Identifying the Personality Traits Using Handwriting Recognition in a Real-Time Environment

The resurgence of handwriting analysis in personality detection holds promise for advancing research, practice, and innovation across various fields. By leveraging the unique insights provided by handwriting analysis, professionals can enhance their understanding of human behavior, improve diagnostic accuracy, and inform evidence-based interventions. Furthermore, the integration of handwriting analysis into digital platforms and technologies enables real-time assessment and feedback, expanding its accessibility and usability in diverse settings. Personality detection using deep learning is an emerging and interesting field of interest for many, due to its uniqueness of identifying a person’s character traits using just his/her handwriting. An individual's distinct identity is defined by a set of distinguishing qualities, habits, and cognitive patterns that make up their personality. It is what defines who the person is, how they would behave or interact in public, how they would work in different environments, cope with various situations, manage their tasks, etc. According to graphology (or handwriting analysis), how someone writes, how they shape their letters, and how they space them can indicate a lot about who they are [1, 2]. Study of handwriting is essential to forensic investigations, where it is used to identify and authenticate handwritten documents, detect forgeries, and link suspects to crime scenes. Graphology frequently shares its findings in courtrooms, employment offices, and police stations, despite the fact that many people view it as a pseudoscience. The graphologist Michelle Dresbold [3] mentioned that "Every single character represents something related to your mental state" and “Handwriting analysis is not like tea leaves, despite what some people believe. Your brain is causing your body to move”. Traditionally graphology has been going on for centuries. Additionally, since 1895, more than 2,200 researchers have presented their findings in articles related to medicine, education, and psychology [4]. A study carried out in the United States by the National Pen organization claims that an individual's writing style can provide information about 5,000 distinct personality features depending on how they sign their name, space their letters, and even connect them. The importance of graphology as a subject started to spread in the 1930s. Dr. Ludwig Klages, a well-known German philosopher and the founder of contemporary graphology, produced several publications. In 1965, Francis T. Hillige devised a method for assessing characteristics in a sample of handwriting. Graphology has also seen many advancements since then, such as not just predicting whose handwriting it is, but also, predicting the age and gender of the person, predicting any mental illness or disorder or imbalance one might possibly possess, predicting the state of mind of the person while they were writing and predicting their personality traits [5, 6].

This proposed system, focuses on the latter advancement of graphology, with the help of technology, namely deep learning (DL). With the use of DL, Human intervention is much less, thus the variance in the result is also much less which ultimately increases precision in the final predictions or results. Not only does our project predict personality traits through handwriting analysis, but it also does that in a real-time environment. Deep learning has been widely used to recognize handwriting [7, 8]. Text is evaluated after it has been written in off-line handwriting recognition. The binary output of a character on a background is the only data that can be studied. Even though switching to a digital stylus for writing now provides more details like pen stroke, pressure, and writing speed, offline solutions are still required when online is unavailable. For historical records, archives, or the widespread digitization of handwritten forms, it is especially important. Real-time tools for finding out the Personality Traits of a person, on the basis of their Handwriting, were not available. Real-time tools would make it much easier and more convenient for the person to know their personality quickly and without much hassle. Such a tool would have no human intervention. Also, inefficiencies caused by image inputs for such tools would no longer be a problem. Usually for someone to know their personality traits, they have to answer a questionnaire, but with this tool, just with the simple handwriting input, one could know their traits, in real-time.

Higher extroversion was linked to a less logical decision-making style, but higher conscientiousness and agreeableness were significantly linked to a more logical decision-making style. While extroversion and openness to new experiences were strongly linked with a higher intuition style, higher conscientiousness, and agreeableness were highly associated with a lower intuition style. A more dependent decision-making style was significantly associated with greater levels of conscientiousness and agreeableness, whereas a less reliant decision-making style was strongly associated with higher levels of openness to new experiences. More agreeableness, conscientiousness, and neuroticism were significantly associated with a less spontaneous decision-making style [9]. When humans are taught how to write, they don't continue to write the same way all their lives. As time progresses, an individual’s handwriting keeps changing. This is due to the change in behavior or personality of an individual. Since handwriting, which is as distinctive as DNA, is the pattern of our psychology showed in symbols on paper. When you get to know someone's handwriting well enough, you can identify their writing just like you would a famous painting or photo. Graphology is based on the idea that each person's handwriting has a distinct personality that is entirely reflective of their individual personality. Therefore, in particular, the style and multi-modality highlight the significance and standing of graphology as a linguistic level of study [10]. Additionally, it is observed that handwriting based on graphology indicates an individual's inner feelings, even though these traits are not obvious in their actions [11].

By examining handwriting features psychologists can gain insights into clients' emotional states, cognitive patterns, and underlying psychological dynamics. This information can inform therapeutic interventions and treatment strategies. The purpose of Graphology is to identify various people from all walks of life, by just analyzing, observing, and studying their handwriting. Due to its extensive applicability in a variety of sectors, including education, psychology, medicine, criminal detection, marriage counseling, commerce, and recruitment, graphology-based behavioral analysis has become more and more popular in recent years [12]. Graphology lets us analyze a personality without considering extra factors such as lifestyle, appearance, social status, physical features, etc. Graphology focuses on handwriting features such as the baseline of the handwriting, the stroke of handwriting, the slant of letters and the words overall, the pressure of the pen, the word spacing, letter spacing, and overall sentence spacing and sometimes also paragraph spacing of the writing. A lot of methods are present for handwriting analysis but are often proven to be biased as it has human intervention, which introduces bias and variation in the results. Conventional personality assessment techniques may not be able to process results instantly and place a significant emphasis on subjective interpretation, which could introduce bias into the analysis. Furthermore, a lot of preprocessing and manual input is often needed for these approaches, which limits their scalability and accessibility. By utilizing a CNN model for real-time handwriting analysis, our suggested system seeks to overcome these drawbacks by offering objective insights into personality traits with little assistance from humans and efficient processing.

