人臉識別準備 -- 基於raspberry pi 3b + movidius

最近準備系統地學習一下深度學習和TensorFlow，就以人臉識別做爲目的。

十年前我作過一些圖像處理相關的項目和研究，涉及到圖像檢索。記得當時使用的是SIFT特徵提取，該特徵算子能很好地抵抗圖像旋轉、仿射變換等變化。能夠說SIFT是圖像特徵工程方面作得很出色的算子。

現現在深度學習特別是CNN，ResNet等模型被研究者發明以後，圖像特徵工程彷佛已經很「沒有必要」了。深度神經網絡經過多層表示可以更抽象地表示圖像的特徵（稱做embedding）。

在人臉識別也得益於深度學習，其中facenet的性能很是出色。facenet基於triplet loss訓練模型輸出128維embedding。訓練時準備M我的，每一個人N張圖像，目標使得同一我的的不一樣人臉的embedding距離儘可能小，而不一樣人的人臉圖像的embedding儘可能大。

本文將描述基於raspberry 3B + movidius做爲硬件平臺，TensorFlow facenet做爲模型實現人臉識別。後續將基於這套edge computing作一套完整的人臉識別系統，例如考勤系統。
本文將不涉及在線人臉檢測過程。

raspberry 3B

當前的系統：

pi@raspberrypi:~ $ uname -a
Linux raspberrypi 4.14.34-v7+ #1110 SMP Mon Apr 16 15:18:51 BST 2018 armv7l GNU/Linux

相關外設：

• 16G tf卡
• 官方攝像頭
• 3.5電阻觸屏

TensorFlow準備

首先在raspberry上安裝TensorFlow。目前raspberry上預裝了python2.7和python3.5.咱們選擇python3.5.
從https://github.com/lhelontra/tensorflow-on-arm/releases下載tensorflow-1.3.1-cp35-none-linux_armv7l.whl並安裝：
pip3 install tensorflow-1.3.1-cp35-none-linux_armv7l.whl
可能須要pip3一些別的：

# numpy issue
sudo apt-get install libatlas-base-dev
# opencv cv2
pip3 install opencv-python
sudo apt-get install libjpeg-dev libtiff5-dev libjasper-dev libpng12-dev

pip3 install sklearn
pip3 install scipy
# qt issue
sudo apt-get install libqtgui4 libqt4-test

測試：

pi@raspberrypi:~ $ python3
Python 3.5.3 (default, Jan 19 2017, 14:11:04)
[GCC 6.3.0 20170124] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow
>>> tensorflow.__version__
'1.3.1'

pi上運行facenet

有了TensorFlow以後咱們能夠編譯facenet並在pi上運行。https://github.com/davidsandberg/facenet/tree/tl_revisited
基於模型20170512-110547運行compare.py來比較多張圖像中人臉的距離。發現速度很是慢。
具體說，首先檢測圖像中的人臉，這裏運行了mtnet網絡，而後再經過facenet網絡inference。單獨測試inference的時間開銷20+秒（inference時人臉圖像都是160x160）。相比之下用dlib的開銷在2秒左右。這樣的性能很讓人沮喪？
爲了將facenet進行到底，我選擇加速，movidius是神經計算神器，inference速度很是快。

movidius sdk 安裝

clone代碼git clone -b ncsdk2 https://github.com/movidius/ncsdk.git
由於咱們事先安裝了TensorFlow，因此修改ncsdk.conf，再也不安裝TensorFlow，可是還須要caffe

INSTALL_DIR=/opt/movidius
INSTALL_CAFFE=yes
CAFFE_FLAVOR=ssd
CAFFE_USE_CUDA=no
INSTALL_TENSORFLOW=no
INSTALL_TOOLKIT=yes
PIP_SYSTEM_INSTALL=no
VERBOSE=yes
USE_VIRTUALENV=no
#MAKE_NJOBS=1

make install

ncs model編譯

clone代碼：git clone -b ncsdk2 https://github.com/movidius/ncappzoo.git
在tensorflow/facenet下，根據README一步一步編譯。最終獲得facenet_celeb_ncs.graph文件，這個文件是movidius識別的圖模型文件。

