本文作者： 小森同学
阿里云Elasticsearch客户真实实践分享
文中涉及到的图片特征提取，使用了yongyuan.name的VGGNet库，再此表示感谢！
“图片搜索”是作为导购类网站比较常见的一种功能，其实现的方式有很多，比如“哈西指纹+汉明距离计算”、“特征向量+milvus”，但在实际的应用场景中，要做到快速、精准、简单等特性是比较困难的事情。
“图片搜索”方式优缺点对比
方案三查询效果：
四步搭建“以图搜图”搜索引擎
以下是基于 阿里云 Elasticsearch 6.7 版本，通过安装阿里云 Elasticsearch 向量检索插件【aliyun-knn】 实现，且设计图片向量特征为512维度。
如果自建 Elasticsearch ,是无法使用aliyun-knn插件的,自建建议使用开源 Elasticsearch 7.x版本,并安装fast-elasticsearch-vector-scoring插件(https://github.com/lior-k/fast-elasticsearch-vector-scoring/)
一、 Elasticsearch 索引设计
1.1、索引结构
1. # 创建一个图片索引
2. PUT images_v2
3. {
4. "aliases": {
5. "images": {}
6. },
7. "settings": {
8. "index.codec": "proxima",
9. "index.vector.algorithm": "hnsw",
10. "index.number_of_replicas":1,
11. "index.number_of_shards":3
12. },
13. "mappings": {
14. "_doc": {
15. "properties": {
16. "feature": {
17. "type": "proxima_vector",
18. "dim": 512
19. },
20. "relation_id": {
21. "type": "keyword"
22. },
23. "image_path": {
24. "type": "keyword"
25. }
26. }
27. }
28. }
29. }
1.2、DSL 语句
1. GET images/_search
2.
3. "query": {
4. "hnsw": {
5. "feature": {
6. "vector": [255,....255],
7. "size": 3,
8. "ef": 1
9. }
10. }
11. },
12. "from": 0,
13. "size": 20,
14. "sort": [
15. {
16. "_score": {
17. "order": "desc"
18. }
19. }
20. ],
21. "collapse": {
22. "field": "relation_id"
23. },
24. "_source": {
25. "includes": [
26. "relation_id",
27. "image_path"
28. ]
29. }
二、 图片特征
extract_cnn_vgg16_keras.py
1. # -*- coding: utf-8 -*-
2. # Author: yongyuan.name
3. import numpy as np
4. from numpy import linalg as LA
5. from keras.applications.vgg16 import VGG16
6. from keras.preprocessing import image
7. from keras.applications.vgg16 import preprocess_input
8. from PIL import Image, ImageFile
9. ImageFile.LOAD_TRUNCATED_IMAGES = True
10. class VGGNet:
11. def __init__(self):
12. # weights: 'imagenet'
13. # pooling: 'max' or 'avg'
14. # input_shape: (width, height, 3), width and height should >= 48
15. self.input_shape = (224, 224, 3)
16. self.weight = 'imagenet'
17. self.pooling = 'max'
18. self.model = VGG16(weights = self.weight, input_shape = (self.input_shape[0], self.input_shape[1], self.input_shape[2]), pooling = self.pooling, include_top = False)
19. self.model.predict(np.zeros((1, 224, 224 , 3)))
20. '''
21. Use vgg16 model to extract features
22. Output normalized feature vector
23. '''
24. def extract_feat(self, img_path):
25. img = image.load_img(img_path, target_size=(self.input_shape[0], self.input_shape[1]))
26. img = image.img_to_array(img)
27. img = np.expand_dims(img, axis=0)
28. img = preprocess_input(img)
29. feat = self.model.predict(img)
30. norm_feat = feat[0]/LA.norm(feat[0])
31. return norm_feat
1. # 获取图片特征
2. from extract_cnn_vgg16_keras import VGGNet
3. model = VGGNet()
4. file_path = "./demo.jpg"
5. queryVec = model.extract_feat(file_path)
6. feature = queryVec.tolist()
三、 图片特征写入阿里云 Elasticsearch
helper.py
1. import re
2. import urllib.request
3. def strip(path):
4. """
5. 需要清洗的文件夹名字
6. 清洗掉Windows系统非法文件夹名字的字符串
7. :param path:
8. :return:
9. """
10. path = re.sub(r'[？\\*|“<>:/]', '', str(path))
11. return path
12.
13. def getfilename(url):
14. """
15. 通过url获取最后的文件名
16. :param url:
17. :return:
18. """
19. filename = url.split('/')[-1]
20. filename = strip(filename)
21. return filename
22.
