【TensorFlow】獲取object detection API訓練模型的輸出座標

　　以下圖，谷歌開源的object detection API提供了五種網絡結構的fine-tuning訓練權重，方便咱們針對目標檢測的需求進行模型訓練，本文詳細介紹下導出訓練模型後，如何得到目標檢測框的座標。若是對使用object detection API訓練模型的過程不瞭解，能夠參考博文：http://www.javashuo.com/article/p-kxnwxjax-cr.html

　　　　　　　　　　　　　　　　　　　　　　　　　　

　　新建一個測試文件object_detection_test.py，該腳本讀取咱們已經訓練好的模型文件和測試圖片，進行測試，代碼以下，

  1 import numpy as np
  2 import os
  3 import six.moves.urllib as urllib
  4 import sys
  5 import tarfile
  6 import tensorflow as tf
  7 import zipfile
  8 
  9 from collections import defaultdict
 10 from io import StringIO
 11 from matplotlib import pyplot as plt
 12 from PIL import Image
 13 
 14 ## This is needed to display the images.
 15 #%matplotlib inline
 16 
 17 # This is needed since the notebook is stored in the object_detection folder.
 18 sys.path.append("..")
 19 
 20 from utils import label_map_util
 21 
 22 from utils import visualization_utils as vis_util
 23 # What model to download.
 24 #MODEL_NAME = 'ssd_mobilenet_v1_coco_2017_11_17'
 25 #MODEL_FILE = MODEL_NAME + '.tar.gz'
 26 #DOWNLOAD_BASE = #'http://download.tensorflow.org/models/object_detection/'
 27 MODEL_NAME = 'data'  # 訓練過程當中保存模型文件的文件夾路徑
 28 
 29 # Path to frozen detection graph. This is the actual model that is used for the object detection.
 30 PATH_TO_CKPT = MODEL_NAME + '/frozen_inference_graph.pb' # 訓練完成導出的pb模型文件
 31 
 32 # List of the strings that is used to add correct label for each box.
 33 PATH_TO_LABELS = 'E:/TensorFlow/Box-object-detection/data/label_map.pbtxt' # label_map.pbtxt文件
 34 
 35 NUM_CLASSES = 2   # 類別總數
 36 
 37 #Load a (frozen) Tensorflow model into memory. 加載模型
 38 detection_graph = tf.Graph()
 39 with detection_graph.as_default():
 40   od_graph_def = tf.GraphDef()
 41   with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
 42     serialized_graph = fid.read()
 43     od_graph_def.ParseFromString(serialized_graph)
 44     tf.import_graph_def(od_graph_def, name='')
 45 #Loading label map 加載label_map
 46 label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
 47 categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)
 48 category_index = label_map_util.create_category_index(categories)
 49 #Helper code
 50 def load_image_into_numpy_array(image):
 51   (im_width, im_height) = image.size
 52   return np.array(image.getdata()).reshape(
 53       (im_height, im_width, 3)).astype(np.uint8)
 54 
 55 # For the sake of simplicity we will use only 2 images:
 56 # image1.jpg
 57 # image2.jpg
 58 # If you want to test the code with your images, just add path to the images to the TEST_IMAGE_PATHS.
 59 PATH_TO_TEST_IMAGES_DIR = 'test_images'   # 測試圖片的路徑
 60 #TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, 3) ]
 61 TEST_IMAGE = sys.argv[1]
 62 print("the test image is:", TEST_IMAGE)
 63 
 64 # Size, in inches, of the output images.
 65 IMAGE_SIZE = (12, 8)
 66 with detection_graph.as_default():
 67   with tf.Session(graph=detection_graph) as sess:
 68     #for image_path in TEST_IMAGE_PATHS:
 69     image = Image.open(TEST_IMAGE)  # 打開圖片
 70     # the array based representation of the image will be used later in order to prepare the
 71     # result image with boxes and labels on it.
 72     image_np = load_image_into_numpy_array(image)
 73     # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
 74     image_np_expanded = np.expand_dims(image_np, axis=0)
 75     image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')   # 獲取圖片張量
 76     # Each box represents a part of the image where a particular object was detected.
 77     boxes = detection_graph.get_tensor_by_name('detection_boxes:0')   # 獲取檢測框張量
 78     # Each score represent how level of confidence for each of the objects.
 79     # Score is shown on the result image, together with the class label.
 80     scores = detection_graph.get_tensor_by_name('detection_scores:0')   # 獲取每一個檢測框的分數，即機率
 81     classes = detection_graph.get_tensor_by_name('detection_classes:0')  # 獲取類別名稱id，與label_map中的ID對應
 82     num_detections = detection_graph.get_tensor_by_name('num_detections:0')  # 獲取檢測總數
 83     # Actual detection.
 84     (boxes, scores, classes, num_detections) = sess.run(  
 85         [boxes, scores, classes, num_detections],
 86         feed_dict={image_tensor: image_np_expanded})
 87     # Visualization of the results of a detection.結果可視化
 88     vis_util.visualize_boxes_and_labels_on_image_array(
 89         image_np,
 90         np.squeeze(boxes),
 91         np.squeeze(classes).astype(np.int32),
 92         np.squeeze(scores),
 93         category_index,
 94         use_normalized_coordinates=True,
 95         line_thickness=8)
 96 
 97     print(boxes) # 打印檢測框座標
 98     print(scores)   #打印每一個檢測框的機率
 99     print(classes)   # 打印檢測框對應的類別
100     print(category_index)  # 打印類別的索引，其是一個嵌套的字典
101    
102     final_score = np.squeeze(scores)    
103     count = 0
104     for i in range(100):
105         if scores is None or final_score[i] > 0.5: # 顯示大於50%機率的檢測框
106             count = count + 1
107     print("the count of objects is: ", count )   
108 
109     plt.figure(figsize=IMAGE_SIZE)
110     plt.imshow(image_np)
111     plt.show()  

