本文來源電子發(fā)燒友社區(qū)，作者：碼農(nóng)愛學(xué)習(xí)， 帖子地址：https://bbs.elecfans.com/jishu_2308719_1_1.html



上篇文章，測(cè)試了EASY EAI Nano的人臉檢測(cè)功能，本篇進(jìn)行人臉識(shí)別功能。

人臉檢測(cè)，只是將圖像中的人臉的位置檢測(cè)出來，人臉識(shí)別，又增加了一部，在檢測(cè)到人臉后，還要識(shí)別出這張臉是誰的臉。

本篇參考官方文檔：https://www.easy-eai.com/document_details/3/110

1 API介紹

主要來看下人臉識(shí)別組件。

face_recognition.h的主要內(nèi)容

//人臉識(shí)別初始化函數(shù)
int face_recognition_init(rknn_context *ctx, const char * path);
?
//人臉識(shí)別執(zhí)行函數(shù)
int face_recognition_run(rknn_context ctx, cv::Mat *face_image, float (*feature)[512]);
?
//人臉識(shí)別特征比對(duì)函數(shù)
float face_recognition_comparison(float *feature_1, float *feature_2, int output_len);
?
//人臉識(shí)別釋放函數(shù)
int face_recognition_release(rknn_context ctx);

一些參數(shù)：

• ctx:輸入?yún)?shù),rknn_context句柄
• path:輸入?yún)?shù),算法模型路徑
• face_image:輸入?yún)?shù),圖像數(shù)據(jù)輸入(cv::Mat是Opencv的類型)
• feature:輸出參數(shù)，算法輸出的人臉特征碼

face_alignment.h的主要內(nèi)容

cv::Mat face_alignment(cv::Mat img, cv::Point2f* points);

geometry.h的主要內(nèi)容

typedef struct{
    float x;
    float y;
}fPoint_t;
?
typedef struct{
    float left;   //x1
    float top;    //y1
    float right;  //x2
    float bottom; //y2
}fRect_t;
?
typedef struct{
    int32_t x;
    int32_t y;
}s32Point_t;
?
typedef struct{
    int32_t left;   //x1
    int32_t top;    //y1
    int32_t right;  //x2
    int32_t bottom; //y2
}s32Rect_t;
?
//判斷點(diǎn)是否在矩形內(nèi)
extern bool point_in_rect(s32Point_t point, s32Rect_t rect);
?
//計(jì)算矩形面積
extern int32_t calc_rect_square(s32Rect_t rect);
?
//找出面積較小矩形
extern s32Rect_t min_rect(s32Rect_t rect1, s32Rect_t rect2);
?
//找出面積較大矩形
extern s32Rect_t max_rect(s32Rect_t rect1, s32Rect_t rect2);
?
//判斷矩形是否相交或相切
extern bool rect_is_intersect(s32Rect_t rect1, s32Rect_t rect2);
?
//計(jì)算兩矩形相交部分面積(若相切，面積也為0)
extern int32_t calc_rect_intersect_square(s32Rect_t rect1, s32Rect_t rect2);
?
//計(jì)算[兩矩形相交部分面積]與[小矩形面積]之比
extern double calc_intersect_of_min_rect(s32Rect_t rect1, s32Rect_t rect2);
?
//計(jì)算兩矩形的交并比
extern double calc_intersect_of_union(s32Rect_t rect1, s32Rect_t rect2);

2 代碼分析與修改

主要的修改是將紅外攝像頭采集和活體檢測(cè)去掉，改用外接RGB攝像頭，另外，識(shí)別結(jié)果的顯示，將更多的信息（識(shí)別的id，人物名稱，耗時(shí)等）展示在屏幕上。

2.1 圖像采集與顯示線程(主線程)

重新定義識(shí)別結(jié)果：

typedef struct{
    bool bHasFace;     //是否檢測(cè)到人臉
    bool bMatch;       //是否與注冊(cè)的人臉匹配
    char idStr[128];   //匹配的人臉的id
    char nameStr[128]; //匹配的人臉的名稱
    float similarity;  //匹配的相似度
    uint64_t useTime;  //匹配用時(shí)
    uint32_t x1;       //識(shí)別到的人臉框的4個(gè)點(diǎn)
uint32_t y1;
uint32_t x2;
uint32_t y2;
}Result_t;

