Application and Feasibility Assessment of Artificial Intelligence in Forensic Identification and Document Examination

In the context of the “Internet +” era, artificial intelligence continues to give rise to new economic forms, providing an environment for mass entrepreneurship and innovation, and also making the Internet and the law have a variety of links. As an important part of public legal services, forensic identification and document examination work is also driven by the ubiquitous computing, data, knowledge and continuous innovation and development [1-4].

Document examination is a common way and means in forensic procedures, and plays an important role in various judicial proceedings in China. The results of document examination is usually from the perceptual understanding of the rise of rational understanding of the concluding observations, document examination conclusions are essentially an analysis of the document and judgment of the opinion [5-8]. Document examination conclusion can also be referred to as a document examination results, belonging to the results of judicial appraisal of a kind of scientific evidence. In recent years, the number of forensic identification business and the value of fees have maintained a sustained growth trend, forensic identification work in the litigation activities play an increasingly important role in the service and protection [9-12]. However, while bringing help to the process of social rule of law, the imperfect system and unstandardized practice have also seriously affected judicial justice and brought unstable factors to the whole society [13-15]. Therefore, in such a situation, according to the development trend of forensic identification work, the advantageous role of big data, artificial intelligence and other technologies in the Internet era should be given full play to provide intelligent support for forensic identification work, enhance the intrinsic motivation of its development, ensure that the forensic identification opinions are scientific and fair, and adapt to the reform of the litigation system that is centered on the trial [16-19].

In this paper, first of all, the forensic document examination of camouflage handwriting inspection and identification is briefly described, and the specific methods of camouflage handwriting inspection and identification in forensic document examination are explored. It is proposed to utilize the Hough line transform for image feature detection to realize the physical evidence image preprocessing. The text lines are quickly extracted by horizontal projection histogram, and at the same time, combined with the improved SWT stroke width algorithm, the Chinese text image is detected. Using the minimum error method to extract the difference between the target object and the background in gray scale, the image is divided into regions with their own characteristics. Aiming at the difficult problems of altered document inspection, artificial intelligence is proposed to be used for document inspection. Wavelet transform and Fourier transform are selected as the control group to verify the advantage of Hough line transform in image compression. The proposed Chinese text detection algorithm is used for document homogeneity analysis, and experiments under different padding conditions are designed. The algorithm detects the difference in stroke width and the difference in indentation features are compared to examine the reliability of the algorithm under different conditions. Based on two-factor analysis, the effectiveness of the minimum error method for extracting features is examined. Raman spectral analysis is introduced to explore the influence of image segmentation processing on the final recognition results.

2

Common types of disguised handwriting in forensic document examination

In the specific practice of carrying out forensic document examination, a comprehensive understanding of the common types of disguised handwriting, can to a large extent improve the quality and efficiency of the work of disguised handwriting examination and identification. At present, forensic document examination of common camouflage handwriting is mainly manifested in four aspects, respectively, camouflage low level handwriting, change the shape of the word, destroy the structure of the text and left-handed camouflage. 1)

Camouflage low level handwriting

In the process of writing, the handwriting forger will deliberately reduce the proficiency of handwriting and writing speed, reduce the continuity between strokes and strokes, so that the font presents a dull, hard visual effect; even some handwriting forger will deliberately write some misspelled words, making the whole text looks like an elementary school student. At present, the realization of this low level of handwriting disguise is very easy, the comprehensive quality of handwriting disguise is not too high level requirements. However, it is worth noting that low-level handwriting disguise often leaves a lot of traces.

2)

Changing the shape of characters

After thousands of years of evolution, Chinese characters have developed an intrinsic characteristic of being horizontal and vertical. For most people, they will write Chinese characters in a square shape, while some will write them in a rectangle, flat circle or round shape. Relevant research data show that there exists a very close connection between character shape and the psychological state, external image, aesthetic interest and daily training of the writers. In the process of handwriting disguise, the camouflage will change the original shape of the text, so that the handwriting in the document and the daily handwriting there are obvious differences. This type of handwriting camouflage is highly skillful, and the slightest mistake can change the smoothness and overall aesthetics of the text. Taking the change from square to rectangular glyphs as an example, if the handwriting forger has not been trained for a long time, then there will be a situation where square and rectangular Chinese characters are intermingled with each other.