打開cmd，輸入以下命令，

python object_detection_test.py ./test_images/2.png

運行結果以下，

目標檢測框box的座標，此處的座標是座標除以相應圖片的長寬所獲得的小數，排列順序爲[ymin , xmin , ymax , xmax]，即box檢測框左上角和右下角的座標，

同時顯示的是目標檢測框box的機率：

Box的標籤索引和每一個索引所表明的標籤，如第一個box的索引爲1,1的標籤名爲「box」，即檢測框裏的是「箱子」

檢測圖：

由於源碼中將座標與圖片的長寬相除，因此顯示的是小數，爲了獲得準確的座標，只要乘上相應的長寬數值就能夠獲得座標了，上圖的檢測圖座標由計算可得

[ymin , xmin , ymax , xmax] = [ 614.4 , 410.4 , 764.16 , 569.16 ]，即在y軸的座標和使用pyplot顯示的座標相近（圖中紅線標出）。

接下來，咱們只要將上面的測試代碼稍加修改便可獲得咱們想要的座標，好比得到每一個檢測物體的中心座標，代碼以下：

  1 import numpy as np
  2 import os
  3 import six.moves.urllib as urllib
  4 import sys
  5 import tarfile
  6 import tensorflow as tf
  7 import zipfile
  8 import time
  9 
 10 from collections import defaultdict
 11 from io import StringIO
 12 from matplotlib import pyplot as plt
 13 #plt.switch_backend('Agg')
 14 from PIL import Image
 15 
 16 ## This is needed to display the images.
 17 #%matplotlib inline
 18 
 19 # This is needed since the notebook is stored in the object_detection folder.
 20 sys.path.append("..")
 21 
 22 from utils import label_map_util
 23 
 24 from utils import visualization_utils as vis_util
 25 # What model to download.
 26 #MODEL_NAME = 'ssd_mobilenet_v1_coco_2017_11_17'
 27 #MODEL_FILE = MODEL_NAME + '.tar.gz'
 28 #DOWNLOAD_BASE = #'http://download.tensorflow.org/models/object_detection/'
 29 MODEL_NAME = 'E:/Project/object-detection-Game-2018-5-31/data-20180607'  # model.ckpt路徑,包括frozen_inference_graph.pb文件
 30 
 31 # Path to frozen detection graph. This is the actual model that is used for the object detection.
 32 PATH_TO_CKPT = MODEL_NAME + '/frozen_inference_graph.pb'
 33 
 34 # List of the strings that is used to add correct label for each box.
 35 PATH_TO_LABELS = MODEL_NAME+'/label_map.pbtxt'
 36 #E:/Project/object-detection-Game-2018-5-31
 37 
 38 NUM_CLASSES = 6
 39 start = time.time()
 40 #Load a (frozen) Tensorflow model into memory.
 41 detection_graph = tf.Graph()
 42 with detection_graph.as_default():
 43   od_graph_def = tf.GraphDef()
 44   #loading ckpt file to graph
 45   with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
 46     serialized_graph = fid.read()
 47     od_graph_def.ParseFromString(serialized_graph)
 48     tf.import_graph_def(od_graph_def, name='')
 49 #Loading label map
 50 label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
 51 categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)
 52 category_index = label_map_util.create_category_index(categories)
 53 #Helper code
 54 def load_image_into_numpy_array(image):
 55   (im_width, im_height) = image.size
 56   return np.array(image.getdata()).reshape(
 57       (im_height, im_width, 3)).astype(np.uint8)
 58 
 59 
 60 # If you want to test the code with your images, just add path to the images to the TEST_IMAGE_PATHS.
 61 #PATH_TO_TEST_IMAGES_DIR = 'test_images'
 62 #TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, 3) ]
 63 TEST_IMAGE = sys.argv[1]
 64 print("the test image is:", TEST_IMAGE)
 65 
 66 # Size, in inches, of the output images.
 67 IMAGE_SIZE = (12, 8)
 68 with detection_graph.as_default():
 69   with tf.Session(graph=detection_graph) as sess:
 70     # Definite input and output Tensors for detection_graph
 71     image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
 72     # Each box represents a part of the image where a particular object was detected.