官方例程用到了紅外攝像頭，用于活體檢測(cè)，此次測(cè)試，為了使用電腦屏幕上的人物圖片進(jìn)行人臉檢測(cè)，去掉了活體檢測(cè)功能，并換做只使用USB攝像頭采集RGB圖像進(jìn)行人臉識(shí)別。

主函數(shù)邏輯如下，和上篇進(jìn)行人臉檢測(cè)的代碼邏輯類似。

#define COLOR_RED Scalar(255, 0, 0)
#define COLOR_GREEN Scalar(0, 255, 0)
?
int main(int argc, char **argv)
{
int ret = 0;
    int rgbRet = 0;
    disp_screen_t screen = {0};

char *pRGBbuf = NULL;
int skip = 0;

pthread_t mTid;
Result_t *pResult = NULL;

Mat image;
// 1.打開USB攝像頭
ret = usbcamera_init(USB2_0, USB_DIRECT, CAMERA_WIDTH, CAMERA_HEIGHT, 180);
if (ret) {
printf("error: %s, %dn", __func__, __LINE__);
goto exit_donothing;
}

pRGBbuf = NULL;
pRGBbuf = (char *)malloc(IMAGE_SIZE);
if (!pRGBbuf) {
printf("error: %s, %dn", __func__, __LINE__);
ret = -1;
goto exit_freeusb;
}

// 跳過前10幀
skip = 10;
while(skip--) {
ret = usbcamera_getframe(USB2_0, USB_DIRECT, pRGBbuf);
if (ret) {
printf("error: %s, %dn", __func__, __LINE__);
goto exit_freeusb_freebuf;
}
}
?
// 2.創(chuàng)建識(shí)別線程，以及圖像互斥鎖
pthread_mutex_init(&img_lock, NULL);
pResult = (Result_t *)malloc(sizeof(Result_t));
if(NULL == pResult){
goto exit_free_all;
}
memset(pResult, 0, sizeof(Result_t));
if(0 != CreateNormalThread(detect_thread_entry, pResult, &mTid)){
free(pResult);
}
?
// 3.顯示初始化
    screen.screen_width = SCREEN_WIDTH;
    screen.screen_height = SCREEN_HEIGHT;
    screen.wins[0].enable = 1;
    screen.wins[0].in_fmt = IMAGE_TYPE_RGB888;
    screen.wins[0].in_w = CAMERA_WIDTH;
    screen.wins[0].in_h = CAMERA_HEIGHT;
    screen.wins[0].rotation = 90;
    screen.wins[0].win_x = 0;
    screen.wins[0].win_y = 0;
    screen.wins[0].win_w = 720;
    screen.wins[0].win_h = 1280;
    ret = disp_init_pro(&screen);
if (ret) {
printf("error: %s, %dn", __func__, __LINE__);
goto exit_free_all;
}
    
// 4.(取流 + 顯示)循環(huán)
while(1)
    {
// 4.1、取流
pthread_mutex_lock(&img_lock);
ret = usbcamera_getframe(USB2_0, USB_DIRECT, pRGBbuf);
        if (0 != rgbRet) 
        {
            printf("error: %s, %dn", __func__, __LINE__);
pthread_mutex_unlock(&img_lock);
            continue;
        }
algorithm_image = Mat(CAMERA_HEIGHT, CAMERA_WIDTH, CV_8UC3, pRGBbuf);
image = algorithm_image.clone();
pthread_mutex_unlock(&img_lock);
?
// 4.2、顯示
        if (pResult->bHasFace) //檢測(cè)到人臉
        {
            Scalar color;
            //識(shí)別到已注冊(cè)的人臉
            if(pResult->bMatch)
            {
                color = COLOR_GREEN;
                cv::putText(image, std::string("idStr: ") + std::string(pResult->idStr), cv::Point2f(30, 50), cv::FONT_HERSHEY_SIMPLEX, 1.45, CV_RGB(255,0,0),3.0);
                cv::putText(image, std::string("name: ") + std::string(pResult->nameStr), cv::Point2f(30, 100), cv::FONT_HERSHEY_SIMPLEX, 1.45, CV_RGB(255,0,0),3.0);
                cv::putText(image, std::string("similarity: ") + std::to_string(pResult->similarity), cv::Point2f(30, 150), cv::FONT_HERSHEY_SIMPLEX, 1.45, CV_RGB(255,0,0),3.0);
                cv::putText(image, std::string("use time: ") + std::to_string(pResult->useTime), cv::Point2f(30, 200), cv::FONT_HERSHEY_SIMPLEX, 1.45, CV_RGB(255,0,0),3.0);
            }
            else
            {
                color = COLOR_RED;
                cv::putText(image, std::string("unknow face"), cv::Point2f(30, 50), cv::FONT_HERSHEY_SIMPLEX, 1.45, CV_RGB(0,0,255),3.0);
            }
            // 畫框
            rectangle(image, Point(pResult->x1, pResult->y1), Point(pResult->x2, pResult->y2), color, 3);
        }
        disp_commit(image.data, IMAGE_SIZE);

