Mask R-CNN Trong Bài Toán Nhận Dạng Và Phân Vùng Đối Tượng

 1import numpy as np
 2import os
 3import sys
 4import tarfile
 5import tensorflow as tf
 6
 7from collections import defaultdict
 8from io import StringIO
 9from matplotlib import pyplot as plt
10from PIL import Image
11import PIL.ImageDraw as ImageDraw
12import PIL.ImageFont as ImageFont
13import cv2
14
15import pprint
16
17import PIL.Image as Image
18import PIL.ImageColor as ImageColor
19
20# Model preparation
21
22
23# Path to frozen detection graph. This is the actual model that is used for the object detection.
24PATH_TO_CKPT = 'mask_rcnn_inception_v2_coco_2018_01_28' + '/frozen_inference_graph.pb'
25
26# List of the strings that is used to add correct label for each box.
27#PATH_TO_LABELS = 'mscoco_label_map.pbtxt'
28
29NUM_CLASSES = 1
30
31
32# categories
33
34category_index = {1: {'id': 1, 'name': 'person'},
35# 3: {'id': 3, 'name': 'car'},
36 }
37
38detection_graph = tf.Graph()
39with 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='')

 1category_index = {1: {'id': 1, 'name': 'person'},
 2 2: {'id': 2, 'name': 'bicycle'},
 3 3: {'id': 3, 'name': 'car'},
 4 4: {'id': 4, 'name': 'motorcycle'},
 5 5: {'id': 5, 'name': 'airplane'},
 6 6: {'id': 6, 'name': 'bus'},
 7 7: {'id': 7, 'name': 'train'},
 8 8: {'id': 8, 'name': 'truck'},
 9 9: {'id': 9, 'name': 'boat'},
10 10: {'id': 10, 'name': 'traffic light'},
11 11: {'id': 11, 'name': 'fire hydrant'},
12 13: {'id': 13, 'name': 'stop sign'},
13 14: {'id': 14, 'name': 'parking meter'},
14 15: {'id': 15, 'name': 'bench'},
15 16: {'id': 16, 'name': 'bird'},
16 17: {'id': 17, 'name': 'cat'},
17 18: {'id': 18, 'name': 'dog'},
18 19: {'id': 19, 'name': 'horse'},
19 20: {'id': 20, 'name': 'sheep'},
20 21: {'id': 21, 'name': 'cow'},
21 22: {'id': 22, 'name': 'elephant'},
22 23: {'id': 23, 'name': 'bear'},
23 24: {'id': 24, 'name': 'zebra'},
24 25: {'id': 25, 'name': 'giraffe'},
25 27: {'id': 27, 'name': 'backpack'},
26 28: {'id': 28, 'name': 'umbrella'},
27 31: {'id': 31, 'name': 'handbag'},
28 32: {'id': 32, 'name': 'tie'},
29 33: {'id': 33, 'name': 'suitcase'},
30 34: {'id': 34, 'name': 'frisbee'},
31 35: {'id': 35, 'name': 'skis'},
32 36: {'id': 36, 'name': 'snowboard'},
33 37: {'id': 37, 'name': 'sports ball'},
34 38: {'id': 38, 'name': 'kite'},
35 39: {'id': 39, 'name': 'baseball bat'},
36 40: {'id': 40, 'name': 'baseball glove'},
37 41: {'id': 41, 'name': 'skateboard'},
38 42: {'id': 42, 'name': 'surfboard'},
39 43: {'id': 43, 'name': 'tennis racket'},
40 44: {'id': 44, 'name': 'bottle'},
41 46: {'id': 46, 'name': 'wine glass'},
42 47: {'id': 47, 'name': 'cup'},
43 48: {'id': 48, 'name': 'fork'},
44 49: {'id': 49, 'name': 'knife'},
45 50: {'id': 50, 'name': 'spoon'},