 73     detection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
 74     # Each score represent how level of confidence for each of the objects.
 75     # Score is shown on the result image, together with the class label.
 76     detection_scores = detection_graph.get_tensor_by_name('detection_scores:0')
 77     detection_classes = detection_graph.get_tensor_by_name('detection_classes:0')
 78     num_detections = detection_graph.get_tensor_by_name('num_detections:0')
 79     #for image_path in TEST_IMAGE_PATHS:
 80     image = Image.open(TEST_IMAGE)
 81     # the array based representation of the image will be used later in order to prepare the
 82     # result image with boxes and labels on it.
 83     image_np = load_image_into_numpy_array(image)
 84     # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
 85     image_np_expanded = np.expand_dims(image_np, axis=0)
 86     image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
 87     # Each box represents a part of the image where a particular object was detected.
 88     boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
 89     # Each score represent how level of confidence for each of the objects.
 90     # Score is shown on the result image, together with the class label.
 91     scores = detection_graph.get_tensor_by_name('detection_scores:0')
 92     classes = detection_graph.get_tensor_by_name('detection_classes:0')
 93     num_detections = detection_graph.get_tensor_by_name('num_detections:0')
 94     # Actual detection.
 95     (boxes, scores, classes, num_detections) = sess.run(
 96         [boxes, scores, classes, num_detections],
 97         feed_dict={image_tensor: image_np_expanded})
 98     # Visualization of the results of a detection.
 99     vis_util.visualize_boxes_and_labels_on_image_array(
100         image_np,
101         np.squeeze(boxes),
102         np.squeeze(classes).astype(np.int32),
103         np.squeeze(scores),
104         category_index,
105         use_normalized_coordinates=True,
106         line_thickness=8)
107 
108     #print(boxes)
109     # for i in range(len(scores[0])):
110     #     if scores[0][i]>0.5:
111     #         print(scores[0][i])
112     #print(scores)      
113     #print(classes)
114     #print(category_index)
115     final_score = np.squeeze(scores)    
116     count = 0
117     for i in range(100):
118         if scores is None or final_score[i] > 0.5:
119             count = count + 1
120     print()
121     print("the count of objects is: ", count )
122     (im_width, im_height) = image.size
123     for i in range(count):
124         #print(boxes[0][i])
125         y_min = boxes[0][i][0]*im_height
126         x_min = boxes[0][i][1]*im_width
127         y_max = boxes[0][i][2]*im_height
128         x_max = boxes[0][i][3]*im_width
129         print("object{0}: {1}".format(i,category_index[classes[0][i]]['name']),
130                          ',Center_X:',int((x_min+x_max)/2),',Center_Y:',int((y_min+y_max)/2))
131         #print(x_min,y_min,x_max,y_max)
132     end = time.time()
133     seconds = end - start
134     print("Time taken : {0} seconds".format(seconds))
135 
136     # plt.figure(figsize=IMAGE_SIZE)
137     # plt.imshow(image_np)
138     # plt.show()

運行結果以下，

 

轉載請註明出處：http://www.javashuo.com/article/p-mobuytjy-dc.html