TensorFlowのPythonAPIで画像分析する際に抽出できる情報について

前提・実現したいこと

TensorFlowのInception-v3で画像を分類してみた(Python API編)にしたがって、
TensorFlowで画像分析をしようとしています。
実行環境はGPUではなくCPUです。

発生している問題・エラーメッセージ

TensorFlowの公式サイトに書かれているように、
PythonAPIをGitHubからcloneして実行すると
サンプルのパンダの画像では以下の結果が得られるのですが、上位5件だけでなく、
画像の色など他の属性（パンダの場合は白と黒など）も抽出することが可能かのか、
そして可能であればどのようにすればいいのか知りたいです。

bash
1classify_image.py

結果

giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca (score = 0.88493)
indri, indris, Indri indri, Indri brevicaudatus (score = 0.00878)
lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens (score = 0.00317)
custard apple (score = 0.00149)
earthstar (score = 0.00127)

試したこと

classify_image.pyの中身をcatで確認しました。

    node_lookup = NodeLookup()

    top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]
    for node_id in top_k:
      human_string = node_lookup.id_to_string(node_id)
      score = predictions[node_id]
      print('%s (score = %.5f)' % (human_string, score))
....
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

この辺りが関係ありそうだと感じたのですが、APIなのでどこまで手を加えられるのか分かっていない状態です・

python
1from __future__ import absolute_import
2from __future__ import division
3from __future__ import print_function
4
5import argparse
6import os.path
7import re
8import sys
9import tarfile
10
11import numpy as np
12from six.moves import urllib
13import tensorflow as tf
14
15FLAGS = None
16
17# pylint: disable=line-too-long
18DATA_URL = 'http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz'
19# pylint: enable=line-too-long
20
21
22class NodeLookup(object):
23  """Converts integer node ID's to human readable labels."""
24
25  def __init__(self,
26               label_lookup_path=None,
27               uid_lookup_path=None):
28    if not label_lookup_path:
29      label_lookup_path = os.path.join(
30          FLAGS.model_dir, 'imagenet_2012_challenge_label_map_proto.pbtxt')
31    if not uid_lookup_path:
32      uid_lookup_path = os.path.join(
33          FLAGS.model_dir, 'imagenet_synset_to_human_label_map.txt')
34    self.node_lookup = self.load(label_lookup_path, uid_lookup_path)
35
36  def load(self, label_lookup_path, uid_lookup_path):
37    """Loads a human readable English name for each softmax node.
38
39    Args:
40      label_lookup_path: string UID to integer node ID.
41      uid_lookup_path: string UID to human-readable string.
42
43    Returns:
44      dict from integer node ID to human-readable string.
45    """
46    if not tf.gfile.Exists(uid_lookup_path):
47      tf.logging.fatal('File does not exist %s', uid_lookup_path)
48    if not tf.gfile.Exists(label_lookup_path):
49      tf.logging.fatal('File does not exist %s', label_lookup_path)
50
51    # Loads mapping from string UID to human-readable string
52    proto_as_ascii_lines = tf.gfile.GFile(uid_lookup_path).readlines()
53    uid_to_human = {}
54    p = re.compile(r'[n\d]*[ \S,]*')
55    for line in proto_as_ascii_lines:
56      parsed_items = p.findall(line)
57      uid = parsed_items[0]
58      human_string = parsed_items[2]
59      uid_to_human[uid] = human_string
60
61    # Loads mapping from string UID to integer node ID.
62    node_id_to_uid = {}
63    proto_as_ascii = tf.gfile.GFile(label_lookup_path).readlines()
64    for line in proto_as_ascii:
65      if line.startswith('  target_class:'):
66        target_class = int(line.split(': ')[1])
67      if line.startswith('  target_class_string:'):
68        target_class_string = line.split(': ')[1]
69        node_id_to_uid[target_class] = target_class_string[1:-2]
70
71    # Loads the final mapping of integer node ID to human-readable string
72    node_id_to_name = {}
73    for key, val in node_id_to_uid.items():
74      if val not in uid_to_human:
75        tf.logging.fatal('Failed to locate: %s', val)
76      name = uid_to_human[val]
77      node_id_to_name[key] = name
78
79    return node_id_to_name
80
81  def id_to_string(self, node_id):
82    if node_id not in self.node_lookup:
83      return ''
84    return self.node_lookup[node_id]