        usleep(20*1000);
}
?
exit_free_all:
pthread_mutex_destroy(&img_lock);
exit_freeusb_freebuf:
free(pRGBbuf);
pRGBbuf = NULL;
exit_freeusb:
usbcamera_exit(USB2_0, USB_DIRECT);
exit_donothing:
    return ret;
}

2.2 人臉識(shí)別處理

去掉紅外攝像頭的活體檢測(cè)邏輯，只使用USB攝像頭的圖像進(jìn)行人臉識(shí)別

void *detect_thread_entry(void *para)
{
int ret;
uint64_t start_time,end_time;
Result_t *pResult = (Result_t *)para;
    
// 初始化人臉檢測(cè)
rknn_context detect_ctx;
std::vector detect_result;
Point2f points[5];
    s32Rect_t rgbRect;
printf("face detect init!n");
ret = face_detect_init(&detect_ctx, "./face_detect.model");
if( ret < 0)
    {
printf("face_detect fail! ret=%dn", ret);
return NULL;
}
    
// 初始化人臉識(shí)別
rknn_context recognition_ctx;
float face_feature[512];
printf("face recognition init!n");
ret =  face_recognition_init(&recognition_ctx, "./face_recognition.model");
if( ret < 0)
    {
printf("face_recognition fail! ret=%dn", ret);
return NULL;
}
?
    // 初始化數(shù)據(jù)庫(kù)
    database_init();
    // 同步數(shù)據(jù)庫(kù)所有數(shù)據(jù)到內(nèi)存
faceData_t *pFaceData = (faceData_t *)malloc(MAX_USER_NUM * sizeof(faceData_t));
    memset(pFaceData, 0, MAX_USER_NUM * sizeof(faceData_t));
    int peopleNum = database_getData_to_memory(pFaceData);

// 初始化按鍵事件
keyEvent_init();
set_event_handle(dataBase_opt_handle);
    
Mat image;
Mat face_algin;
float similarity; //特征值相似度比對(duì)
int face_index = 0;
while(1)
{
if(g_delete_all_record)
        {
g_delete_all_record = false;
// 刪除庫(kù)
database_delete_all_record();
// 重載數(shù)據(jù)庫(kù)
peopleNum = database_getData_to_memory(pFaceData);
}

        if(algorithm_image.empty()) 
        {
usleep(5);
            continue;
        }

pthread_mutex_lock(&img_lock);
image = algorithm_image.clone();
pthread_mutex_unlock(&img_lock);
        
