This paper devoted the research with an aim of enhancement of information technology for person identification based on image quality features. Nowadays, face recognition and identification technologies are rapidly spread and widely used in various fields, from access control of sensitive data to building war crime cases. So, it is critically important to investigate such technologies, so that mistakes made by the software or those using it could not have an impact on human lives by dire consequences. Due to the previous research, it was found out that the accuracy rate of the algorithm applied for face recognition and identification can significantly vary on different datasets. It is common that databases of different organizations contain images of different quality. Comparison of the high-quality image (such as passport image in law enforcement database) to the low-quality image (such as surveillance camera image) may cause the failure of the identification process. This fact means that it is necessary to research the image quality features in order to unify images and increase the accuracy rate in identification process. Experimental research, described in this work, was performed in three stages with an aim to determine the dependency of the identification accuracy rate on properties of the database images and testing images sample, such as: image format, image resolution, face region that covers the image.
Face recognition is a widely used technology that has been developed in usage from the function of smartphone unlocking or tagging people in social networks to the law enforcement technology of person identification to prevent crime and finding missing children. Many countries have made tremendous progress in applying this technology in various fields. Face recognition information is currently being used by law enforcement to make policing decisions, and might in the future be employed by public or private bodies to make decisions as to e.g. your suitability for a job, a university course or loans, overseas travel, or entry into a festival.
According to the Comparitech report [
(O. Bychkov); 0000-0002-8526-8211 (O. Ivanchenko); 0000-0001-6347-5191 (K. Merkulova), ️©
2022 Copyright for this paper by its authors. appliance, with the possible further use of the data obtained in this way by law enforcement, facе recognition algorithms must be highly accurate in order to allow further identification of the person and avoid false or false positive identification results.
Nowadays, face recognition and identification technologies are widely used by Ukrainian
government and independent organizations during the Russian-Ukrainian War with the purpose of
identifying Russian soldiers and to verify that travelers in Ukraine are who they claim [
In previous work [
For the problem that the face image from surveillance video stream with law quality cannot be
accurately identified in comparison with high quality database images, this paper proposes a study of
information technology for person identification in video stream proposed in [
First stage of the algorithm consists in capturing the snapshot of the person in video stream and transforming it into image. Then face detection and localization processes are performed on this image. An example of the algorithm performance on this stage is presented in Figure 1.
Face image obtained after the first stage preprocesses with anisotropic diffusion method [
( ( , , ) 2), where div is the divergence operator, c(x, y, t) is a family of derived images obtained by convolving the original image I0(x, y) with a Gaussian kernel G(x, y, t0), Δ is the Laplacian operator.
After applying of anisotropic diffusion on face image, it processes by Gabor wavelet transform with
various parameters changes in the wavelet function allowing to obtain 16 transformed variations of the
input face image. An example of image processing with Gabor wavelet transform is presented in Figure
(1)
2. The result of this stage is a global image obtained by summing of those images into one global image.
Gabor wavelet transform in spatial domain is based on the following formula [
( , ) = ( , ) ( , ), ( , ) =
( (2 ( 0 + 0 ) + )), where (u0, y0) is the spatial frequency and P is the sinewave phase.
Gaussian 2-dimensional function wr, which is the envelope function, can be represented with the following: operation: where (x0, y0) is the function peak, a and b are the scaling parameters of Gaussian, r defines the rotation
After all previously mentioned notations, complex Gabor function in the spatial domain can be
written as follows [
( (2 ( 0 + 0 ) + ),
vectors by histogram of oriented gradients (HOG) and local binary patterns in 1-dimensional space
(1DLBP). Using the method of histogram of oriented gradients allows to obtain gradient values for each
cell of the matrix, that represents the image, normalize and group those values into the 512-valued
feature vector [
. ( , ) = ∙
(− ( 2( – 0)2 + 2( – 0)2)), ( – 0) = ( – 0) + ( – 0)
, ( – 0) = −( – 0) + ( – 0), (2) (3) (4) (5) (6) (7) (8) (9) where R(x) is function of pixel value defining; pn and p0 are the values of current pixel and central pixel, respectively. Result of 1DLBP method is the 512-valued feature vector.
