基于内容的图像检索(CBIR, Content Based Image Retrieval)是相对成熟的技术领域,在工业界也有广泛的应用场景,如搜索引擎(Google、百度)的以图搜图功能,各电商网站(淘宝、Amazon、ebay)的相似商品搜索,社交平台(Pinterest)的相似内容推荐等。本文从图像检索流程出发,结合我们团队在社交应用中的相似图片、视频检索中的实践经验,介绍构建基于内容的图像检索系统所涉及的算法技术,包括特征提取、索引构建、近邻搜索等技术,供相关领域研发人员参考。
在介绍视觉内容检索流程前,先来回顾下文本检索流程。
相似文本检索可以分成构建词库、构建索引和检索三部分,如下图所示。
#知乎打大括号好难,所以此处用'伪伪代码',凑合看吧
if q(x) == q(y) && h(b(x), b(y))<= h_t:
f(x,y) = tfidf(q(x))
else:
f(x,y) = 0
[1]. Video google: a text retrieval approach to object matching in videos, 2003.
[2]. Aggregating local descriptors in to a compact image representation, 2010
[3]. Large scale image retrieval with compressed Fisher Vectors, 2010
[4]. Triangular embedding, 2014
[5]. Nueral codes for image retrieval, 2014
[6]. Aggregating deep convolutional features for image retrieval, 2015
[7]. Spectral hashing, 2018
[8]. Deep hashing with category mast for fast video retrieval, http://cn.arxiv.org/abs/1712.08315
[9]. Deep learning of binary hash codes for fast image retrieval, 2015.
[10]. p-stable LSH: locality-sensitive hashing scheme based on p-stable distributions, 2004.
[11]. Hamming embedding and weak geometric consistency for large scale image search, 2008.
[12]. https://blog.csdn.net/icvpr/article/details/12342159
[13]. The inverted multi-index, 2012.
[14]. Efficient indexing of billion-scale datasets of deep descriptors, 2016.
[15]. Product quantization for nearest neighbor search, 2011.
[16]. Visual search at ebay, 2017.
[17]. Visual search at Pinterest, 2017.
[18]. https://github.com/facebookresearch/faiss