Movidius人臉識別

這裏我先不考慮在線人臉檢測。先準備一張照片，離線人臉檢測並保存人臉圖像做爲比對目標。先以一張人臉爲例，多我的臉圖像實際上是同樣的。
在線檢測時咱們將攝像頭的resolution設置小一些，例如280x280。在線識別是，人臉儘可能靠近攝像頭，這樣能夠認爲這張照片就是人臉照片。或者也能夠限定人臉在顯示屏上給定的一個區域。
目前inference的速度~100ms，當前對ncs還不是很瞭解，待進一步研究後再優化。

代碼以下（保存在ncappzoo/tensorflow/facenet）

• VALIDATED_IMAGES_DIR + '/my1.png' 是一張人臉圖像，經過人臉檢測獲得後保存的結果

#! /usr/bin/env python3

import sys
sys.path.insert(0, "../../ncapi2_shim")
import mvnc_simple_api as mvnc

import numpy
import cv2
import sys
import os

from picamera.array import PiRGBArray
from picamera import PiCamera
import time


# initialize the camera and grab a reference to the raw camera capture
camera = PiCamera()

camera.resolution = (280, 280)
camera.framerate = 32
rawCapture = PiRGBArray(camera, size=(280, 280))

frame_name=''
EXAMPLES_BASE_DIR='../../'
IMAGES_DIR = './'

VALIDATED_IMAGES_DIR = IMAGES_DIR + 'validated_images/'
validated_image_filename = VALIDATED_IMAGES_DIR + 'my1.png'

GRAPH_FILENAME = "facenet_celeb_ncs.graph"

# name of the opencv window
CV_WINDOW_NAME = "FaceNet"



# the same face will return 0.0
# different faces return higher numbers
# this is NOT between 0.0 and 1.0
FACE_MATCH_THRESHOLD = 1.2


# Run an inference on the passed image
# image_to_classify is the image on which an inference will be performed
#    upon successful return this image will be overlayed with boxes
#    and labels identifying the found objects within the image.
# ssd_mobilenet_graph is the Graph object from the NCAPI which will
#    be used to peform the inference.
def run_inference(image_to_classify, facenet_graph):

    # get a resized version of the image that is the dimensions
    # SSD Mobile net expects
    resized_image = preprocess_image(image_to_classify)

    # ***************************************************************
    # Send the image to the NCS
    # ***************************************************************
    facenet_graph.LoadTensor(resized_image.astype(numpy.float16), None)

    # ***************************************************************
    # Get the result from the NCS
    # ***************************************************************
    output, userobj = facenet_graph.GetResult()

    return output


# overlays the boxes and labels onto the display image.
# display_image is the image on which to overlay to
# image info is a text string to overlay onto the image.
# matching is a Boolean specifying if the image was a match.
# returns None
def overlay_on_image(display_image, image_info, matching):
    rect_width = 10
    offset = int(rect_width/2)
    if (image_info != None):
        cv2.putText(display_image, image_info, (30, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 0, 0), 1)
    if (matching):
        # match, green rectangle
        cv2.rectangle(display_image, (0+offset, 0+offset),
                      (display_image.shape[1]-offset-1, display_image.shape[0]-offset-1),
                      (0, 255, 0), 10)
    else:
        # not a match, red rectangle
        cv2.rectangle(display_image, (0+offset, 0+offset),
                      (display_image.shape[1]-offset-1, display_image.shape[0]-offset-1),
                      (0, 0, 255), 10)


# whiten an image
def whiten_image(source_image):
    source_mean = numpy.mean(source_image)
    source_standard_deviation = numpy.std(source_image)
    std_adjusted = numpy.maximum(source_standard_deviation, 1.0 / numpy.sqrt(source_image.size))
    whitened_image = numpy.multiply(numpy.subtract(source_image, source_mean), 1 / std_adjusted)
    return whitened_image

# create a preprocessed image from the source image that matches the
# network expectations and return it
def preprocess_image(src):
    # scale the image
    NETWORK_WIDTH = 160
    NETWORK_HEIGHT = 160
    preprocessed_image = cv2.resize(src, (NETWORK_WIDTH, NETWORK_HEIGHT))