// 人臉檢測(cè)，計(jì)算出人臉位置
ret = face_detect_run(detect_ctx, image, detect_result);
if(ret <= 0)
        {
// 識(shí)別結(jié)果數(shù)據(jù)，復(fù)位
memset(pResult, 0 , sizeof(Result_t));
g_input_feature = false;
usleep(1000);
continue;
}
        
        
        rgbRect.left   = (uint32_t)(detect_result[0].box.x);
        rgbRect.top    = (uint32_t)(detect_result[0].box.y);
        rgbRect.right  = (uint32_t)(detect_result[0].box.x + detect_result[0].box.width);
        rgbRect.bottom = (uint32_t)(detect_result[0].box.y + detect_result[0].box.height);
pResult->bHasFace = true;
pResult->x1 = rgbRect.left;
pResult->y1 = rgbRect.top;
pResult->x2 = rgbRect.right;
pResult->y2 = rgbRect.bottom;
for (int i = 0; i < (int)detect_result[0].landmarks.size(); ++i) {
points[i].x = (int)detect_result[0].landmarks[i].x;
points[i].y = (int)detect_result[0].landmarks[i].y;
}
        
// 人臉校正(從圖像中裁出人臉)
face_algin = face_alignment(image, points);
        
// 人臉識(shí)別，計(jì)算特征值
start_time = get_timeval_ms();
face_recognition_run(recognition_ctx, &face_algin, &face_feature);
end_time = get_timeval_ms();
        pResult->useTime = end_time - start_time;
printf("n>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>n");
printf("face_recognition_run use time: %llun", pResult->useTime); // 【打印耗時(shí)】
        
        // 特征值比對(duì)，得出id
similarity = -0.5;
if(peopleNum > 0)
        {
for(face_index = 0; face_index < peopleNum; ++face_index)
            {
similarity = face_recognition_comparison(face_feature, (float *)((pFaceData + face_index)->feature), 512);
if(similarity > 0.5) 
                {
                    break;
                }
}
}
        pResult->similarity = similarity;
printf("similarity:%fn", similarity); //【打印得分】
if((face_index < peopleNum)&&(similarity > 0.5))
        {
            pResult->bMatch = true;
            
// 用id，找名字
            memcpy(pResult->idStr, (pFaceData + face_index)->id, 128);
printf("idStr : %sn", pResult->idStr); //【打印id】
database_id_is_exist(pResult->idStr, pResult->nameStr, sizeof(pResult->nameStr));
printf("person name : %sn", pResult->nameStr); //【打印名字】

            // 按鍵被按下，更新特征值
if(g_input_feature)
            {
g_input_feature = false;
// 特征值入庫(kù)
database_add_record(pResult->idStr, pResult->nameStr, (char *)face_feature, sizeof(face_feature));
// 重載數(shù)據(jù)庫(kù)
peopleNum = database_getData_to_memory(pFaceData);
}
}
        else
        {
            pResult->bMatch = false;
            
// 按鍵被按下，錄入特征值
if(g_input_feature)
            {
g_input_feature = false;
char idStr[32]={0};
char nameStr[32]={0};
sprintf(idStr, "%05d", face_index+1);
sprintf(nameStr, "people_%d", face_index+1);
// 特征值入庫(kù)
database_add_record(idStr, nameStr, (char *)face_feature, sizeof(face_feature));
// 重載數(shù)據(jù)庫(kù)
peopleNum = database_getData_to_memory(pFaceData);
}
}
        usleep(16*1000);
}

/* 數(shù)據(jù)庫(kù)內(nèi)存數(shù)據(jù)釋放 */
free(pFaceData);
/* 數(shù)據(jù)庫(kù)釋放 */
    database_exit();
/* 人臉識(shí)別釋放 */
face_recognition_release(recognition_ctx);
/* 人臉檢測(cè)釋放 */
face_detect_release(detect_ctx);
return NULL;
}

3 測(cè)試

測(cè)試視頻見文章底部視頻，這里再放一張測(cè)試圖片：

4 總結(jié)

本篇介紹了EASY EAI Nano的人臉識(shí)別功能，與上篇的人臉檢測(cè)相比，在檢測(cè)到有人臉的基礎(chǔ)上，通過先錄入人臉顯示到數(shù)據(jù)庫(kù)，可以對(duì)比當(dāng)前識(shí)別的誰的臉，實(shí)際測(cè)試，去掉活體檢測(cè)功能后，通過外接USB攝像頭來識(shí)別電腦屏幕上的3個(gè)人的多張不同人臉，可以分辨出不同人的人臉。


-----------


附上EASY EAI Nano多人臉識(shí)別測(cè)試【視頻】，詳見作者原帖子內(nèi)容。