After HOG and 1DLBP feature vectors have been received, it is necessary to normalize them with the purpose of the following concatenation. Normalization performs by min-max method that transforms the feature vector values in the range from 0 to 1. Normalized vectors get concatenated to form the global feature vector of 1024 values. Global feature vector can be used in further classification process. Three face databases were used to perform the research of the proposed technology: The Database of Faces, Facial Recognition Technology (FERET) database and Surveillance Cameras Face Database (SCface). Analysis of the experimentational results indicated that the proposed information technology allows to obtain high identification accuracy rate on low-resolution face images. The highest identification accuracy rate of 92.5% was obtained on the images from SCface database, and identification accuracy rates from 70 to 72.5% were obtained on the Database of Faces and FERET database. As a consequence of such variation of the obtained results, the following problem arose: if the algorithm provides the high identification accuracy rate on the set of certain images, what affecting factors reduce its efficiency on the sets of images from other databases. With this problem in mind, it was decided to conduct the experimental research to determine the dependency of the identification accuracy rate on the following image properties: image format, image resolution, face region that covers the image. Additionally, it was resolved to examine the algorithm results on the images that contain the occlusion and on the images that were preprocessed supplementary methods.
Experimental research was performed in three stages with an aim to determine the dependency of the identification accuracy rate on properties of the database images and testing images sample, such as: image format, image resolution, face region that covers the image.
For the experiments three face image databases were used: prepared by AT&T Laboratories
Cambridge the Database of Faces (DoF) [
First set of experiments was performed with the conversion of original images from the previously mentioned databases to the BMP, PNG (compressed) and JPG (uncompressed) image formats. Results of these experiments are presented in the Table 1. As can be concluded from the comparative diagram in Figure 4, the format conversion did not affect the experimentational results for the SCface database - identification accuracy rate inalterably equals to 92.5%. Considering the experimentational results for images from the Database of Faces, can be seen that the conversion of the original image format, which is PGM, to the JPG improved the obtained result from 70 to 75%. On the contrary, result for the FERET database decreased on 2.5% with the image conversion to JPG format, while being stably equal to 72.5% with the image conversion to the BMP and PNG formats. 3.2.
With the aim to determine the dependency of the identification rate on the image resolution, it was decided to perform the experiments with the various resolution values. Total images /
individuals Number of images Identification rate
Total images /
individuals Number of images Identification rate
Total images /
individuals Number of images Identification rate
As far as all databases used in this work contain images with different initial resolution, and to convert these images to the single resolution it is required either increase or decrease their height and width values, it is impossible to select only one conversion method. Therefore, it was decided to perform the experiment that compare these two methods and converts the images to the resolution, that is acceptable for the images from all databases, regardless of whether image resolution needed to be increased or decreased. The image format was selected on the basis of previously obtained results. So, the setup for this experiment is the following: images resolution is width x 100, images format – JPG. The experimental results are divergent for both methods. On different databases one of the methods can exceed the identification accuracy rate of the other method as well as inferior it. Due to this result, for the performance of the experiment on the image resolution it was decided to use both of these methods considering if increasing or decreasing of the image resolution is needed.
As following from the analysis of previous experimental research, the highest identification accuracy rate was obtained as the result of algorithm appliance on the images from SCface database, that contains image with the resolution of 75x100, 96x128 and 108x144 pixels. The same resolution values were chosen to perform the experiments with the resolution conversion on the other databases, as well as for the SCface database with all images conversion to the unitary resolution.
For image resolution conversion the thumbnail function was used on early stages of the research. During the experiments it was found that the images from the Databases of Faces cannot be converted to the size of 75x100 with thumbnail function only, instead images were converted to the resolution of 75x91 pixels. Since the obtained results on this resolution seemed promising, it was decided to use this resolution for the experiments as well.
In preparation of the experiments, it was important to convert images to the same resolution values. But, as far as all used databases contain initial images with different resolution, the aspect ratio of the images could have been changed. Consequently, the image features could have been changed as well, including the features of the face depicted on the images, which is essential in face recognition tasks. So, the image resolution was converted by the single height keeping the aspect ratio unchanged with automatically defined value of the image width. As the result, experiments were performed with the following resolution values: 75x91, 75x100, 96x128 and 108x144 – for the Database of Faces; 61x91, 67x100, 85x128 and 96x144 – for the FERET database; 68x91, 75x100, 96x128 and 108x144 – for the SCface database. Also, the format of the images was changed accordingly to the first set of the experiments. The results of experiments are presented in the Table 2.