23. def urllib_download(url, filename):
24. """
25. 下载
26. :param url:
27. :param filename:
28. :return:
29. """
30. return urllib.request.urlretrieve(url, filename)
train.py
1. # coding=utf-8
2. import mysql.connector
3. import os
4. from helper import urllib_download, getfilename
5. from elasticsearch5 import Elasticsearch, helpers
6. from extract_cnn_vgg16_keras import VGGNet
7. model = VGGNet()
8. http_auth = ("elastic", "123455")
9. es = Elasticsearch("http://127.0.0.1:9200", http_auth=http_auth)
10. mydb = mysql.connector.connect(
11. host="127.0.0.1", # 数据库主机地址
12. user="root", # 数据库用户名
13. passwd="123456", # 数据库密码
14. database="images"
15. )
16. mycursor = mydb.cursor()
17. imgae_path = "./images/"
18. def get_data(page=1):
19. page_size = 20
20. offset = (page - 1) * page_size
21. sql = """
22. SELECT id, relation_id, photo FROM images LIMIT {0},{1}
23. """
24. mycursor.execute(sql.format(offset, page_size))
25. myresult = mycursor.fetchall()
26. return myresult
27.
28. def train_image_feature(myresult):
29. indexName = "images"
30. photo_path = "http://域名/{0}"
31. actions = []
32. for x in myresult:
33. id = str(x[0])
34. relation_id = x[1]
35. # photo = x[2].decode(encoding="utf-8")
36. photo = x[2]
37. full_photo = photo_path.format(photo)
38. filename = imgae_path + getfilename(full_photo)
39. if not os.path.exists(filename):
40. try:
41. urllib_download(full_photo, filename)
42. except BaseException as e:
43. print("gid:{0}的图片{1}未能下载成功".format(gid, full_photo))
44. continue
45. if not os.path.exists(filename):
46. continue
47. try:
48. feature = model.extract_feat(filename).tolist()
49. action = {
50. "_op_type": "index",
51. "_index": indexName,
52. "_type": "_doc",
53. "_id": id,
54. "_source": {
55. "relation_id": relation_id,
56. "feature": feature,
57. "image_path": photo
58. }
59. }
60. actions.append(action)
61. except BaseException as e:
62. print("id:{0}的图片{1}未能获取到特征".format(id, full_photo))
63. continue
64. # print(actions)
65. succeed_num = 0
66. for ok, response in helpers.streaming_bulk(es, actions):
67. if not ok:
68. print(ok)
69. print(response)
70. else:
71. succeed_num += 1
72. print("本次更新了{0}条数据".format(succeed_num))
73. es.indices.refresh(indexName)
74.
75. page = 1
76. while True:
77. print("当前第{0}页".format(page))
78. myresult = get_data(page=page)
79. if not myresult:
80. print("没有获取到数据了，退出")
81. break
82. train_image_feature(myresult)
83. page += 1
四、 搜索图片
1. import requests
2. import json
3. import os
4. import time
5. from elasticsearch5 import Elasticsearch
6. from extract_cnn_vgg16_keras import VGGNet
7. model = VGGNet()
8. http_auth = ("elastic", "123455")
9. es = Elasticsearch("http://127.0.0.1:9200", http_auth=http_auth)
10. #上传图片保存
11. upload_image_path = "./runtime/"
12. upload_image = request.files.get("image")
13. upload_image_type = upload_image.content_type.split('/')[-1]
14. file_name = str(time.time())[:10] + '.' + upload_image_type
15. file_path = upload_image_path + file_name
16. upload_image.save(file_path)
17. # 计算图片特征向量
18. queryVec = model.extract_feat(file_path)
19. feature = queryVec.tolist()
20. # 删除图片
21. os.remove(file_path)
22. # 根据特征向量去ES中搜索
23. body = {
24. "query": {
25. "hnsw": {
26. "feature": {
27. "vector": feature,
28. "size": 5,
29. "ef": 10
30. }
31. }
32. },
33. # "collapse": {
34. # "field": "relation_id"
35. # },
36. "_source": {"includes": ["relation_id", "image_path"]},
37. "from": 0,
38. "size": 40
39. }
40. indexName = "images"
41. res = es.search(indexName, body=body)
42. # 返回的结果，最好根据自身情况，将得分低的过滤掉...经过测试, 得分在0.65及其以上的,比较符合要求
依赖的包
1. mysql_connector_repackaged
2. elasticsearch
3. Pillow
4. tensorflow
5. requests
6. pandas
7. Keras
8. numpy
总结：
从“用户体验”角度考虑，在可感知层面，速度和精准度决定了产品在用户使用过程中，是否满足“好用”的感觉，通过阿里云 Elasticsearch 向量检索（aliyun-knn）简单四步搭建的“以图搜图”搜索引擎，不仅满足“好用”，同时操作简单一步到位的特征，也加分不少。
相关活动
更多折扣活动，请访问阿里云 Elasticsearch 官网
• 阿里云 Elasticsearch 商业通用版，1核2G首月免费
• 阿里云 Elasticsearch 日志增强版，首月六折，年付六折
• 阿里云 Logstash 2核4G首月免费
如有相关实践的同学，欢迎关注我们公众号投稿