3)

Destroying the structure of characters

In the process of disguising handwriting, destroying the text structure is mainly manifested in changing the internal structural relationship of the text, changing the order of the strokes of the text, changing the proportion of the strokes of the text, deliberately shortening specific text strokes and deliberately lengthening specific text strokes. In forensic document examination, the destruction of the text structure of the handwriting camouflage way is also more common.

4)

Left-handed Disguise

If a person is accustomed to use the right hand in daily life, then the left-handed writing of the text will have a significant gap with the daily handwriting, and the whole text will look very awkward, or the proportion of the text is very incongruous. It is important to note here that left-handed camouflage does not specifically refer to camouflaging handwriting with the left hand, but rather to camouflaging handwriting with an unaccustomed hand. For example, if a person’s habitual hand is his left hand, then his choice to disguise his handwriting with his right hand is also known as left-handed camouflage.

3

Artificial intelligence-based methods for the examination of forensic documents

3.1

Physical evidence image pre-processing

3.1.1

Image processing software

In the process of physical evidence image processing, should be in accordance with a certain proportion of the image file for zoom processing, at this time will be applied to a computer software called Imagesize, the role of this software tool is mainly with the role of the computer image processing software to zoom in and zoom out of the photo or image. Specific implementation process is as follows: First of all, open the computer image processing software and Imagesize, the need to deal with the image on the basis of this open, and toolbox, select the “metric tool” to measure the operation. Secondly, after measuring a specific proportion of the image, to determine its specific length, and its length of the specific value of the input to Imagesize, and then read the specific value of the image processing software, and this value is input to the Imagesize, so as to determine the new length of the image file, to reset the size of the image.

3.1.2

Image Feature Detection Based on Hough Line Transformation

Hough transform as a common method of image feature extraction, is widely used in image geometric feature extraction, common geometric shapes, such as circles, straight lines, etc. can be used to extract features using the Hough transform method. Meanwhile, in the detection of red header files, the large long red line of the red header file can also be used as a strong feature point, which is based on the Hough line detection method, through the detection of the tilt angle between the straight line and the image level baseline, tilt correction, compared to the method based on the two-dimensional discrete Fourier transform, the method is based on the support of the OpenCV image processing library, the binary image as the processing object, the implementation is simpler. Processing object, the implementation is easier, but as a tilt correction method, does not have a good robustness, most of the printed text files, in the case of not correctly extracted straight line features, tilt correction of tilted printed text is less than ideal.

Principle of Hough line detection: in two-dimensional space, two points can determine a straight line, usually can be expressed by the formula y = kx + b, but the straight line approaching the vertical direction, the slope k will tend to ∞, this time with the above formula will be difficult to express, for the slope k will tend to ∞ occasions, generally use the Cartesian polar coordinates to express, the following formula (1) shows: (1) $ρ = x \cos θ + y \sin θ$

The Hough line polar coordinate representation is shown in Fig. 1, where ρ is the distance from the origin to line m, which is the length of the red line in Fig. 1, and θ is the angle between the red line and the x-axis.

3.2

Chinese text image detection in the application of document inspection

3.2.1

Improved stroke width algorithm

In the judicial appraisal of document inspection, the use of artificial intelligence for Chinese text image inspection can verify the consistency of the content of the document, to identify whether the document has been altered. Most of the traditional text detection algorithms for the detection of English text, the detection of Chinese is often less than ideal, this paper in the SWT algorithm on the basis of the use of the algorithm will be the width of the text strokes as text as an intrinsic feature to extract the text text, while combining the Chinese inherent characteristics of the structure of the strokes, the design of the following four inspirational rules: 1)

Increase the ratio of the number of foreground pixels in the text line to the connected area to further filter out the strong interference noise, so as to make the detection of Chinese text more robust, and it is determined that the effect is optimal when the number of foreground pixels accounts for the ratio of the connected area of the overall text line SwtRulel > 0.6. The formula is shown in (2): (2) $S w t R u l e 1 = \frac{S_{t e x t}}{S_{r e c t}} > 0.6$

