Open CV (i.e.; Open Source Computer Vision) [4] is a free graphics library aimed mainly at the vision by computer in real time. It was originally developed by Intel and is now supported by Willow Garage and It see z. It is free for use under the open source license BSD. The library is a multiplatform with more than 2500 optimized algorithms. It includes a comprehensive set of algorithms for classical learning and the latest innovations in computer vision.

These algorithms can be used in several areas, such as: the detection and recognition of faces and the identification of objects, etc.

The image matching implements Algorithms for analysis and comparison, the purpose of which is to perform the following actions:

- Extraction of the characteristics from images

- Representation of data (model)

- Putting in correspondence with test data

- Recognition

Its main modules are [5] :

Core: the basic features: This library allows the user to manipulate the basic structures, perform operations on matrices, draw images, and save and load data in XML files …

Imgproc: Image Processing: The functions and structures of this module are related to transformations of images, filtering and detection of contours …

Features2d: descriptors: This module concerns mainly the extraction of descriptors according to two common approaches (SURF and Star detector), which are used when we are interested in the characteristics of the images.

Objdetect: detection of objects: This library allows to recognize objects in an image in the middle of the Adaboost algorithm (Viola & Jones, 2001), whose goal is the learning and the recognition of forms.

Video: treatment of video streams: These functions are used to segment and follow the moving objects in a video.

Highgui: inputs-outputs and user interface: Open CV integrates its own library top-level for opening, saving, and displaying images and video stream. The latter also contains a number of functions to achieve the graphical interfaces very simply, but this largely sufficient to test our program.

Calib3d: calibration, estimate of installation and stereovision: This module contains functions to reconstruct a scene in 3D from images acquired with several cameras simultaneously.

a) The SIFT algorithm:

The Scale Invariant Feature Transform(SIFT) method (transformation of visual characteristics invariant to the scale) [6] , is a method developed by David Lowe in 2004, allowing to transform an image in a set of vectors of characteristics that are usually invariant to geometric transformations (dilation, rotation) and in a less reliable manner to affine transformations and to lightening.

The SIFT algorithm has come to compensate for the limitations of the methods of extraction of remarkable points. In fact, it has contributed to the improvement of the techniques of extraction of information in an image by providing a robust algorithm and satisfying the properties that require the processes of artificial vision including the registration of images. It is a technique to find a geometric transformation to move from an image (called source) to another image (called target).

The implementation of the SIFT method requires two main steps:

First, it is necessary to extract the characteristics of an object and to calculate its descriptors, that is to say, to detect the characteristics that are the most likely to represent this object, to define and discriminate it in relation to the other.

Secondly, we must put in place a mapping procedure (“matching”) which is the ultimate goal of the method.

The first phase itself involves five steps which can be summarized in Figure 1.

b) The PCA-SIFT algorithm:

PCA-SIFT is a variation of the SIFT. PCA is a standard technique for the reduction of dimension, which is well suited to represent the patches of the key points. In other words, PCA-SIFT use PCA instead of a histogram to normalize the gradient patch. Its vector of characteristics is significantly smaller than that of the standard SIFT, and it can be used with the same matching algorithms. Like SIFT, PCA-SIFT also use the Euclidean distance to determine if the two vectors correspond to the same key point in different images [7] .

PCA-SIFT has fewer components that require less storage and this is achieved at a more rapid adjustment.

c) The SURF algorithm:

Speeded Up Robust Features (SURF)] [8] is a robust algorithm for extraction of interest points, which calculate the invariant local descriptors associated with these feature points.

The SURF method is inspired from the SIFT algorithm which is the forerunner in the field of extraction of the invariant points.

The general principle of this method follows the steps below:

1. Detect the points of interest.

2. Define a region of interest around each characteristic point.

3. Calculate the local descriptors from the standard areas.

4. Map the descriptors.

This experiment showed the influence of rotation on the three methods. As indicated in Figure 2, the image rotates 5 or 10 degrees. The result shows the details of the 10 first key points matched between the two images.

