自制AI图像搜索引擎[笔记]

一些资料：

谷歌搜索关键字：Deep learning web image search engine  github

pratheeksh/Deep-Image-Search-Engine

https://github.com/pratheeksh/Deep-Image-Search-Engine

来自于课程：NYU Courant课程

https://cs.nyu.edu/courses/spring17/CSCI-GA.3033-006/

https://github.com/sethuiyer/Image-to-Image-search 

https://github.com/matsui528/sis 

https://github.com/sVujke/img_classification_deep_learning 

https://github.com/paucarre/tiefvision 

今天看到一个书本推送，挺有意思，图搜

就是以图搜图

《自制AI图像搜索引擎》明恒毅 著

图像搜索引擎有两种实现方式—基于图像上下文文本特征的方式和基于图像视觉内容特征的方式。本书所指的图像搜索引擎是基于内容特征的图像检索，也就是通常所说的“以图搜图”来检索相似图片。本书主要讲解了搜索引擎技术的发展脉络、文本搜索引擎的基本原理和搜索引擎的一般结构，详细讲述了图像搜索引擎各主要组成部分的原理和实现，并构建了一个基于深度学习的Web图像搜索引擎。

第 1章 从文本搜索到图像搜索 1 
1．1　文本搜索引擎的发展　1 
1．2　文本搜索引擎的结构与实现　2 
1．2．1　文本预处理　3 
1．2．2　建立索引　5 
1．2．3　对索引进行搜索　7 
1．3　搜索引擎的一般结构　10 
1．4　从文本到图像　10 
1．5　现有图像搜索引擎介绍　12 
1．5．1　Google图像搜索引擎　12 
1．5．2　百度图像搜索引擎　13 
1．5．3　TinEye图像搜索引擎　14 
1．5．4　淘宝图像搜索引擎　15 
1．6　本章小结　16 
第　2章 传统图像特征提取　17 
2．1　人类怎样获取和理解一幅图像　17 
2．2　计算机怎样获取和表示一幅图像　18 
2．2．1　采样　18 
2．2．2　量化　19 
2．2．3　数字图像的存储　19 
2．2．4　常用的位图格式　20 
2．2．5　色彩空间　20 
2．2．6　图像基本操作　21 
2．3　图像特征的分类　29 
2．4　全局特征　30 
2．4．1　颜色特征　30 
2．4．2　纹理特征　41 
2．4．3　形状特征　67 
2．5　局部特征　82 
2．5．1　SIFT描述符　82 
2．5．2　SURF描述符　86 
2．6　本章小结　88 
第3章　深度学习图像特征提取　89 
3．1　深度学习　89 
3．1．1　神经网络的发展　89 
3．1．2　深度神经网络的突破　92 
3．1．3　主要的深度神经网络模型　95 
3．2　深度学习应用框架　97 
3．2．1　TensorFlow　97 
3．2．2　Torch　98 
3．2．3　Caffe　98 
3．2．4　Theano　98 
3．2．5　Keras　99 
3．2．6　DeepLearning4J　99 
3．3　卷积神经网络　99 
3．3．1　卷积　99 
3．3．2　卷积神经网络概述　103 
3．3．3　经典卷积神经网络结构　110 
3．3．4　使用卷积神经网络提取图像特征　130 
3．3．5　使用迁移学习和微调技术进一步提升提取特征的精度　134 
3．4　本章小结　141 
第4章　图像特征索引与检索　142 
4．1　图像特征降维　142 
4．1．1　主成分分析算法降维　142 
4．1．2　深度自动编码器降维　150 
4．2　图像特征标准化　153 
4．2．1　离差标准化　153 
4．2．2　标准差标准化　153 
4．3　图像特征相似度的度量　154 
4．3．1　欧氏距离　154 
4．3．2　曼哈顿距离　155 
4．3．3　海明距离　155 
4．3．4　余弦相似度　155 
4．3．5　杰卡德相似度　156 
4．4　图像特征索引与检索　157 
4．4．1　从最近邻（NN）到K最近邻（KNN）　157 
4．4．2　索引构建与检索　158 
4．5　本章小结　173 
第5章　构建一个基于深度学习的Web图像搜索引擎　174 
5．1　架构分析与技术路线　174 
5．1．1　架构分析　174 
5．1．2　技术路线　175 
5．2　程序实现　175 
5．2．1　开发环境搭建　175 
5．2．2　项目实现　176 
5．3　优化策略　204 
5．4　本章小结　205