Where: S_text is the area of the number of foreground pixels, and S_rect is the area of the overall text line connectivity field. 2)

Increase the text line connectivity domain aspect ratio with the formula shown in (3): (3) $S w t R u l e 2 = ((s (h) \geq 10 & & s (w) \geq 30) & & \frac{s (w)}{s (h)} \geq 3)$

Where: s(h), s(w) are the height and width of the connected domain of the text line, respectively. 3)

For the original algorithm to find the direction of the width of the text stroke threshold π/6, appropriate amendments, this paper in the detection of the research process found that the original algorithm threshold π/6, in the English text such as text morphology transformation of the simple language, the effect is still possible, in the Chinese language such as morphology of the text of the complexity of the threshold is set too strict, resulting in the retrieval of the direction of the width of the stroke of the direction of the stroke is often overstepped the boundary phenomenon, after the study of the Chinese printed image found that when the detection of the width of the direction threshold of 5, the effect is optimal, the formula as in (4). After the study of Chinese printed images, it is found that when the threshold of detecting stroke width direction is π/3, the effect is optimal, and the formula is shown in (4): (4) $S w t R u l e 3 = | d_{x} + d_{y} | < \frac{π}{3}$

Where: x is the stroke edge pixel point, d_x is its gradient direction; y is the edge pixel point in the corresponding retrieval direction, d_y is its gradient direction; 4)

The ratio of the variance of the character connectivity domain strokes to the mean of the connectivity domain is reset for the Chinese stroke characteristics, and the formula is shown in (5). (5) $S w t R u l e 4 = \frac{S w t V a r i a n c e}{S w t M e a n} > 3.4$

Where: SwtVariance is the variance of the character connectivity domain strokes and SwtMean is the mean of the character connectivity domain strokes.

3.2.2

Optimization of Chinese Print Detection Algorithm

For Chinese printed images, this paper adopts the method of detecting text lines by quickly extracting them through the horizontal projection histogram, and at the same time setting rules to exclude the projection areas in Chinese printed materials that are obviously not text (similar to the red lines inherent in red-headed documents), as shown in the following formula (6): (6) $Re s u l t 1 = p r o E i n d e x - p r o Sin d e x > = 20$

Where: proEindex is the end index of the text line of the horizontal projection map and proSindex is the beginning index of the text line.

In order to improve the detection speed, the size of the input image as well as the detected text area are set and optimized respectively, and it is found that the detection effect is optimal when the input image with width and height $(W * H)$ greater than 650 × 850 pixels is set according to Equation (7). (7) $i m a g e S i z e = {\begin{matrix} 650 \times 850, i n p u t S i z e > 650 \times 850 \\ i n p u t S i z e, i n p u t S i z e \leq 650 \times 850 \end{matrix}}$

Where: 650, 850 are the width and height of the image, respectively, in pixels; inputSize is the input image size; imageSize is the processing image size;

3.3

Application of image segmentation in document inspection

The machine source authentication technology based on print file identification is a crucial forensic technology in the field of information security. The texture characteristics of print file character images are affected by both texture and structure factors, but for Chinese print files, it is difficult to obtain a large number of identical characters in a file to realize the separation of character factors. Image segmentation of printed character images can eliminate the influence of character structure by extracting features from different regions.

Image segmentation is the technique and process of dividing an image into regions with different characteristics and extracting the target of interest. Threshold segmentation is the most common parallel segmentation method for direct detection of regions, image segmentation method based on threshold selection is to extract the difference between the target object and the background in terms of gray level, and to divide the image into a combination of target and background regions with different gray levels.