85
86
87def create_graph():
88  """Creates a graph from saved GraphDef file and returns a saver."""
89  # Creates graph from saved graph_def.pb.
90  with tf.gfile.FastGFile(os.path.join(
91      FLAGS.model_dir, 'classify_image_graph_def.pb'), 'rb') as f:
92    graph_def = tf.GraphDef()
93    graph_def.ParseFromString(f.read())
94    _ = tf.import_graph_def(graph_def, name='')
95
96
97def run_inference_on_image(image):
98  """Runs inference on an image.
99
100  Args:
101    image: Image file name.
102
103  Returns:
104    Nothing
105  """
106  if not tf.gfile.Exists(image):
107    tf.logging.fatal('File does not exist %s', image)
108  image_data = tf.gfile.FastGFile(image, 'rb').read()
109
110  # Creates graph from saved GraphDef.
111  create_graph()
112
113  with tf.Session() as sess:
114    # Some useful tensors:
115    # 'softmax:0': A tensor containing the normalized prediction across
116    #   1000 labels.
117    # 'pool_3:0': A tensor containing the next-to-last layer containing 2048
118    #   float description of the image.
119    # 'DecodeJpeg/contents:0': A tensor containing a string providing JPEG
120    #   encoding of the image.
121    # Runs the softmax tensor by feeding the image_data as input to the graph.
122    softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')
123    predictions = sess.run(softmax_tensor,
124                           {'DecodeJpeg/contents:0': image_data})
125    predictions = np.squeeze(predictions)
126
127    # Creates node ID --> English string lookup.
128    node_lookup = NodeLookup()
129
130    top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]
131    for node_id in top_k:
132      human_string = node_lookup.id_to_string(node_id)
133      score = predictions[node_id]
134      print('%s (score = %.5f)' % (human_string, score))
135
136
137def maybe_download_and_extract():
138  """Download and extract model tar file."""
139  dest_directory = FLAGS.model_dir
140  if not os.path.exists(dest_directory):
141    os.makedirs(dest_directory)
142  filename = DATA_URL.split('/')[-1]
143  filepath = os.path.join(dest_directory, filename)
144  if not os.path.exists(filepath):
145    def _progress(count, block_size, total_size):
146      sys.stdout.write('\r>> Downloading %s %.1f%%' % (
147          filename, float(count * block_size) / float(total_size) * 100.0))
148      sys.stdout.flush()
149    filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, _progress)
150    print()
151    statinfo = os.stat(filepath)
152    print('Successfully downloaded', filename, statinfo.st_size, 'bytes.')
153  tarfile.open(filepath, 'r:gz').extractall(dest_directory)
154
155
156def main(_):
157  maybe_download_and_extract()
158  image = (FLAGS.image_file if FLAGS.image_file else
159           os.path.join(FLAGS.model_dir, 'cropped_panda.jpg'))
160  run_inference_on_image(image)
161
162
163if __name__ == '__main__':
164  parser = argparse.ArgumentParser()
165  # classify_image_graph_def.pb:
166  #   Binary representation of the GraphDef protocol buffer.
167  # imagenet_synset_to_human_label_map.txt:
168  #   Map from synset ID to a human readable string.
169  # imagenet_2012_challenge_label_map_proto.pbtxt:
170  #   Text representation of a protocol buffer mapping a label to synset ID.
171  parser.add_argument(
172      '--model_dir',
173      type=str,
174      default='/tmp/imagenet',
175      help="""\
176      Path to classify_image_graph_def.pb,
177      imagenet_synset_to_human_label_map.txt, and
178      imagenet_2012_challenge_label_map_proto.pbtxt.\
179      """
180  )
181  parser.add_argument(
182      '--image_file',
183      type=str,
184      default='',
185      help='Absolute path to image file.'
186  )
187  parser.add_argument(
188      '--num_top_predictions',
189      type=int,
190      default=5,
191      help='Display this many predictions.'
192  )
193  FLAGS, unparsed = parser.parse_known_args()
194  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)