46 51: {'id': 51, 'name': 'bowl'},
47 52: {'id': 52, 'name': 'banana'},
48 53: {'id': 53, 'name': 'apple'},
49 54: {'id': 54, 'name': 'sandwich'},
50 55: {'id': 55, 'name': 'orange'},
51 56: {'id': 56, 'name': 'broccoli'},
52 57: {'id': 57, 'name': 'carrot'},
53 58: {'id': 58, 'name': 'hot dog'},
54 59: {'id': 59, 'name': 'pizza'},
55 60: {'id': 60, 'name': 'donut'},
56 61: {'id': 61, 'name': 'cake'},
57 62: {'id': 62, 'name': 'chair'},
58 63: {'id': 63, 'name': 'couch'},
59 64: {'id': 64, 'name': 'potted plant'},
60 65: {'id': 65, 'name': 'bed'},
61 67: {'id': 67, 'name': 'dining table'},
62 70: {'id': 70, 'name': 'toilet'},
63 72: {'id': 72, 'name': 'tv'},
64 73: {'id': 73, 'name': 'laptop'},
65 74: {'id': 74, 'name': 'mouse'},
66 75: {'id': 75, 'name': 'remote'},
67 76: {'id': 76, 'name': 'keyboard'},
68 77: {'id': 77, 'name': 'cell phone'},
69 78: {'id': 78, 'name': 'microwave'},
70 79: {'id': 79, 'name': 'oven'},
71 80: {'id': 80, 'name': 'toaster'},
72 81: {'id': 81, 'name': 'sink'},
73 82: {'id': 82, 'name': 'refrigerator'},
74 84: {'id': 84, 'name': 'book'},
75 85: {'id': 85, 'name': 'clock'},
76 86: {'id': 86, 'name': 'vase'},
77 87: {'id': 87, 'name': 'scissors'},
78 88: {'id': 88, 'name': 'teddy bear'},
79 89: {'id': 89, 'name': 'hair drier'},
80 90: {'id': 90, 'name': 'toothbrush'}}

  1
  2    draw  =  ImageDraw.Draw(image)
  3    im_width,  im_height  =  image.size
  4    if  use_normalized_coordinates:
  5        (left,  right,  top,  bottom)  =  (xmin  *  im_width,  xmax  *  im_width,
  6                                                                    ymin  *  im_height,  ymax  *  im_height)
  7    else:
  8        (left,  right,  top,  bottom)  =  (xmin,  xmax,  ymin,  ymax)
  9    draw.line([(left,  top),  (left,  bottom),  (right,  bottom),
 10                          (right,  top),  (left,  top)],  width=thickness,  fill=color)
 11    try:
 12        font  =  ImageFont.truetype('arial.ttf',  24)
 13    except  IOError:
 14        font  =  ImageFont.load_default()
 15
 16    #  If  the  total  height  of  the  display  strings  added  to  the  top  of  the  bounding
 17    #  box  exceeds  the  top  of  the  image,  stack  the  strings  below  the  bounding  box
 18    #  instead  of  above.
 19    display_str_heights  =  [font.getsize(ds)[1]  for  ds  in  display_str_list]
 20    #  Each  display_str  has  a  top  and  bottom  margin  of  0.05x.
 21    total_display_str_height  =  (1  +  2  *  0.05)  *  sum(display_str_heights)
 22
 23    if  top  >  total_display_str_height:
 24        text_bottom  =  top
 25    else:
 26        text_bottom  =  bottom  +  total_display_str_height
 27    #  Reverse  list  and  print  from  bottom  to  top.