The minimum error method belongs to one of the threshold segmentation, the grayscale histogram of an image can be regarded as the probability density function of the mixed distribution of the grayscale of the combined object and background pixels p(z). It is usually assumed that the two components of the mixed distribution, p(z/1) and p(z/2), are normally distributed, with means, standard deviations and prior probabilities of μ₁, μ₂, σ₁, σ₂, p₁, and p₂, respectively, which means that there are (8) $p (z) = \sum_{i = 1}^{2} P_{i} p (z / i)$

where $p (z / i) = \frac{1}{\sqrt{2 π} σ_{i}} \exp (- \frac{z - μ_{i}}{2 σ_{i}^{2}})$ , i = 1, 2

Knowing p(z/i) and P_i, there exists a threshold t that satisfies (9) $P_{1} p (z / 1) = P_{2} p (z / 2) {\begin{array}{l} z \leq t \\ z > t \end{array}}$

This gray scale t is the Bayes Minimal Error Threshold. Taking the logarithm of both sides of the above equation has (10) $\frac{(z - μ_{1}^{2})}{σ_{1}^{2}} + \ln σ_{1}^{2} - 2 \ln P_{1} = \frac{(z - μ_{2}^{2})}{σ_{2}^{2}} + \ln σ_{2}^{2} - 2 \ln P_{2}$

Solving this equation finds the optimal threshold. The parameters μ_i, σ_i and p_i(i = 1, 2) of the mixture distribution are related to the image to be segmented, and the following is a simpler way to solve for the optimal threshold.

With a threshold of t and 0 < t < l − 1, there are (11) $P_{1} (t) = \sum_{z = 0}^{t} h (z), P_{2} (t) = \sum_{z = t + 1}^{l - 1} h (z)$ (12) $μ_{1} (t) = \frac{{\sum_{z = 0}^{t} h (z) z}}{P_{1} (t)}, μ_{2} (t) = \frac{{\sum_{z = t + 1}^{l - 1} h (z) z}}{P_{2} (t)}$ (13) $σ_{1}^{2} (t) = \frac{{\sum_{z = 0}^{t} [z - μ_{1} (t) I z - μ_{1} (t)] h (z)}}{P_{1} (t)}$ (14) $σ_{2}^{2} (t) = \frac{{\sum_{z = l + 1}^{l - 1} [z - μ_{2} (t) J z - μ_{2} (t)} (z)}}{P_{2} (t)}$

The conditional probability e(z, t) of the gray level after binarization can be expressed as (15) $e (z, t) = h (z | i, t) P_{i} (t) / h (z) i = {\begin{array}{l} 1 \\ 2 \end{array}} {\begin{array}{l} z \leq t \\ z > t \end{array}}$

The denominator term h(z) in the equation is independent of i and t and can be ignored, then taking the logarithm of both sides of the above equation and multiplying the result by -2, we will have (16) $\begin{array}{rcl} e (z, t) & = & {[\frac{z - μ_{i} (t)}{σ_{i}}]}^{2} + 2 \log σ_{i} (t) - 2 \log P_{i} (t) \\ i & = & {\begin{array}{l} 1 \\ 2 \end{array}} {\begin{array}{l} z \leq t \\ z > t \end{array}} \end{array}$

It is a measure that reflects the correct classification performance. Based on this, a discriminant function can be defined to describe the correct classification performance averaged over the whole image. (17) $J (t) = \sum_{z} h (z) e (z, t)$

For a known threshold value of t, the discriminant function indirectly reflects the amount of overlap between the Gaussian model of the object and background class distributions. Substituting Eq. (17) into Eq. (16) yields (18) $\begin{array}{rcl} J (t) & = & \sum_{z = 0}^{t} h (z) \times {{[\frac{z - μ_{1} (t)}{σ_{1} (t)}]}^{2} + 2 \log σ_{1} (t) - 2 \log P_{1} (t)} \\ + \sum_{z = t + 1}^{l - 1} h (z) \times {{[\frac{z - μ_{2} (t)}{σ_{2} (t)}]}^{2} + 2 \log σ_{2} (t) - 2 \log P_{2} (t)} \end{array}$

Then substituting Eq. (12)~Eq. (14) into the above equation, we get (19) $\begin{array}{rcl} J (t) & = & 1 + 2 [P_{1} (t) \log σ_{1} (t) + P_{2} \log σ_{2} (t)] \\ - 2 [P_{1} (t) \log P_{1} (t) + P_{2} \log P_{2} (t)] \end{array}$

The gray value thre that makes the discriminant function take a very small value, will be the threshold for the minimum error, i.e. (20) $t h r e = A r g M i n [J (t)] 0 < t < l - 1$

4

Application of artificial intelligence in forensic document examination

As the document examination is a systematic project, the forensic expert often rely solely on a method is unable to make a clear identification opinion, then through a variety of ways and with a systematic appraisal of the document, so that according to the views of all aspects of mutual corroboration, so as to ensure that the accuracy of the identification opinion of the scientific nature of the field of document examination of altered is the direction of the development of the current.

