*data, memory, len*sizeof(float));
	fclose(fp);

	if(len == 0 || *data == NULL)
	{
		cout << " read tensor error happend." << "len:" << len << "data address:" << *data << endl;
		exit(-1);
	}

	printf("len %d.\n", len);
	return len;
}

static int detect_retinaface(const cv::Mat& bgr, std::vector<FaceObject>& faceobjects)
{
    ncnn::Net retinaface;

    retinaface.opt.use_vulkan_compute = true;

    // model is converted from
    // https://github.com/deepinsight/insightface/tree/master/RetinaFace#retinaface-pretrained-models
    // https://github.com/deepinsight/insightface/issues/669
    // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models
    //     retinaface.load_param("retinaface-R50.param");
    //     retinaface.load_model("retinaface-R50.bin");
    retinaface.load_param("mnet.25-opt.param");
    retinaface.load_model("mnet.25-opt.bin");

    const float prob_threshold = 0.8f;
    const float nms_threshold = 0.4f;

    int img_w = bgr.cols;
    int img_h = bgr.rows;

    ncnn::Mat in = ncnn::Mat::from_pixels(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h);

    ncnn::Extractor ex = retinaface.create_extractor();

    ex.input("data", in);

    std::vector<FaceObject> faceproposals;

	float *score0 = (float*)calloc(20*20*4, sizeof(float));
	float *bbox0 = (float*)calloc(20*20*8, sizeof(float));
	float *landmark0 = (float*)calloc(20*20*20, sizeof(float));

	float *score1 = (float*)calloc(40*40*4, sizeof(float));
	float *bbox1 = (float*)calloc(40*40*8, sizeof(float));
	float *landmark1 = (float*)calloc(40*40*20, sizeof(float));

	float *score2 = (float*)calloc(80*80*4, sizeof(float));
	float *bbox2 = (float*)calloc(80*80*8, sizeof(float));
	float *landmark2 = (float*)calloc(80*80*20, sizeof(float));
	
	get_tensor_data("./iter_0_output_212_out0_1_4_20_20.tensor", &score0);
	get_tensor_data("./iter_0_output_210_out0_1_8_20_20.tensor", &bbox0);
	get_tensor_data("./iter_0_output_211_out0_1_20_20_20.tensor", &landmark0);

	get_tensor_data("./iter_0_output_215_out0_1_4_40_40.tensor", &score1);
	get_tensor_data("./iter_0_output_213_out0_1_8_40_40.tensor", &bbox1);
	get_tensor_data("./iter_0_output_214_out0_1_20_40_40.tensor", &landmark1);

	get_tensor_data("./iter_0_output_218_out0_1_4_80_80.tensor", &score2);
	get_tensor_data("./iter_0_output_216_out0_1_8_80_80.tensor", &bbox2);
	get_tensor_data("./iter_0_output_217_out0_1_20_80_80.tensor", &landmark2);

    // stride 32
    {
        ncnn::Mat score_blob(20*20*4), bbox_blob(20*20*8), landmark_blob(20*20*20);
#if 0
		ex.extract("face_rpn_cls_prob_reshape_stride32", score_blob);
		ex.extract("face_rpn_bbox_pred_stride32", bbox_blob);
		ex.extract("face_rpn_landmark_pred_stride32", landmark_blob);
#else
        int j = 0;
        for(j = 0; j < 20*20*4; j ++) score_blob[j] = score0[j];
        for(j = 0; j < 20*20*8; j ++) bbox_blob[j] = bbox0[j];
        for(j = 0; j < 20*20*20; j ++) landmark_blob[j] = landmark0[j];
        score_blob.w = score_blob.h = bbox_blob.w = bbox_blob.h = landmark_blob.w = landmark_blob.h = 20;
        score_blob.d = 1;
	    bbox_blob.d = 1;
	    landmark_blob.d = 1;
        score_blob.c = 4;
	    bbox_blob.c = 8;
	    landmark_blob.c = 20;

        score_blob.dims = 3;
	    bbox_blob.dims = 3;
	    landmark_blob.dims = 3;

        score_blob.elempack = 1;
	    bbox_blob.elempack = 1;
	    landmark_blob.elempack = 1;