    #convert to RGB
    preprocessed_image = cv2.cvtColor(preprocessed_image, cv2.COLOR_BGR2RGB)

    #whiten
    preprocessed_image = whiten_image(preprocessed_image)

    # return the preprocessed image
    return preprocessed_image

# determine if two images are of matching faces based on the
# the network output for both images.
def face_match(face1_output, face2_output):
    if (len(face1_output) != len(face2_output)):
        print('length mismatch in face_match')
        return False
    total_diff = 0
    for output_index in range(0, len(face1_output)):
        this_diff = numpy.square(face1_output[output_index] - face2_output[output_index])
        total_diff += this_diff
    print('Total Difference is: ' + str(total_diff))

    if (total_diff < FACE_MATCH_THRESHOLD):
        # the total difference between the two is under the threshold so
        # the faces match.
        return True

    # differences between faces was over the threshold above so
    # they didn't match.
    return False

# handles key presses
# raw_key is the return value from cv2.waitkey
# returns False if program should end, or True if should continue
def handle_keys(raw_key):
    ascii_code = raw_key & 0xFF
    if ((ascii_code == ord('q')) or (ascii_code == ord('Q'))):
        return False

    return True


# start the opencv webcam streaming and pass each frame
# from the camera to the facenet network for an inference
# Continue looping until the result of the camera frame inference
# matches the valid face output and then return.
# valid_output is inference result for the valid image
# validated image filename is the name of the valid image file
# graph is the ncsdk Graph object initialized with the facenet graph file
#   which we will run the inference on.
# returns None
def run_camera(valid_output, validated_image_filename, graph):

    frame_count = 0

    cv2.namedWindow(CV_WINDOW_NAME)

    found_match = False

    for frame in camera.capture_continuous(rawCapture, format="bgr", use_video_port=True):
        # grab the raw NumPy array representing the image, then initialize the timestamp
        # and occupied/unoccupied text
        vid_image = frame.array

        test_output = run_inference(vid_image, graph)


        if (face_match(valid_output, test_output)):
                print('PASS!  File ' + frame_name + ' matches ' + validated_image_filename)
                found_match = True
        else:
            found_match = False
            print('FAIL!  File ' + frame_name + ' does not match ' + validated_image_filename)

        overlay_on_image(vid_image, frame_name, found_match)

        # check if the window is visible, this means the user hasn't closed
        # the window via the X button
        prop_val = cv2.getWindowProperty(CV_WINDOW_NAME, cv2.WND_PROP_ASPECT_RATIO)
        if (prop_val < 0.0):
            print('window closed')
            break

        # display the results and wait for user to hit a key
        cv2.imshow(CV_WINDOW_NAME, vid_image)
        raw_key = cv2.waitKey(1)
        if (raw_key != -1):
            if (handle_keys(raw_key) == False):
                print('user pressed Q')
                break
        # show the frame
        #cv2.imshow("Frame", image)


        key = cv2.waitKey(1) & 0xFF

        # clear the stream in preparation for the next frame
        rawCapture.truncate(0)

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


# This function is called from the entry point to do
# all the work of the program
def main():

    # Get a list of ALL the sticks that are plugged in
    # we need at least one
    devices = mvnc.EnumerateDevices()
    if len(devices) == 0:
        print('No NCS devices found')
        quit()

    # Pick the first stick to run the network
    device = mvnc.Device(devices[0])

    # Open the NCS
    device.OpenDevice()

    # The graph file that was created with the ncsdk compiler
    graph_file_name = GRAPH_FILENAME

    # read in the graph file to memory buffer
    with open(graph_file_name, mode='rb') as f:
        graph_in_memory = f.read()

    # create the NCAPI graph instance from the memory buffer containing the graph file.
    graph = device.AllocateGraph(graph_in_memory)

    validated_image = cv2.imread(validated_image_filename)
    valid_output = run_inference(validated_image, graph)

    run_camera(valid_output, validated_image_filename, graph)

    # Clean up the graph and the device
    graph.DeallocateGraph()
    device.CloseDevice()


# main entry point for program. we'll call main() to do what needs to be done.
if __name__ == "__main__":
    sys.exit(main())