In the field of altered documents test, there are still many difficult to solve the world’s problems, such as the formation of the same time period of forgery documents, or the formation of documents, such as the absolute time of the formation of the current forensic field of research shows that can be completely solved. For the study of altered documents test, not only need to start from the basic work, but also should boldly try to analyze and solve the difficult problems. Therefore, this paper attempts to use artificial intelligence for document examination, to explore its feasibility.

4.1

Physical evidence image pre-processing

Before the document examination, the first physical evidence image preprocessing. Comprehensive consideration of the time and space complexity of the algorithm, for this paper image for lossless compression processing needs, do not need to carry out multi-resolution analysis of quantitative processing images. It mainly seeks the mean and difference of neighboring pixels for the image pixel values to perform the Hough line transformation. After the integer boosting Hough line transform, the pixel values of the transformed image will be reduced in general, and the amplitude range of the values will be narrowed, especially the high-frequency coefficients are the most obvious, so the amount of coded data can be greatly reduced.

In order to explore the feasibility of the idea of using the Hough line transform to compress the image, this paper carries out the experiments of the Hough line transform processing of the image through MATLAB, and analyzes the information entropy and the limit compression ratio of the image after processing. In order to intuitively analyze the change rule of pixel value of the image after the Hough line transformation, the pixel value of the sample original image and the Hough line transformation of this paper were counted by MATLAB respectively. The Hough line transform results of this experiment are shown in Figure 2. It can be seen from the Hough line transformation pixel statistics graph of this paper, the horizontal coordinate is the pixel value, the vertical coordinate is the corresponding number of pixel value, the transformed coefficient value transformation range is small and basically are close to 0 or equal to 0 value. Therefore, the Hough line transform scheme in this paper has a great advantage for realizing image compression.

In order to improve the reliability of the results of this experiment, the sample image was added for the experiment, and the image transformed data were subjected to the calculation of information entropy. The finally obtained comparison of the experimental data results of Hough line transform is shown in Table 1. From the experimental results, it is known that the limit compression ratio of the transformed image obtained by the Hough line transform method used in this paper is significantly higher than the value of the original image after wavelet transform and Fourier transform. The average ultimate compression ratio of the sample image with pixel depth of 8bits after wavelet transform and Fourier transform is 1.7107 and 2.3910, respectively, and the average ultimate compression ratio after the Hough line transform of this paper is 2.9137, therefore, the average ultimate compression ratio of the Hough line transform compared with the wavelet transform and the Fourier transform of this paper is improved by 70.32% and 21.86%, respectively.

Table 1.

Comparison results of Hough line changes

Image	Entropy of information	Ultimate compression ratio
Image 1	6.9735	1.203
Wavelet transform	5.0382	1.703
Fourier transform	4.0284	2.198
Hough line transform	2.7653	2.974
Image 2	7.0843	1.148
Wavelet transform	4.8863	1.718
Fourier transform	3.7826	2.286
Hough line transform	2.5973	2.864
Image 3	7.1084	1.129
Wavelet transform	4.8691	1.711
Fourier transform	3.7742	2.689
Hough line transform	2.5038	2.903

4.2

Document homogeneity test

Add and change the document refers to the original real document to add, rewrite part of the content so as to change the original real document and the formation of the content of false suspicious documents. In the actual case, self-added to make the handwriting characteristics can not be used, and if the writing tools used before and after the addition of the same color composition, it can not be differentiated according to the color difference, most of the original test methods can not play a role, so self-added to change the document test is relatively more difficult.

This paper adopts the proposed Chinese text detection algorithm for document identity analysis. In order to verify the reliability of the algorithm in different situations, the use of handwriting three-dimensional information extraction analyzer on the same person in different liner conditions for the extraction of three-dimensional information on the written text, the algorithm to detect the difference in the width of the strokes and indentation characteristics of the differences in the comparative analysis.