The experimental results are presented in Figure 3 which gives the exact rate of correspondence for each algorithm between two images.

The corresponding rate for SIFT was 1.0 for the 10 first key corresponding points. For SURF, it could only reach 0.9. This new experiment shows that SURF is not very efficient in terms of correct matching.

This experiment shows good performance according to blur invariance when the radius becomes larger. The radius of the blur changes from 0.5 to 9.0.

Figure 5 shows the details of the 10 first key corresponding points―between the original image and the

generated image. As can see in Figure 5, high blur in the second image caused the number of key points extracted from SIFT, PCA-SIFT and SURF to be 9, 7 and 10, respectively. One can also conclude that the performance of SURF is the worst under Softness invariant.

This experiment shows the effects of lighting methods. The brightness of the image became lower and lower.

Figure 6 shows the repetition of changes of illumination. SURF has the largest repetition, PCA-SIFT also shows a good performance.

The time was counted for the complete treatment which included the detection function and correspondence.

The time of evaluation is a relative result, which shows that the trends of time consumption of the three algorithms. There are factors that influenced the results such as the size, the quality of the image, the types of images (for example, landscape, or texture), and the parameters of the algorithm (for example the report of the distance).

SURF saved time in the correspondence phase. It was always the best in all situations. However, the time needed for the extraction was much higher than that of SIFT.

SIFT used the least time in the extraction of characteristics in all situations.

The system we proposed consisted in putting in place a smart solution for the automatic recognition of currency having Android as operating system.

In this next section, we will invoke the printed screens describing this system:

- Launch interface of the application

Once the user launches the application, the camera opens. It captures the first face of the bank note as shown in Figure 7.

After capturing, the system asks the user to validate the photo or to cancel it and take another one (Figure 8).

After that, the camera remains open to take the second face of the same ticket (Figure 9).

Similarly, the system asks the user to validate the photo or to cancel it and take another one (Figure 10).

Once the user completes the capture of the two images, the system switches to the comparison between the captured images and those stored in the database in a little time and with optimal quality. The system responds to the user by a sound result of the value of the bank note captured (Figure 11).

a) Principle and objectives

Remember that the objective is to enable the user to recognize the bank note manually and to rely on our system to obtain information about the value of the currency captured orally. To ensure this purpose, we used the SIFT algorithm under Open CV.

In fact, at first glance in the documentation of the Bookshop SIFT, we find that the method we are interested in is used in the detection of the points of interest:

The SIFT method takes as parameters the width and height of the image, as well as a table of integers representing the color for each pixel. However, we used an instance of the Bitmap class to represent the image, we thus proceeded to the writing of a function allowing the conversion of a bitmap type to a table of pixels:

After the conversion of the image, we created a method of comparison between the two images:

The result of this function can be true if the two images compared have more than 10 identical points of interest otherwise it returns false.

After a few executions of this function, we noticed that the result was not always accurate. It is for this reason that we added another function in order to improve the comparison for it to be more suitable.

This function was used to compare the two images pixels by pixels. If it finds that more than 20% of pixels are equivalent, it returns true. The advantage of this function is to avoid doubt in the result.

All these improvements allowed us to propose our own version of the SIFT algorithm (proposed SIFT).

Even more, we studied the execution time and its impact on the percentage of the identification of images for the two versions of the SIFT.

As well, we notice that the proposed version of the SIFT algorithm allows to reach a higher percentage of identification during a more reduced time of execution compared to the first version of SIFT.

To conclude, we can say that the proposed version of SIFT allows to give better performance in terms of recognition of images captured in a minimal execution time as shown in the curve presented in the next section.

b) Curve:

The curve below shows that the proposed SIFT saves a lot of time and that it is more efficient with regard to the identification of images, compared to the old SIFT version (Figure 12).

In fact, this curve is used as a comparative indicator for each of the SIFT methods since the proposed SIFT gives clearer results and a higher percentage in a reduced amount of time compared to the standard SIFT.