给你几篇参考资料看一下：
CNN 网络直接学习二进制特征
Learning to Hash Paper, Code and Dataset： http://cs.nju.edu.cn/lwj/L2H.html
Learning to Hash for Big Data: A Tutorial

台湾中央研究院资讯科学研究所 Kevin (Ke-Yun) Lin 林可昀有多个开源项目：
Kevin Lin, Jiwen Lu, Chu-Song Chen, Jie Zhou. Learning Compact Binary Descriptors with Unsupervised Deep Neural Networks. CVPR, 2016.
代码是基于 Caffe 的： https://github.com/kevinlin311tw/cvpr16-deepbit

Huei-Fang Yang, Kevin Lin, Chu-Song Chen. Supervised Semantics-Preserving Hash via Deep Convolutional Neural Networks, TPAMI under revision.
SSDH 代码是基于 Caffe 的： https://github.com/kevinlin311tw/Caffe-DeepBinaryCode

K. Lin, H.-F. Yang, J.-H. Hsiao, C.-S. Chen. Deep Learning of Binary Hash Codes for Fast Image Retrieval. CVPR Workshop (CVPRW) on Deep Learning in Computer Vision, DeepVision 2015.
代码是基于 Caffe 的： https://github.com/kevinlin311tw/caffe-cvprw15
Slide： http://www.csie.ntu.edu.tw/~r01944012/deepworkshop-slide.pdf


卷积层特征聚合为全局特征
A Babenko, V Lempitsky. Aggregating Local Deep Features for Image Retrieval. ICCV 2015.
特征简称为 sum-pooled convolutional features (SPoC descriptors)

G. Tolias, R. Sicre, and H. Jegou. Particular object retrieval with integral max-pooling of CNN activations. ICLR 2016.
特征简称为 R-MAC：Regional Maximum Activation of Convolutions
Matlab/MEX 代码： http://cmp.felk.cvut.cz/~toliageo/soft.html

Yannis Kalantidis, Clayton Mellina, Simon Osindero. Cross-dimensional Weighting for Aggregated Deep Convolutional Features. arxiv 2015.
特征简称为 CroW
代码为 Python： https://github.com/yahoo/crow
在 Paris、Oxford 和 Holidays 三个小规模图片搜索数据集上，CroW 的 mAP （ mean Average Precision ）均高于 SPoC 和 R-MAC 等算法。