A total of 50 school students were selected, numbered 1~50. The content of the writing was “RMB 200,000 Yuan has been paid back”, and the 50 writers were instructed to take the bill paper and A4 paper as the direct bearer, and write the experimental content 5 times under 5 kinds of liner conditions, such as the mouse pad, 10 sheets of bill paper, 10 sheets of A4 paper, wooden desk, and glass plate, at the normal writing speed. The indentation depth of handwriting was quantitatively analyzed, the number of sampling points of each stroke, the location of the sampling point, the indentation depth of the sampling point were counted, and the indentation depth curve was drawn according to the indentation depth of the sampling point. Taking the second stroke of “Wan” written by No. 1 writer as an example, the indentation depth curve written by the writer on the ticket paper and A4 paper under 5 kinds of padding conditions is shown in Figure 3.

It can be found that the indentations of the strokes written under five different liner conditions correspond to five different curves. The basic trend of the five curves was the same, with similar indentation depth characteristics occurring at the same locations, and the average position of the indentation depth curves varied with the softness of the liner. On the ticket paper, the indentation depths under the five liner conditions were -0.046±0.009, -0.036±0.006, -0.037±0.007, -0.029±0.006, and -0.025±0.007, respectively. On A4 paper, the indentation depths were -0.045±0.009, -0.036±0.006, -0.036±0.006, -0.027±0.006, and -0.025±0.007 for the five liner conditions. Using the proposed algorithm for detection, it is found that the detection results change with the increase of Δh, and the model prediction error is <1.5 px. Under the same liner condition, the Δh of A4 paper is lower than that of bill paper, and the difference in stroke width also decreases. This proves that the algorithm is able to automatically adjust the detection threshold through paper type classification and has the ability to compensate for material differences.

The indentation depth curves of the second stroke of the Chinese character “Wan” written five times by No.1 writer on the note paper and A4 paper under the liner condition of 10 sheets of A4 paper are shown in Fig.4, respectively. The mean value of Δh is 0.012 μm and the standard deviation σ is 0.0009 μm (CV=7.5%) for five times on the bill paper, and the mean value of Δh is 0.009 μm and the standard deviation σ is 0.006 μm (CV=6.7%) for five times on the A4 paper. And the detection with the proposed algorithm shows that σ=0.18px (86% agreement) on the bill paper and σ=0.22px (85% agreement) on the A4 paper, which proves that the algorithm’s suppression efficiency of writing fluctuation reaches more than 85%.

Therefore, the algorithm proposed in this paper has the ability to compensate for material deformation and cross-boundary mass stability, which is suitable for document examination in forensic identification.

4.3

Printed Document Inspection

Print file identification, the purpose is to identify the machine source of the print file. Printer in the production of the parameter configuration and the use of the process of wear and tear will form the unique characteristics of the printer, by scanning the print file to take a picture, these features are reflected in the print file image, the formation of different texture characteristics. The texture characteristics of print file character images are affected by both texture and structure factors. The idea of this paper is to eliminate the influence of character structure by performing character segmentation based on the minimum error method on print character images and feature extraction on different regions.

By extracting LBP features from a large number of printed character images, it is found that the overall features conform to the normal distribution, so this paper adopts the two-factor variance model to analyze the feature values, and tries to dig out the significant influencing factors among the feature values.

The LBP operator used in this paper for two-factor analysis is the rotationally invariant equivalent mode LBP operator with circular neighborhood radius R = 1 and neighborhood sampling point P = 8. There are a total of 10 modes in the LBP binary code, and the texture features are represented as normalized LBP histograms.

The 10-dimensional LBP features are extracted from the character image to form a feature vector, and each dimensional feature is denoted as $x_{i j}^{l}$ . Assuming that each dimensional feature obeys the normal distribution, there are $x_{i j}^{l} ~ N (μ_{i j}^{l}, σ^{2})$ , $μ_{i j}^{l}$ , and σ² are unknowns, where the subscripts i = 1, 2, ⋯, N denote N different models of printers, j = 1, 2, ⋯, M denote M different characters, and l = 1, 2, ⋯, 10 denote the 10-dimensional component of the LBP feature vector.