 28    for  display_str  in  display_str_list[::-1]:
 29        text_width,  text_height  =  font.getsize(display_str)
 30        margin  =  np.ceil(0.05  *  text_height)
 31        draw.rectangle(
 32                [(left,  text_bottom  -  text_height  -  2  *  margin),  (left  +  text_width,
 33                                                                                                                    text_bottom)],
 34                fill=color)
 35        draw.text(
 36                (left  +  margin,  text_bottom  -  text_height  -  margin),
 37                display_str,
 38                fill='black',
 39                font=font)
 40        text_bottom  -=  text_height  -  2  *  margin
 41
 42
 43
 44def visualize_boxes_and_labels_on_image_array(
 45        image,
 46        boxes,
 47        classes,
 48        scores,
 49        category_index,
 50        instance_masks=None,
 51        instance_boundaries=None,
 52        keypoints=None,
 53        use_normalized_coordinates=False,
 54        max_boxes_to_draw=20,
 55        min_score_thresh=.5,
 56        agnostic_mode=False,
 57        line_thickness=4,
 58        groundtruth_box_visualization_color='black',
 59        skip_scores=False,
 60        skip_labels=False):
 61
 62    box_to_display_str_map = collections.defaultdict(list)
 63    box_to_color_map = collections.defaultdict(str)
 64    box_to_instance_masks_map = {}
 65    box_to_instance_boundaries_map = {}
 66    box_to_keypoints_map = collections.defaultdict(list)
 67    if not max_boxes_to_draw:
 68        max_boxes_to_draw = boxes.shape[0]
 69    #print(boxes)
 70    for i in range(min(max_boxes_to_draw, boxes.shape[0])):
 71        if scores is None or scores[i] > min_score_thresh:
 72            box = tuple(boxes[i].tolist())
 73        if instance_masks is not None:
 74            box_to_instance_masks_map[box] = instance_masks[i]
 75        if instance_boundaries is not None:
 76            box_to_instance_boundaries_map[box] = instance_boundaries[i]
 77        if keypoints is not None:
 78            box_to_keypoints_map[box].extend(keypoints[i])
 79        if scores is None:
 80            box_to_color_map[box] = groundtruth_box_visualization_color
 81        else:
 82            display_str = ''
 83            if not skip_labels:
 84                if not agnostic_mode:
 85                    if classes[i] in category_index.keys():
 86                        class_name = category_index[classes[i]]['name']
 87                    else:
 88                        class_name = 'N/A'
 89                    display_str = str(class_name)
 90            if not skip_scores:
 91                if not display_str:
 92                    display_str = '{}%'.format(int(100 * scores[i]))
 93                else:
 94                    display_str = '{}: {}%'.format(
 95                        display_str, int(100 * scores[i]))
 96            box_to_display_str_map[box].append(display_str)
 97            if agnostic_mode:
 98                box_to_color_map[box] = 'DarkOrange'
 99            else:
100                box_to_color_map[box] = STANDARD_COLORS[classes[i] %
101                                                        len(STANDARD_COLORS)]
102
103    # Draw all boxes onto image.
104    for box, color in box_to_color_map.items():
105        ymin, xmin, ymax, xmax = box
106        if instance_masks is not None:
107            draw_mask_on_image_array(image, box_to_instance_masks_map[box], color=color)
108
109        draw_bounding_box_on_image_array(
110        image,
111        ymin,
112        xmin,
113        ymax,
114        xmax,
115        color=color,
116        thickness=line_thickness,
117        display_str_list=box_to_display_str_map[box],
118        use_normalized_coordinates=use_normalized_coordinates)
119
120    return image
121
122
123def reframe_box_masks_to_image_masks(box_masks,  boxes,  image_height,
124                                     image_width):
125    """Transforms  the  box  masks  back  to  full  image  masks.
126
127    Embeds  masks  in  bounding  boxes  of  larger  masks  whose  shapes  correspond  to
128    image  shape.
129
130    Args:
131        box_masks:  A  tf.float32  tensor  of  size  [num_masks,  mask_height,  mask_width].
132        boxes:  A  tf.float32  tensor  of  size  [num_masks,  4]  containing  the  box
133                      corners.  Row  i  contains  [ymin,  xmin,  ymax,  xmax]  of  the  box
134                      corresponding  to  mask  i.  Note  that  the  box  corners  are  in
135                      normalized  coordinates.
136        image_height:  Image  height.  The  output  mask  will  have  the  same  height  as
137                                    the  image  height.
138        image_width:  Image  width.  The  output  mask  will  have  the  same  width  as  the
139                                  image  width.
140
141    Returns:
142        A  tf.float32  tensor  of  size  [num_masks,  image_height,  image_width].
143    """
144    #  TODO(rathodv):  Make  this  a  public  function.
145    def reframe_box_masks_to_image_masks_default():
146        """The  default  function  when  there  are  more  than  0  box  masks."""