这个书的最后，会提供一个例子，基于web+java+深度学习的图搜Demo

这里我也找了一些图搜的工程：

图像搜索引擎

https://blog.csdn.net/real_myth/article/details/45576319

他是转载自维基百科：

https://en.wikipedia.org/w/index.php?title=List_of_CBIR_engines&oldid=661221480

CBIR research projects/demos/open source projects

Name	Description	External Image Query	Metadata Query	Index Size (Estimate, Millions of Images)	Organization Type	License (Open/Closed)
akiwi	akiwi is a semi-automatic image keywording tool using CBIR techniques. It was developed by HTW Berlin / pixolution GmbH	Yes	Yes	15M	University	Closed
ALIPR	Developed by Penn State University researchers	Yes	Yes		University	Closed
Anaktisi	This Web-Solution implements a new family of CBIR descriptors. These descriptors combine in one histogram color and texture information and are suitable for accurately retrieving images.	Yes	No	0.225M	University	Open
BRISC	BRISC is a recursive acronym for BRISC Really IS Cool, and is (conveniently enough) also an anagram of Content-Based Image Retrieval System.	Yes	No		University	GPL
digiKam	Extensive photo management application build on top of KDE libraries. It provides, besides many other features, reverse searches for images in the local collection, detection of duplicates and a fuzzy search by drawings.	Yes	Yes	Desktop-based	KDE	GPL
Caliph & Emir	Creation and Retrieval of images based on MPEG-7.	Yes	No	Desktop-based	University	GPL
FIRE	Open source query by visual example CBIR system. Developed at RWTH Aachen University. FIRE is a research system developed with extensibility in mind and can easily be combined with textual information retrieval systems.	No	No		University	Open
GNU Image Finding Tool	Query by example image search system.	Yes	No	Desktop-based	GNU	GPL
ISSBP	Similar Image Search by Imense plugin for Adobe Bridge, free beta.	Yes	Yes	free-beta limited to 4k images	Private Company	Closed
img(Rummager)	Image retrieval Engine (Freeware Application).	Yes	No	Desktop-based	Individual	Closed
imgSeek	photo collection manager and viewer with content-based search and many other features.	Yes	No		Individual	GPL
IKONA	Generic CBIR system - INRIA - IMEDIA	Yes	Yes		University	Closed
IOSB	Image retrieval demonstration software of Fraunhofer IOSB (Germany)	Yes	No	Desktop-based	Research Institute	Closed
LIRE	Java GPL library for content based image retrieval based on Lucene including multiple low level global and local features and different indexing strategies including bag of visual words and hashing.	Yes	Yes		University	GPL
Lucignolo	Image similarity search engine using only the native full-text search engine Lucene.	Yes	Yes	106M	Research Institute	Closed
MIFile	Image similarity search engine based on MI File (Metric Inverted File) developed at ISTI-CNR. Source code of the MI File.	No	No	106M	Research Institute	Open
MUVIS	CBIR System at TUT- Tampere University of Technology.	Yes	No	Desktop-based	University	Closed
Pastec	C++ LGPL index and search engine for near-duplicate image retrieval that uses bag of visual words with ORB features.	Yes	Yes		Private company	LGPL
PIRIA	CBIR tool developed at CEA-LIST, LVIC (Vision and Content Engineering Laboratory).	Yes	Yes	130M	University	Closed
PicsLikeThat	Image search using visual similarity search and sorting combined with a recommender system. (Cooperation of pixolution GmbH, fotolia and HTW Berlin)	No	No	12M	University	Closed
Pixcavator	Similar image search based on topological image analysis	Yes	No	Desktop-based	Private company	Closed
QuickLook	Visual information retrieval system with relevance feedback	No	Yes		University	Closed
RETIN	Interactive images retrieval system - CNRS - ETIS Lab., MIDI Team	No	No		University	Closed
Retrievr	Search and explore in a selection of Flickr images by drawing a rough sketch or uploading an image.	No	No		University	Closed
SIMBA	demo of system by the Albert-Ludwigs-Universitet Freiburg (Germany) Inst. for Pattern Recognition and Image Processing	Yes	No	0.002M	University	Closed
TagProp	The demonstration of image annotation tool TagProp in ICCV2009 for image set: Corel 5k ESP Game IAPR TC-12 and MIR Flickr.	No	Yes		Institute	Closed
VIRaL	Visual Image Retrieval and Localization: A visual search engine that, given a query image, retrieves photos depicting the same object or scene under varying viewpoint or lighting conditions. Using Flickr photos of urban scenes, it automatically estimates where a picture is taken, suggests tags, identifies known landmarks or points of interest, and links to relevant Wikipedia articles. It currently supports 39 cities around the world.	Yes	Yes	2.221M	University	Closed
Windsurf	A general framework for efficiently processing content-based image queries with particular emphasis to the region-based paradigm; it provides an environment where different alternatives of the paradigm can be implemented, allowing such implementations to be compared on a fair basis, from the points of view of both effectiveness and efficiency.	Yes	No		University	Open but not free
PIBE	An adaptive image browsing system that provides users with an intuitive, easy-to-use, structured view of an image collection and complements it with ideas from the field of adaptable content-based similarity search. A hierarchical view of images (the Browsing Tree) that can be customized according to user preferences is provided.	Yes	No		University	Closed
SHIATSU	A novel system for automatic video tagging which is based on shot boundaries detection and hierarchical annotation processes. The tagging phase assigns semantic concepts to both shot sequences and whole videos, by exploiting visual features extracted from key frames.	Yes	Yes		University	Closed