The two-factor model is established as follows: (21) $x_{i j}^{l} = μ^{l} + α_{i}^{l} + β_{j}^{l} + ε_{i j}^{l}$

$ε_{i j}^{l} ~ N (0, σ^{2})$ , each $ε_{i j}^{l}$ independent of the other. (22) $i = 1, 2, \dots, N; j = 1, 2, \dots, M; l = 1, 2, \dots, 10$

Where μ^l represents the total average; $α_{i}^{l}$ represents the effect of different printer models on the LBP eigenvalues, called the printer factor effect; $β_{j}^{l}$ represents the effect of different characters on the LBP eigenvalues, called the character factor effect; and $ε_{i j}^{l}$ represents random noise. μ^l, $α_{i}^{l}$ , and $β_{j}^{l}$ can all be obtained using sample estimation: (23) ${\hat{μ}}^{l} = \frac{1}{N M} \sum_{i = 1}^{N} \sum_{j = 1}^{M} x_{i j}^{l}$ (24) ${\hat{μ}}_{i}^{l} = \frac{1}{M} \sum_{j = 1}^{M} x_{i j}^{l}, i = 1, 2, \dots, N$ (25) ${\hat{μ}}_{j}^{l} = \frac{1}{N} \sum_{i = 1}^{N} x_{i j}^{l}, j = 1, 2, \dots, M$ (26) ${\hat{α}}_{i}^{l} = {\hat{μ}}_{i}^{l} - {\hat{μ}}^{l}, i = 1, 2, \dots, N$ (27) ${\hat{β}}_{j}^{l} = {\hat{μ}}_{j}^{l} - {\hat{μ}}^{l}, j = 1, 2, \dots, M$

ANOVA was used to test the significance of the effect of the two types of factors on the eigenvalues of LBP. Notation: 1)

The total sum of squares (28) $S_{T} = \sum_{i = 1}^{N} \sum_{j = 1}^{M} {(x_{i j}^{l} - {\hat{μ}}^{l})}^{2}$

2)

Printer factor effect sum of squares (29) $S_{A} = M \sum_{i = 1}^{N} {({\hat{μ}}_{i}^{l} - {\hat{μ}}^{l})}^{2}$

3)

Sum of squares of character factor effects (30) $S_{B} = N \sum_{j = 1}^{M} {({\hat{μ}}_{j}^{l} - {\hat{μ}}^{l})}^{2}$

Sum of squared errors: (31) $S_{e} = S_{T} - S_{A} - S_{B}$

According to the F-test theory, if (32) $F_{A} = \frac{S_{A} / (N - 1)}{S_{e} / (N - 1) (M - 1)} \geq F_{α} (N - 1, (N - 1) (M - 1))$

Then it means that α₁ ≠ α₂ ≠ ⋯ ≠ α_N significantly holds, i.e., the effect of the printer factor is significant. If (33) $F_{B} = \frac{S_{B} / (M - 1)}{S_{e} / (N - 1) (M - 1)} \geq F_{a} (M - 1, (N - 1) (M - 1))$

Then it means that β₁ ≠ β₂ ≠ ⋯ ≠ β_M significantly holds, i.e., the effect of the character factor is significant.

In order to verify the significance assumptions of the above model for the printer factor and the character factor, 10 different models of printers are selected for the experiments on the printed documents, that is, N = 10, and the level range of the printer factor $α_{i}^{l}$ is i = 1, 2, ⋯, 10; the same 20 different Chinese characters are selected for each printer, that is, M = 20, and the level range of the character factor $β_{j}^{l}$ is j = 1, 2, ⋯, 20; the level of significance is taken to be α = 0.01 for the two-factor analysis.

The two-factor analysis experiments of extracting LBP features for the full map of character images are shown in Fig. 5. The experiment shows that when the significance level is taken as α = 0.01, there are 4-dimensional features which are very significantly affected by the printer factor, but also by the character factor.

In the problem of print document recognition, the character factor is an interfering factor that affects the recognition results. Only by eliminating the influence of the character factor in the LBP features, the recognition accuracy can be improved.

