如何使用 DELF 和 TensorFlow Hub 匹配图像

在 TensorFlow.org 上查看 在 Google Colab 中运行 在 GitHub 中查看源代码 下载笔记本

TensorFlow Hub (TF-Hub) 是一个分享打包在可重用资源(尤其是预训练的模块)中的机器学习专业知识的平台。

在此 Colab 中,我们将使用打包 DELF 神经网络和逻辑的模块来处理图像,从而识别关键点及其描述符。神经网络的权重在地标图像上训练,如这篇论文所述。

设置

pip install -q scikit-image
from absl import logging

import matplotlib.pyplot as plt
import numpy as np
from PIL import Image, ImageOps
from scipy.spatial import cKDTree
from skimage.feature import plot_matches
from skimage.measure import ransac
from skimage.transform import AffineTransform
from six import BytesIO

import tensorflow as tf

import tensorflow_hub as hub
from six.moves.urllib.request import urlopen

数据

在下一个代码单元中,我们指定要使用 DELF 处理的两个图像的网址,以便进行匹配和对比。

Choose images

下载、调整大小、保存并显示图像。

def download_and_resize(name, url, new_width=256, new_height=256):
  path = tf.keras.utils.get_file(url.split('/')[-1], url)
  image = Image.open(path)
  image = ImageOps.fit(image, (new_width, new_height), Image.ANTIALIAS)
  return image
image1 = download_and_resize('image_1.jpg', IMAGE_1_URL)
image2 = download_and_resize('image_2.jpg', IMAGE_2_URL)

plt.subplot(1,2,1)
plt.imshow(image1)
plt.subplot(1,2,2)
plt.imshow(image2)
Downloading data from https://upload.wikimedia.org/wikipedia/commons/2/28/Bridge_of_Sighs%2C_Oxford.jpg
7020544/7013850 [==============================] - 0s 0us/step
Downloading data from https://upload.wikimedia.org/wikipedia/commons/c/c3/The_Bridge_of_Sighs_and_Sheldonian_Theatre%2C_Oxford.jpg
14172160/14164194 [==============================] - 0s 0us/step
<matplotlib.image.AxesImage at 0x7fe64cb63ba8>

png

将 DELF 模块应用到数据

DELF 模块使用一个图像作为输入,并使用向量描述需要注意的点。以下单元包含此 Colab 逻辑的核心。

delf = hub.load('https://tfhub.dev/google/delf/1').signatures['default']
def run_delf(image):
  np_image = np.array(image)
  float_image = tf.image.convert_image_dtype(np_image, tf.float32)

  return delf(
      image=float_image,
      score_threshold=tf.constant(100.0),
      image_scales=tf.constant([0.25, 0.3536, 0.5, 0.7071, 1.0, 1.4142, 2.0]),
      max_feature_num=tf.constant(1000))
result1 = run_delf(image1)
result2 = run_delf(image2)

使用位置和描述向量匹配图像

TensorFlow is not needed for this post-processing and visualization

def match_images(image1, image2, result1, result2):
  distance_threshold = 0.8

  # Read features.
  num_features_1 = result1['locations'].shape[0]
  print("Loaded image 1's %d features" % num_features_1)

  num_features_2 = result2['locations'].shape[0]
  print("Loaded image 2's %d features" % num_features_2)

  # Find nearest-neighbor matches using a KD tree.
  d1_tree = cKDTree(result1['descriptors'])
  _, indices = d1_tree.query(
      result2['descriptors'],
      distance_upper_bound=distance_threshold)

  # Select feature locations for putative matches.
  locations_2_to_use = np.array([
      result2['locations'][i,]
      for i in range(num_features_2)
      if indices[i] != num_features_1
  ])
  locations_1_to_use = np.array([
      result1['locations'][indices[i],]
      for i in range(num_features_2)
      if indices[i] != num_features_1
  ])

  # Perform geometric verification using RANSAC.
  _, inliers = ransac(
      (locations_1_to_use, locations_2_to_use),
      AffineTransform,
      min_samples=3,
      residual_threshold=20,
      max_trials=1000)

  print('Found %d inliers' % sum(inliers))

  # Visualize correspondences.
  _, ax = plt.subplots()
  inlier_idxs = np.nonzero(inliers)[0]
  plot_matches(
      ax,
      image1,
      image2,
      locations_1_to_use,
      locations_2_to_use,
      np.column_stack((inlier_idxs, inlier_idxs)),
      matches_color='b')
  ax.axis('off')
  ax.set_title('DELF correspondences')
match_images(image1, image2, result1, result2)
Loaded image 1's 233 features
Loaded image 2's 262 features
Found 50 inliers

png