Hashing for Image Retrieval

Web implement of hashing methods for image retrieval in Python and DPLM method was applied in this version:

    Shen F, Zhou X, Yang Y, et al. A fast optimization method for general binary code learning  
    IEEE Transactions on Image Processing, 2016, 25(12): 5610-5621.

In further versions, I will add more methods, especially deep end-to-end framework methods. If you have any question, you can ask me through Issues.

Getting started

1. Ensure you have installed Caffe on you platform and you can you can refer to this document if not.
2. There are two query methods: memory and database. If you choose memory method, you can ingnore the database configuration and just load data from .pkl file. Configure enviroment variant such as caffe root path, database, query source via configuration.conf.
3. Import the database named irs_sun397.sql in folder DB to your own database or directly use hash binary code and corresponding path saved in .pkl file in folder hashing_model, The details are in Database and PKL file.
   Remark: database is necessary to large-scale image dataset but .pkl file is convenient to experiment.
4. Download the hashing model from here(code is zqye) and place in folder caffe_model.(ONLY DPLM MODEL)
5. Download the SUN397 dataset from here and place in static/media/dataset folder.
6. After correcttly placing all files , run app.py by using command below

    python app.py

then browse http://127.0.0.1:5000.

Dataset and deep feature

The SUN397 dataset VGG16's 'fc7' feature extracted by using deep learning framework Caffe. There are a page where offer details about SUN397 dataset and download link. The VGG16 pre-trained model can be download from here. The raw features can be download from here.(code is b98q)

Hashing binary code

After some pre-processing(center and normlize), 4096-d feature was transformed into 128-bit hashing binary code by using hashing method.

PKL file

There are two pkl files named DPLM128Path.pkl, DPLM128Code.pkl and stored image path and corresponding hashing binary codes. There are all stored in list, but path's element is str and hashing binary code is tuple. You can load the data as follows:

>>> import pickle
>>> code = pickle.load(open('/your/own/path/hashing_model/DPLM128Code.pkl','rb'))
>>> path = pickle.load(open('/your/own/path/hashing_model/DPLM128Path.pkl','rb'))
>>> code[0]
(1249808051, 2609047194, 471897213, 2675925156)
>>> path[0]
'/SUN397/p/picnic_area/sun_bpcvysbqhstfifcb.jpg'

Database

The hashing binary code are stored in the database with mysql. The database is named irs_sun397.sql and the table 'img' construction is bellow:

Field	Type	Null	Key	Default	Extra
id	int(11)	NO	PRI	NULL	auto_increment
path	varchar(150)	YES		NULL
code_1	bigint(20)	YES		NULL
code_2	bigint(20)	YES		NULL
code_3	bigint(20)	YES		NULL
code_4	bigint(20)	YES		NULL

A detailed data is as follows

id	path	code_1	code_2	code_3	code_4
10000	/SUN397/m/manufactured_home/sun_bwnvkpvdgnxhkjst.jpg	4076301309	917904382	1211050903	2286664558

The database can be found in folder DB and imported to your own database by using command:

    source irs_sun397.sql

A hamming distance function has to be created in order to retrieval simililar images with hashing binary code.

CREATE FUNCTION HAMMINGDISTANCE(
  A0 BIGINT, A1 BIGINT, A2 BIGINT, A3 BIGINT, 
  B0 BIGINT, B1 BIGINT, B2 BIGINT, B3 BIGINT
)
RETURNS INT DETERMINISTIC
RETURN 
  BIT_COUNT(A0 ^ B0) +
  BIT_COUNT(A1 ^ B1) +
  BIT_COUNT(A2 ^ B2) +
  BIT_COUNT(A3 ^ B3);