Figures 5-7 depict results obtained after experiments on the image resolution, the following conclusions can be drawn. The appliance of the resolution conversion methods on the images from the Database of Faces increased the identification accuracy rate from 70% on initial images to 77.5% on images with resolution 75x91 pixels converted to BMP and PNG images. In relation to the results of single format conversion obtained results increased on 2.5%. Results on converted resolution of 82x100, 105x128 and 118x144 in compare to previously obtained results decreased to 55-72.5%.
Identification accuracy rates obtained on the FERET database significantly decreased after image resolution conversion to 61x91 and 67x100 pixels – from 72.5% to 52.5-62.5%. On the other hand, the results are steadily equal to the previously obtained result of 72.5% on BMP and PNG images with the
BMP
PNG resolution of 85x128 and 96x144 pixels. And the result for 96x144 pixeled JPG images is the highest among all obtained on the FERET database, equals to 75%.
Results for the SCface database images decreased after being converted to the resolution of 68x91, 75x100 and 96x128 pixels. Identification accuracy rate varies from 62.5% to 87.5%. On contrary, image resolution conversion to 108x144 pixels allowed to obtain the result of the initial 92.5% on BMP and PNG images and the highest result of 95% on JPG images.
On the next stage of the experimental research, it was decided to define if the identification accuracy rate depends on the size of the face region that images contain. For this purpose, the experiment that provided the highest identification accuracy rate result during the previous research, in particular, the experiment on the JPG images that SCface database contains, was analyzed. As a result of analysis, it was determined that the images with the face region of 47x47 and 78x78 pixels were correctly identified in the majority of experiments. Size of the face region on the images from the SCface database varies due to the fact that initial etalon and testing images have different sizes as well. Hence, it was decided to use mentioned sizes for the purpose of the experimental research in this work. Also, to determine the variety of the identification rate beyond the selected sizes, it was decided to conduct the experiments on the images that contain face region below and beyond the threshold values (32x32 and 128x128), and between the selected values and threshold values (64x64). Thereby, the experiments on the face image region were performed as: face detection on the image, conversion of the obtained face region to the selected size, input of the image containing only face region of the size of 32х32, 47х47, 64х64, 78х78 or 128х128 pixels to the algorithm. Table 3 contains the results of the described experiments.
Consequently, as can be seen on the Figures 8-10, the identification accuracy rate does not depend on whether the etalon images and testing images samples contain any other detail than the face region itself. The algorithm contains methods for face detection and image resizing, that solve the task independently, regardless of the quality of the input data. Considering the variety of results on the same 10 25% 11 28 70% 26 65% 34 85% 32 80% 30 75% 29 12 30% 14 35% 15% 6 8 20% 35% 13 32.5%
FERET resolutions of the face region images, it still necessary to research methods of face region image conversion in order to unify it.
Also, during the research it was found out that identification accuracy rate does not depend on whether the etalon images and testing images samples contain any other detail than the face region.
For further research of investigated algorithm that is underlies in face recognition and identification technology it is needed to continue the experiments with a change of parameters in methods of unifying image formats and resolutions, as well as reconsider selection of preprocessing methods and image classification methods.
This paper is devoted to the research of information technology for person identification in order to enhance its efficiency by studying features of the comparing images. The image features that were studied are the format, resolution and face region covering the image.
The format conversion to the JPG improved the obtained result for the Database of faces from 70 to 75% and did not affect the results on the images from other databases.
The appliance of the resolution conversion methods on the images from the Database of Faces increased the identification accuracy rate from 70% on initial images to 77.5% on images with resolution 75x91 pixels converted to BMP and PNG images. The result for 96x144 pixeled JPG images is the highest among all obtained on the FERET database and equals to 75%. Image resolution conversion to 108x144 pixels allowed to obtain the result of the initial 92.5% on BMP and PNG images and the highest result of 95% on JPG images.