In order to verify that the feature extraction method proposed in this paper eliminates the influence of the character factor on the LBP features, the character image is segmented into character regions, and then the LBP features of the three regions are extracted separately. The 10-dimensional LBP features of each region are subjected to two-factor analysis, and the results are shown in Fig. 6, Fig. 6(a), Fig. 6(b), and Fig. 6(c), which show the analysis results of the interior, edge, and background of the character image, respectively.

The experiments of two-factor analysis show that: 1)

Whether the LBP features are extracted for the whole image or for the three regions after segmentation, the 1st, 2nd, and 10th dimensional features are invalid because the 1st and 2nd dimensional features represent the LBP patterns of “bright spots”, i.e., the points where the center pixel value is smaller than the neighboring pixel value, and the 10th dimensional features indicate that the jumps from 0 to 1 are more than the neighboring pixel value, while the 10th dimensional features indicate that the jumps from 0 to 1 are more than the neighboring pixel value. The 10th dimensional feature indicates that the LBP pattern has more than two jumps from 0 to 1. The statistical histograms of these two types of patterns are very similar in most of the character images, and therefore fail in the two-way ANOVA.

2)

Comparing the results of two-factor analysis of LBP features before and after character region segmentation, it can be seen that after character region segmentation is performed, the 3rd to 9th dimensional features in LBP features are only affected by the printer factor but not by the character factor, which indicates that the feature extraction method based on the minimum error method for character region segmentation proposed in this paper is effective.

In order to explore the effect of image segmentation processing on the final recognition results, Raman spectral analysis is introduced for further verification. Three documents are recognized using the algorithm of this paper, showing that document 1 and document 3 come from the same printer. Following the spectral analysis using Raman spectroscopy, the results are shown in Fig. 7. The Raman spectra of file 1 and file 3 are basically similar, verifying the effectiveness of the method proposed in this paper.

5

Conclusion

In this paper, an artificial intelligence algorithm is used for document examination, and its feasibility is verified through experiments.

The average ultimate compression ratio of two sample images with pixel depth of 8bits is 1.7107 and 2.3910 after wavelet transform and Fourier transform, respectively, and the average ultimate compression ratio after Hough line transform processing is 2.9137, which is nearly 70.32% and 21.86% higher than the average ultimate compression ratio of wavelet transform and Fourier transform. The proposed Chinese text image detection algorithm is utilized for detection, and it is found that the detection results change with the increase of Δh, and the model prediction error is <1.5px. Under the same liner condition, the Δh of A4 paper is lower than that of bill paper, and the difference in the width of the strokes also decreases. It proves that the algorithm can automatically adjust the detection threshold through paper type classification and has the ability to compensate for material differences. The mean value of Δh for five times writing on bill paper is 0.012 μm, and the standard deviation σ is 0.0009 μm (CV=7.5%). The mean value of Δh for five times writing on A4 paper is 0.009 μm, and the standard deviation σ is 0.006 μm (CV=6.7%). And the detection with the proposed algorithm shows that σ=0.18px (86% agreement) on the bill paper and σ=0.22px (85% agreement) on the A4 paper, which proves that the algorithm’s suppression efficiency of writing fluctuation reaches more than 85%. After utilizing the minimum error method for character region segmentation, the 3rd~9th dimensional features in the LBP features are only affected by the printer factor, but not by the character factor.

The research results in this paper provide a solution with both deformation adaptability and high differentiation for document examination in forensic identification, which has important application value.

Language:: English

Publication timeframe:: 1 times per year
Journal Subjects:: Life Sciences, Life Sciences, other, Mathematics, Applied Mathematics, General Mathematics, Physics, Physics, other

Journal RSS Feed

Application and Feasibility Assessment of Artificial Intelligence in Forensic Identification and Document Examination

Jing Ye

Leya Zhang

Published Online: Sep 29, 2025

Received: Jan 17, 2025

Accepted: Apr 27, 2025

DOI: https://doi.org/10.2478/amns-2025-1127

KeywordsForensic identification, Document inspection, Hough line transformation, Stroke width algorithm, Minimum error method

© 2025 Jing Ye and Leya Zhang, published by Sciendo.

This work is licensed under the Creative Commons Attribution 4.0 International License.

Keywords
Forensic identification, Document inspection, Hough line transformation, Stroke width algorithm, Minimum error method