使用Azure認知服務快速搭建一個目標檢測平臺

前言

??博主參與了由CSDN和微軟共同舉辦的0元試用Azure人工智能認知服務的活動，體驗了一下Azure的計算機視覺方面的服務，在這里記錄一下如何基于Azure計算機視覺服務快速搭建一個目標檢測平臺，

1. 認知服務

??認知服務使每個開發人員和資料科學家都可以使用 AI，借助領先的模型，可以解鎖各種用例，只需要一個 API 呼叫，就可以將看、聽、說、搜索、理解和加快高級決策制定的能力嵌入到應用中，讓所有技能水平的開發人員和資料科學家都能輕松在其應用中添加 AI 功能，

??Azure的計算機視覺服務具體包括以下服務：

??本篇主要介紹的是影像分析服務中的目標檢測功能，Azure的計算機視覺服務對輸入影像的要求如下：

• 影像檔案格式必須是 JPEG、PNG、GIF 或 BMP
• 影像的檔案大小不能超過 4 MB
• 影像的尺寸必須大于 50 x 50 像素，對于讀取 API，影像的尺寸必須介于 50 x 50 和 10000 x 10000 像素之間，

2. 環境配置

2.1 創建資源

??按照下圖所示步驟來添加計算機視覺所需要的資源：

??創建完畢后，進入資源可以看到自己的終結點endpoint和密鑰subscription key：

2.2 創建python環境

	# 安裝Azure Computer Vision庫
	pip install --upgrade azure-cognitiveservices-vision-computervision
	# 安裝pillow庫
	pip install pillow
	# 安裝 matplotlib庫
	pip install matplotlib
	# 安裝opencv庫
	pip install opencv-python
	# 安裝ffmpeg庫
	pip install ffmpeg

3. 代碼實作

3.1 圖片檢測

from azure.cognitiveservices.vision.computervision import ComputerVisionClient
from azure.cognitiveservices.vision.computervision.models import OperationStatusCodes
from azure.cognitiveservices.vision.computervision.models import VisualFeatureTypes
from msrest.authentication import CognitiveServicesCredentials

from PIL import Image
import matplotlib.patches as patches
import matplotlib.pyplot as plt
import numpy as np
import cv2
import glob


subscription_key = "xxxxxxxxxxx"	# your subscription key 
endpoint = "https://xiaxiaoyou-detection.cognitiveservices.azure.cn/"	# your endpoint 

# create computer vision client 
computervision_client = ComputerVisionClient(endpoint, CognitiveServicesCredentials(subscription_key))

# test image
img_file = 'iu.png'

print("===== Detect Objects - local =====")
# Get local image with different objects in it
local_image_objects = open(img_file, "rb")

# Call API with local image
detect_objects_results_local = computervision_client.detect_objects_in_stream(local_image_objects)

object_dict = {}
count = 0
# Print detected objects results with bounding boxes
print("Detecting objects in local image:")
if len(detect_objects_results_local.objects) == 0:
    print("No objects detected.")
else:
    for object in detect_objects_results_local.objects:
        # print(object)
        object_dict[count] = {
            'object_property': object.object_property,
            'rectangle': [(object.rectangle.x, object.rectangle.y), object.rectangle.w, object.rectangle.h],
            'confidence': object.confidence
        }
        count += 1
        print("object at location {}, {}, {}, {}".format(
            object.rectangle.x, object.rectangle.x + object.rectangle.w,
            object.rectangle.y, object.rectangle.y + object.rectangle.h))

# create random color
colors = plt.cm.hsv(np.linspace(0, 1, 10)).tolist()

img = Image.open(img_file)

fig = plt.figure()
plt.imshow(img)
plt.axis('off')
currentAxis = plt.gca()

color_dict = {}
for i, info in enumerate(object_dict.items()):
    print(info)
    color = color_dict.get(info[1]['object_property'], None)
    if not color:
        color = colors[i]
        color_dict[info[1]['object_property']] = color
    rect = patches.Rectangle(*info[1]['rectangle'], edgecolor=color, linewidth=3, fill=False)
    currentAxis.add_patch(rect)
    currentAxis.text(info[1]['rectangle'][0][0], info[1]['rectangle'][0][1], info[1]['object_property'] + ' ' + str(info[1]['confidence']),
                     color='white', size=20, weight='bold', backgroundcolor=color, family='cursive')

# remove blank
fig.set_size_inches(img.size[0]/100, img.size[1]/100)
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top=1, bottom=0, left=0, right=1, hspace=0, wspace=0)
plt.margins(0, 0)
plt.savefig('object-detection-iu.png')
plt.show()

3.2 視頻檢測

??相比圖片檢測，視頻檢測多了兩個視頻轉圖片和圖片轉視頻的操作，具體如下：

def toimg(video_file):
    if not os.path.exists('imgs'):
        os.mkdir('imgs')

    cap = cv2.VideoCapture(video_file)
    isopened = cap.isOpened

    fps = cap.get(cv2.CAP_PROP_FPS)  # 幀率
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    print('fps: {0}, width: {1}, height: {2}'.format(fps, width, height))
    i = 0
    while isopened:
        if i == 1000:
            break
        (flag, frame) = cap.read()  # 讀取每一張 flag frame
        filename = 'imgs/' + str(i).zfill(5) + '.png'
        if flag == True:
            cv2.imwrite(filename, frame, [cv2.IMWRITE_JPEG_QUALITY, 100])
        i += 1
    cap.release()


def tovideo(img_path):
    if not os.path.exists('video'):
        os.mkdir('video')
    fourcc = cv2.VideoWriter_fourcc(*'mp4v')
    videoWrite = cv2.VideoWriter('video/iu.mp4', fourcc, 30, (1920, 1080))
    files = sorted(glob.glob(os.path.join(img_path, '*.png')))
    for file in files:
        img = cv2.imread(file)
        videoWrite.write(img)
    videoWrite.release()

4. 檢測效果

??博主測驗了幾張圖片，效果還是非常不錯的：

結束語

??總體感覺檢測的準確率還是挺高的，雖然不知道Azure基于的什么模型，但就某些細節來看，種類還挺豐富，比如上述圖片的牛仔褲，視頻中IU的唇膏(化妝品)都可以檢測到，而且速度也很棒，可能這就是商用的叭，很不戳！！！！