147        def transform_boxes_relative_to_boxes(boxes,  reference_boxes):
148            boxes = tf.reshape(boxes,  [-1,  2,  2])
149            min_corner = tf.expand_dims(reference_boxes[:,  0:2],  1)
150            max_corner = tf.expand_dims(reference_boxes[:,  2:4],  1)
151            transformed_boxes = (boxes - min_corner) / \
152                (max_corner - min_corner)
153            return tf.reshape(transformed_boxes,  [-1,  4])
154
155        box_masks_expanded = tf.expand_dims(box_masks,  axis=3)
156        num_boxes = tf.shape(box_masks_expanded)[0]
157        unit_boxes = tf.concat(
158            [tf.zeros([num_boxes,  2]),  tf.ones([num_boxes,  2])],  axis=1)
159        reverse_boxes = transform_boxes_relative_to_boxes(unit_boxes,  boxes)
160        return tf.image.crop_and_resize(
161            image=box_masks_expanded,
162            boxes=reverse_boxes,
163            box_ind=tf.range(num_boxes),
164            crop_size=[image_height,  image_width],
165            extrapolation_value=0.0)
166    image_masks = tf.cond(
167        tf.shape(box_masks)[0] > 0,
168        reframe_box_masks_to_image_masks_default,
169        lambda:  tf.zeros([0,  image_height,  image_width,  1],  dtype=tf.float32))
170    return tf.squeeze(image_masks,  axis=3)

 1
 2def detect_frame(image_np, sess, detection_graph):
 3
 4    with detection_graph.as_default():
 5
 6        ops = tf.get_default_graph().get_operations()
 7        all_tensor_names = {output.name for op in ops for output in op.outputs}
 8        tensor_dict = {}
 9        for key in [
10            'num_detections', 'detection_boxes', 'detection_scores',
11            'detection_classes', 'detection_masks'
12        ]:
13            tensor_name = key + ':0'
14            if tensor_name in all_tensor_names:
15                tensor_dict[key] = tf.get_default_graph(
16                ).get_tensor_by_name(tensor_name)
17        if 'detection_masks' in tensor_dict:
18            # The following processing is only for single image
19            detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])
20            detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])
21            # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
22            real_num_detection = tf.cast(
23                tensor_dict['num_detections'][0], tf.int32)
24
25            detection_boxes = tf.slice(detection_boxes, [0, 0], [
26                                       real_num_detection, -1])
27            detection_masks = tf.slice(detection_masks, [0, 0, 0], [
28                                       real_num_detection, -1, -1])
29            detection_masks_reframed = reframe_box_masks_to_image_masks(
30                detection_masks, detection_boxes, image_np.shape[0], image_np.shape[1])
31            detection_masks_reframed = tf.cast(
32                tf.greater(detection_masks_reframed, 0.5), tf.uint8)
33            # Follow the convention by adding back the batch dimension
34            tensor_dict['detection_masks'] = tf.expand_dims(
35                detection_masks_reframed, 0)
36        image_tensor = tf.get_default_graph().get_tensor_by_name('image_tensor:0')
37
38      # Run inference
39        output_dict = sess.run(tensor_dict,
40                               feed_dict={image_tensor: np.expand_dims(image_np, 0)})
41
42      # all outputs are float32 numpy arrays, so convert types as appropriate
43        output_dict['num_detections'] = int(output_dict['num_detections'][0])
44        #print("num detect "+str(output_dict['num_detections']))
45        output_dict['detection_classes'] = output_dict['detection_classes'][0].astype(
46            np.uint8)
47        output_dict['detection_boxes'] = output_dict['detection_boxes'][0]
48        output_dict['detection_scores'] = output_dict['detection_scores'][0]
49        if 'detection_masks' in output_dict:
50            output_dict['detection_masks'] = output_dict['detection_masks'][0]
51
52        visualize_boxes_and_labels_on_image_array(
53            image_np,
54            output_dict['detection_boxes'],
55            output_dict['detection_classes'],
56            output_dict['detection_scores'],
57            category_index,
58            instance_masks=output_dict.get('detection_masks'),
59            use_normalized_coordinates=True,
60            line_thickness=1,
61            max_boxes_to_draw=min(output_dict['num_detections'],20)
62            )
63
64    return image_np

1image = cv2.imread('img2.jpg')
2with detection_graph.as_default():
3    with tf.Session(graph=detection_graph) as sess:
4        image_np = detect_frame(image, sess, detection_graph)
5
6cv2.imwrite('output.jpg', image)