        score_blob.cstep = 400;
	    bbox_blob.cstep = 400;
	    landmark_blob.cstep = 400;

	    landmark_blob.elemsize = bbox_blob.elemsize = score_blob.elemsize = 4;

        free(score0);
        free(bbox0);
        free(landmark0);
#endif

        const int base_size = 16;
        const int feat_stride = 32;
        ncnn::Mat ratios(1);
        ratios[0] = 1.f;
        ncnn::Mat scales(2);
        scales[0] = 32.f;
        scales[1] = 16.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects32;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects32);

        faceproposals.insert(faceproposals.end(), faceobjects32.begin(), faceobjects32.end());
    }

    // stride 16
    {
        ncnn::Mat score_blob(40*40*4), bbox_blob(40*40*8), landmark_blob(40*40*20);
		
#if 0
		ex.extract("face_rpn_cls_prob_reshape_stride16", score_blob);
		ex.extract("face_rpn_bbox_pred_stride16", bbox_blob);
		ex.extract("face_rpn_landmark_pred_stride16", landmark_blob);
#else
        int j = 0;
        for(j = 0; j < 40*40*4; j ++) score_blob[j] = score1[j];
        for(j = 0; j < 40*40*8; j ++) bbox_blob[j] = bbox1[j];
        for(j = 0; j < 40*40*20; j ++) landmark_blob[j] = landmark1[j];
        score_blob.w = score_blob.h = bbox_blob.w = bbox_blob.h = landmark_blob.w = landmark_blob.h = 40;
        score_blob.d = 1;
	    bbox_blob.d = 1;
	    landmark_blob.d = 1;
        score_blob.c = 4;
	    bbox_blob.c = 8;
	    landmark_blob.c = 20;
        free(score1);
        free(bbox1);
        free(landmark1);
        score_blob.dims = 3;
	    bbox_blob.dims = 3;
	    landmark_blob.dims = 3;
        score_blob.elempack = 1;
	    bbox_blob.elempack = 1;
	    landmark_blob.elempack = 1;

        score_blob.cstep = 1600;
	    bbox_blob.cstep = 1600;
	    landmark_blob.cstep = 1600;
	    landmark_blob.elemsize = bbox_blob.elemsize = score_blob.elemsize = 4;
#endif

        const int base_size = 16;
        const int feat_stride = 16;
        ncnn::Mat ratios(1);
        ratios[0] = 1.f;
        ncnn::Mat scales(2);
        scales[0] = 8.f;
        scales[1] = 4.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects16;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects16);

        faceproposals.insert(faceproposals.end(), faceobjects16.begin(), faceobjects16.end());
    }

    // stride 8
    {
        ncnn::Mat score_blob(80*80*4), bbox_blob(80*80*8), landmark_blob(80*80*20);
#if 0
		ex.extract("face_rpn_cls_prob_reshape_stride8", score_blob);
		ex.extract("face_rpn_bbox_pred_stride8", bbox_blob);
		ex.extract("face_rpn_landmark_pred_stride8", landmark_blob);
#else
        int j = 0;
        for(j = 0; j < 80*80*4; j ++) score_blob[j] = score2[j];
        for(j = 0; j < 80*80*8; j ++) bbox_blob[j] = bbox2[j];
        for(j = 0; j < 80*80*20; j ++) landmark_blob[j] = landmark2[j];
        score_blob.w = score_blob.h = bbox_blob.w = bbox_blob.h = landmark_blob.w = landmark_blob.h = 80;
        score_blob.d = 1;
	    bbox_blob.d = 1;
	    landmark_blob.d = 1;
        score_blob.c = 4;
	    bbox_blob.c = 8;
	    landmark_blob.c = 20;
        free(score2);
        free(bbox2);
        free(landmark2);
        score_blob.dims = 3;
	    bbox_blob.dims = 3;
	    landmark_blob.dims = 3;

        score_blob.elempack = 1;
	    bbox_blob.elempack = 1;
	    landmark_blob.elempack = 1;

        score_blob.cstep = 6400;
	    bbox_blob.cstep = 6400;
	    landmark_blob.cstep = 6400;
	    landmark_blob.elemsize = bbox_blob.elemsize = score_blob.elemsize = 4;
#endif

        const int base_size = 16;
        const int feat_stride = 8;
        ncnn::Mat ratios(1);
        ratios[0] = 1.f;
        ncnn::Mat scales(2);
        scales[0] = 2.f;
        scales[1] = 1.f;
        ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);

        std::vector<FaceObject> faceobjects8;
        generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects8);

        faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());
    }

    // sort all proposals by score from highest to lowest
    qsort_descent_inplace(faceproposals);

