Skip to main content

Deep Learning based CBIR 101 (Part 2): the basics

In part 1 of this series a definition of CBIR has been given. Let's now understand what's the typical flow for it. The diagram in figure 1 shows that there are two parts, one that happens offline and another which is online:


Figure 1

Starting from an image storage, a preliminary trained Deep Neural Network is used to extract the features from images. Extracted features are then stored in a feature database. This happens offline any time new images are added to the storage. What happens online is the search process itself. Any time a user uploads a query image, the same Deep Neural Network used for feature extraction is used to extract the input image features. Then the distance from the query image features and the features in the database are computed. The closer the distance, the higher the relevance is. The closest features are then sorted and the corresponding images are returned as results.

Basic implementation: data set preparation

To make things more clear, let's start a first simple implementation using a supervised Deep Learning approach for feature extraction.
The data set used for this example is the Food 101 data set. It contains images of food recipes from international cuisine classified in 101 categories (classes). Its size is about 5 GB and contains 1000 images for each class. The structure of the ZIP archive includes:
  • an image directory, which contains all the images grouped into sub-directories, one for each food class;
  • a meta directory, which contains four text files:
    • classes.txt: the list of class names;
    • labels.txt: a list of verification labels;
    • train.txt: the list of the training images;
    • test.txt: the list of the test images.
Figure 2 shows a sample of the images in this data set.


Figure 2

Unfortunately there are no separate directories for training and test images: you have to split them by yourself (a sample Python script would be provided in GitHub later on). Once the train/test split has been done, the training of the Deep Neural Network (a CNN) could start.
The next post will walk through the details of the feature extraction.

Comments

Popular posts from this blog

Exporting InfluxDB data to a CVS file

Sometimes you would need to export a sample of the data from an InfluxDB table to a CSV file (for example to allow a data scientist to do some offline analysis using a tool like Jupyter, Zeppelin or Spark Notebook). It is possible to perform this operation through the influx command line client. This is the general syntax: sudo /usr/bin/influx -database '<database_name>' -host '<hostname>' -username '<username>'  -password '<password>' -execute 'select_statement' -format '<format>' > <file_path>/<file_name>.csv where the format could be csv , json or column . Example: sudo /usr/bin/influx -database 'telegraf' -host 'localhost' -username 'admin'  -password '123456789' -execute 'select * from mem' -format 'csv' > /home/googlielmo/influxdb-export/mem-export.csv

jOOQ: code generation in Eclipse

jOOQ allows code generation from a database schema through ANT tasks, Maven and shell command tools. But if you're working with Eclipse it's easier to create a new Run Configuration to perform this operation. First of all you have to write the usual XML configuration file for the code generation starting from the database: <?xml version="1.0" encoding="UTF-8" standalone="yes"?> <configuration xmlns="http://www.jooq.org/xsd/jooq-codegen-2.0.4.xsd">   <jdbc>     <driver>oracle.jdbc.driver.OracleDriver</driver>     <url>jdbc:oracle:thin:@dbhost:1700:DBSID</url>     <user>DB_FTRS</user>     <password>password</password>   </jdbc>   <generator>     <name>org.jooq.util.DefaultGenerator</name>     <database>       <name>org.jooq.util.oracle.OracleDatabase</name>     ...

Turning Python Scripts into Working Web Apps Quickly with Streamlit

 I just realized that I am using Streamlit since almost one year now, posted about in Twitter or LinkedIn several times, but never wrote a blog post about it before. Communication in Data Science and Machine Learning is the key. Being able to showcase work in progress and share results with the business makes the difference. Verbal and non-verbal communication skills are important. Having some tool that could support you in this kind of conversation with a mixed audience that couldn't have a technical background or would like to hear in terms of results and business value would be of great help. I found that Streamlit fits well this scenario. Streamlit is an Open Source (Apache License 2.0) Python framework that turns data or ML scripts into shareable web apps in minutes (no kidding). Python only: no front‑end experience required. To start with Streamlit, just install it through pip (it is available in Anaconda too): pip install streamlit and you are ready to execute the working de...