Zefirrat.ImageComparer 1.0.0

Zefirrat.ImageComparer

Recognition of similar or same images, with varying degrees of accuracy.
For example, for images with noise, the similarity of compressed images and the original ones, to search for the original image by its compressed copy, to search for images with small changes or distortions.

Test demonstrations

First image for compare (source):

Comparing with compressed:

Result:

Accuracy: VeryLow
    Similar: True ✔      
    Equal: True ✔        
-------------------------
Accuracy: Low
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Medium
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: High
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: VeryHigh       
    Similar: False ✖
    Equal: False ✖
-------------------------

Added transparent drawing at right bottom corner:

Result:

Accuracy: VeryLow
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Low
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Medium
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: High
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: VeryHigh
    Similar: False ✖
    Equal: False ✖
-------------------------

Added few red curves:

Result:

Accuracy: VeryLow
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Low
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Medium
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: High
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: VeryHigh
    Similar: False ✖
    Equal: False ✖
-------------------------

Added more red curves:

Result:

Accuracy: VeryLow
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Low
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Medium
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: High
    Similar: False ✖     
    Equal: False ✖       
-------------------------
Accuracy: VeryHigh       
    Similar: False ✖
    Equal: False ✖
-------------------------

Added mush more red curves:

Result:

Accuracy: VeryLow
    Similar: True ✔      
    Equal: False ✖       
-------------------------
Accuracy: Low
    Similar: False ✖     
    Equal: False ✖       
-------------------------
Accuracy: Medium
    Similar: False ✖     
    Equal: False ✖       
-------------------------
Accuracy: High
    Similar: False ✖     
    Equal: False ✖       
-------------------------
Accuracy: VeryHigh
    Similar: False ✖
    Equal: False ✖
-------------------------

Installation

PM > Install-Package Zefirrat.ImageComparer

Usage

Base

    var imageComparer = new ImageComparer();

    using var image1 = Image.Load("../../../../Test_Images/test_image_source.png");
    using var image2 = Image.Load("../../../../Test_Images/test_image_source_compressed.jpg");

    var similar = imageComparer.AreSimilar(image1, image2);
    var equal = imageComparer.AreEqual(image1, image2);

Accuracy options

Выше значение - выше допускаемый разброс.

public cllass ImageComparer
{
    // ...
        public static class AccuracyValues
        {
            public const byte VeryHigh = 1;
            public const byte High = 20;
            public const byte Medium = 50;
            public const byte Low = 90;
            public const byte VeryLow = byte.MaxValue;
        }
}

    var imageComparer = new ImageComparer(ImageComparerOptions.AccuracyValues.VeryHigh);

Dependency Injection

In Configure:

    serviceCollection.AddImageComparer

And then request as interface:

public YourControllerConstructor(IImageComparer imageComparer)
{
    _imageComparer = imageComparer;
}

Explanation:

Inspired by the work on the "Manhattan distance" algorithm, which was based on calculating the distance between similar pixels, and based on this value, the probability of similarity of images was compared by the mean square deviation (standard deviation).

I took from this approach, the work with the mean square deviation (standard deviation).

Prepare:

• To begin with, all images are resized to 16x16.
• Then we take the rgba pixel values as coordinate values
• Convert pixels to vectors
• We take the lengths of the vectors
• We get an array of vector lengths for each image being compared

For equality:

• Comparison of the vector difference of each pixel with an acceptable discrepancy

For similarity:

• Getting the quadratic deviation (standard deviation) between the vectors for each image, and the sum of all vectors
• The difference between the total deviation and the deviation for each image should not exceed the allowable discrepancy