    // apply nms with nms_threshold
    std::vector<int> picked;
    nms_sorted_bboxes(faceproposals, picked, nms_threshold);

    int face_count = picked.size();

    faceobjects.resize(face_count);
    for (int i = 0; i < face_count; i++)
    {
        faceobjects[i] = faceproposals[picked[i]];

        // clip to image size
        float x0 = faceobjects[i].rect.x;
        float y0 = faceobjects[i].rect.y;
        float x1 = x0 + faceobjects[i].rect.width;
        float y1 = y0 + faceobjects[i].rect.height;

        x0 = std::max(std::min(x0, (float)img_w - 1), 0.f);
        y0 = std::max(std::min(y0, (float)img_h - 1), 0.f);
        x1 = std::max(std::min(x1, (float)img_w - 1), 0.f);
        y1 = std::max(std::min(y1, (float)img_h - 1), 0.f);

        faceobjects[i].rect.x = x0;
        faceobjects[i].rect.y = y0;
        faceobjects[i].rect.width = x1 - x0;
        faceobjects[i].rect.height = y1 - y0;
    }

    return 0;
}

static void draw_faceobjects(const cv::Mat& bgr, const std::vector<FaceObject>& faceobjects)
{
    cv::Mat image = bgr.clone();

    for (size_t i = 0; i < faceobjects.size(); i++)
    {
        const FaceObject& obj = faceobjects[i];

        fprintf(stderr, "%.5f at %.2f %.2f %.2f x %.2f\n", obj.prob,
                obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);

        cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));

        cv::circle(image, obj.landmark[0], 2, cv::Scalar(0, 255, 255), -1);
        cv::circle(image, obj.landmark[1], 2, cv::Scalar(0, 255, 255), -1);
        cv::circle(image, obj.landmark[2], 2, cv::Scalar(0, 255, 255), -1);
        cv::circle(image, obj.landmark[3], 2, cv::Scalar(0, 255, 255), -1);
        cv::circle(image, obj.landmark[4], 2, cv::Scalar(0, 255, 255), -1);

        char text[256];
        sprintf(text, "%.1f%%", obj.prob * 100);

        int baseLine = 0;
        cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);

        int x = obj.rect.x;
        int y = obj.rect.y - label_size.height - baseLine;
        if (y < 0)
            y = 0;
        if (x + label_size.width > image.cols)
            x = image.cols - label_size.width;

        cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),
                      cv::Scalar(255, 255, 255), -1);

        cv::putText(image, text, cv::Point(x, y + label_size.height),
                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));
    }

    cv::imshow("image", image);
    cv::waitKey(0);
}

int main(int argc, char** argv)
{
    if (argc != 2)
    {
        fprintf(stderr, "Usage: %s [imagepath]\n", argv[0]);
        return -1;
    }

    const char* imagepath = argv[1];

    cv::Mat m = cv::imread(imagepath, 1);
    if (m.empty())
    {
        fprintf(stderr, "cv::imread %s failed\n", imagepath);
        return -1;
    }

    std::vector<FaceObject> faceobjects;
    detect_retinaface(m, faceobjects);

    draw_faceobjects(m, faceobjects);

    return 0;
}

在推理芯原工具inference出的tensor的时候遇到问题，只找到了三位美女的脸和五官，问题待查.

部署自己YOLOV5S

参考下面博客将yolov5s.pt转换为ONNX格式.

pytorch yolov5 推理和训练环境搭建_papaofdoudou的博客-优快云博客

下一步将onnx格式转换为NCNN格式：

参考资料

GitHub - wzj5133329/retinaface_caffe: 完整版caffe-cpp实现

---

结束！