Images Similarity

Workflow Images Similarity in HGrid247-3.5

The purpose of this workflow is:

Comparing an image with a group of other images
Calculating the similarity value between images
Sort results by highest similarity level

Usage example:

similar image search
image duplication detection
image recommendation system

Step 1 – Creating a Workflow

Right click on the workshop package
Select New Workflow
Enter the class name:
```
    WFImageSimilarity
    
```

After clicking Finish , a workflow canvas will appear to design the pipeline.

Step 2 – Defining Source Data

Add Custom Source (

) from palette.

This source is used to read image files .
Arrangement:

    Input Format:
    com.solusi247.mapred.image.ImageInputFormat

This means the system reads data in the form of image files as input.

Step 3 – Defining Input and Output Types

Add Transformator ()

Connect:

    Custom_input_format_1 → Transformator_1

Add output field:

    image1 : Object
    image2 : Object

Use the function:
```
    ImageFromArgument
```
This function is used to retrieve the reference image from the program parameters.

Data relations:

    line → ImageFromArgument → image1
    line → image2

It means:

    image1 = image from argument
    image2 = image from dataset

Step 4 – Calculating Image Similarity

Add Transformormator (

).
Connect:

    Transformator_1 → Transformator_2

Add output field:

    image_name1 : String
    image_name2 : String
    similarity  : Double

Add the following function:

1 GetImageName

Used to retrieve the image file name .

Relation:

    image1 → GetImageName → image_name1
    image2 → GetImageName → image_name2

2 ImageSimilarity

This function calculates the similarity level of images.

Input:

    image1 
    image2

Output:

    similarity

Relation:

    image1 ─┐
            ├── ImageSimilarity ──> similarity
    image2 ─┘

Preferred method:

    SIFT (Scale Invariant Feature Transform)

This method detects keypoints in the image and then compares them.

Step 5 – Grouping dan Sorting

Add GroupBy Assembly (

).
Connect:

    Transformator_2 → GroupBy_1

Arrangement:

    Group Field:
    image_name1

Sorting:

    Sort Field:
    similarity (descending)

This means that the results will be sorted from the highest similarity value .

Add Transformaator (

).
Connect:

     GroupBy_1 → Transformator_3

Use options:

    Identity All

This function simply passes the data to the next level.

Step 6 – Output Workflow

Add HFS Sink (

) to save results to HDFS. Connect:

    Transformator_3 → Hfs_sink_1

The output will be stored in a flat file format on HDFS.

Step 7 – Compile Workflow

Before compiling, add additional libraries:

    image-video

Menu:

    File → Additional Library

Then click the button:
Generate (

)

HGrid247 will produce:

    workshop_mr.jar

Step 8 – Running the Workflow

Workflow is executed on Hadoop cluster using command:

    hadoop jar workshop_mr.jar workshop.WFImageSimilarity book output_sim book/cover4.jpg

Parameter:

    Parameter           Description
    ----------------------------------------
    book                image dataset folder
    output_sim          folder hasil output
    book/cover4.jpg	reference image

Example Output Results

    book/cover4.jpg     book/cover4.jpg     1.0
    book/cover4.jpg     book/cover3.jpg     0.49198125324304637
    book/cover4.jpg     book/cover5.jpg     0.2866741744110343
    book/cover4.jpg     book/cover1.jpg     0.18940610319724102
    book/cover4.jpg     book/cover2.jpg     0.1566145899555032

Explanation:

    image1	image2      similarity
    -------------------------------
    cover4	cover4      1.0 (identical)
    cover4	cover3      0.49
    cover4	cover5      0.28

The closer to 1.0 , the more similar the images are.

Note : The file of workflow WFImageSimilarity can be found in the following link : WFImageSimilarity.csd