Bit-Scalable Deep Hashing with Regularized Similarity Learning for Image Retrieval and Person Re-ide

总结：（1）对基于三元组学习的损失函数利用拉普拉斯矩阵增加了规则化项

（2）通过给hashing code 的每一Bit 给定一个权重，从而实现Bit-Scalable

（3）设计tanh-like layer

（4）设计基于图像的优化算法

Hash—> fast approximate similarity search

Abstract

Two crucial steps in image retrieval:

(1)Extracting informative image features

(2)Learning effective approximate hashing function

Conventional methods:

(1)Learn hash functions from a predefined hand-craft feature space

(2)Preset the bit lengths of output hashing codes

This paper:

we propose a supervised learning framework to generate compact and bit-scalable hashing

codes directly from raw images.—>Pose hashing learning as a problem of regularized similarity learning.

 Organize the training images to triplet samples

 Maximize the margin

 Regularization term –>adjacency consistency

 Unequally weights for bits of hashing codes.

Experiments:

Outperforms state-of-arts on public benchmarks of image retrieval.

Shows promising results in application of Person re-identification.

Bit-scalable hasing codes preserve the discriminative powers with shorter code lengths.

Introduction

(1) Traditional methods:

The feature representation may not be tailored to the objective of hashing learning.

Moreover, the hand-crafted feature engineering often requires much domain knowledge and heavy tuning.

(2)Most approaches–>preset lengths

However, one often requires hashing codes of different lengths under different scenarios.

E.g. ,mobile devices

To cope with such requirements:

one solution is store several versions of hashing codes in different lengths.–>extra computation and storage.

Bit-scalable hashing methods: generate hashing codes bit by bit in a significance decent way.–> need carefully design the embedded feature space

the performances may dramatically fall when shortening the hashing codes.

This paper:

Supervised Bit-Scalable Deep Hashing framework:

(1)CNN

(2)Each bit of hashing codes is weighted according to their significance

Main components:

(1)Present a novel formulation of relative similarity comparison based on the triplet-based model.—–>

we extend the triplet-based relative comparison by incorporating a regularization term, partially motivated by the recently proposed Laplacian Sparse Coding.

(2)Adopt the deep CNN architecture to extract the discriminative features from the input images.

One FC layer and one tanh-like layer—->output the binary hashing codes

An element-wise layer–>to weight the bit of hashing code.

(3)Implements the learning algorithm in a batch-process fashion.

Organize the triplet samples from a randomly selected subset ( 150~200) of the training images.

Use stochastic gradient decent (SGD) for parameter learning.

(Calculate the partial derivative on images instead of on triplets samples.—->reduce the computational cost which is linear to the selected subset of images.)

Contributions

(1) It unifies feature learning and hash function learning via deep neural networks, and the proposed bit-scalable hashing learning can effectively improves the flexibility of image retrieval.

(2) It presents a novel formulation (i.e., the regularized triplet-based comparison) for Hashing learning.

(3)Extensive experiments on standard benchmarks demonstrate:

The learned hashing codes well preserve instance-level similarity

Outperforms state-of-the-art hashing learning approaches.

Successfully apply the method to person re-identification.

Related work

Hashing methods: data-independent and data-dependent.

Data-independent: without any training , make the codes scattered

Data-dependent: How to learn compact hashing codes from the training data .

(1)projecting the high dimensional features onto the lower dimensional space.

(2)quantizing the real-valued representation into binary codes.

Deep learning methods:

[2] proposed a learning-to-rank framework based on multi-scale neural networks, and showed promising performance on capturing fine-grained image similarity.

Pre-training on ImageNet.

[35] utilized CNN for supervised hashing.

Produced the hashing codes by decomposing the pairwise similarity matrix, then learned the mapping fucntions from images to the codes.—>may fail to deal with large-scale data due to the matrix decomposition operation.

Experiments

Datasets : MNIST , CIFAR-10 , CIFAR-20 and NUS-WIDE

Split images into a training set and a query set

Leave -one-out

CHUK03

Several variants of the framework:

(1) without element-wise layer

DRSCH (Deep Regularized Similarity Comparison Hashing)

DSCH

To verify the effectiveness of the regularization term.

(2)remove tanh-like layer —>evaluate the effectiveness of tanh-like

Euclidean

(3)Bit-scale

BS-DRSCH

BS-DSCH

Four types of evaluation tools:

mAP: http://www.zhihu.com/question/41540197

precision@500

precision@k

HAM2