There are techniques that use a signature for identification in order to analyze behaviors including feelings, emotions, and human identifications [13, 14]. But unlike graphology, which takes into account a portion of a character, these techniques take into account the entire signature when making predictions. Furthermore, research on aesthetic analysis identifies human behaviors like beautiful, non-beautiful, great, and poor writing using handwritten letters or image attributes [15, 16]. These techniques are restricted to particular sorts of behavior, though. Techniques for evaluating a person's personality, including their interests, attitudes, and relationships with their family and community, have been offered [17, 18]. These techniques analyze user-typed texts; they do not analyze handwriting documents. Generally speaking, the techniques call for full-text words in order to evaluate personality. Therefore, graphology-based handwriting analysis becomes vital to forecast personal behaviors that may be imperceptible to the human eye [19]. The Big Five factor [20] is a popular model for characterizing a person's personality. Five fundamental personality traits serve as its foundation. Additionally, authors created the first dataset in the literature that includes both specified and arbitrary texts and correlates the Big Five personality handwriting types gathered from 128 people [21]. Alamsyah et. al. explained in his research three important features size, slant-ness and character stroke which is well-known features for any type of handwriting [22]. They had implemented CNN model to recognized Big five personality traits and achieved 80.88% of accuracy. The systematic review is performed between the handwritten writing and the main five personality traits. Various features for each category are explained by Elngar et al. [23]. Using deep learning CNN model the prediction is performed. The finding suggested the limitation of the diverse handwritten data samples and limited research in other regional languages for personality recognition.

Although behavioral assessments and personal questionnaires for self-reporting are the fundamental components of conventional personality assessment techniques, they are time-consuming, biased, and frequently unreliable. Moreover, the complicated and varied features of an individual's personality are often not well communicated by these methods. On the other hand, handwriting analysis offers a unique opportunity to gain an understanding of personality traits; yet, manual analysis is labor-intensive and requires ability. Thus, this research's primary goal is to develop a computerized system that can recognize traits of personality from handwriting with greater accuracy than existing methods. It improves the handwriting assessment process's speed, scalability, objectivity, and reliability. Based on the literature survey conducted for the proposed system, the conclusions were derived for the features of the 5 personality predictions as shown in Figure 1 using the CNN model. CNN is very effective for handling images, identifying universal traits, and sorting through them as compared to other DL models like RNN. Mainly CNN models are simple to design and our data is in image format and not required time series data like RNN. Hence our proposed system is implemented using CNN.

301.png

Figure 1. The big five personalities and their features based on the literature survey

Further in the paper, Dataset Creation, Network Architecture, Execution Details, Comparative Analysis, and Results for the Proposed System have been discussed. The framework of the paper is as follows: Section 2 Details about Dataset Corpus Creation Section 3 Discussion about the Proposed System, the Preprocessing details and CNN Architecture Section 4 Experiments Results and Discussion contains Model Training details, Execution Details, and Comparative Analysis and Section 5 Conclusion.

The suggested approach employs machine learning to analyze handwriting in order to predict personality. Essentially, it's an improvement on graphology. The use of machine learning reduces the amount of human intervention, which reduces the variance in findings and eventually improves the accuracy of the final predictions or results. Not only does the project predict personality traits through handwriting analysis, but it also does that in a real-time environment.

3.1 Preprocessing

Preprocessing is an important step in handwriting character recognition to prepare image data for model input. Preprocessing plays a crucial role, enabling the machine model to learn and extract the given features as output. As the CNN is fully connected layers, it needs all the input images in an array of the same size. Preprocessed images boost the speed of the model and work efficiently [24, 25]. In our system, Preprocessing on raw data or image data includes grey scaling, noise reduction, resizing, and background removal. The preprocessing procedure improves the quality of the visual input, allowing the model to understand and produce the best possible result.

3.1.1 Image background

As mentioned above, CNN takes the input images in an array of the same size; so, we pass the input image shown in Figure 2, on the white background canvas such that its aspect ratio will remain unchanged. By applying a white background, it reduces the unwanted noise and helps to enhance the network’s ability to extract meaningful features as shown in Figure 3. Maintaining the aspect ratio is very important to avoid the deformations of features. Due to the whitening of the background, structures or patterns are boldly visible which will also improve the efficiency of feature extractions.

302.png

Figure 2. Raw input image

303.png

Figure 3. Image after pasting the raw input on a white canvas (background)

3.1.2 Gray scaling

Gray Scaling is an image conversion technique. It converts the colored image into a monochromatic image i.e., into a single-colored image. By applying grey scaling features to our image data it avoids the false classifications and reduces the color complexities to make it easier for the model training. Gray Scaling of image data does not reduce the size of the image as it takes the same space as the colored image as depicted in Figure 4.

304.png

Figure 4. Input image after gray scaling

3.1.3 Noise reduction (median blur)

The process of reducing background noise from an image is known as noise reduction. Image noise reduction is very important as there are some factors that affect the quality of the image. Noise Reduction Algorithm alters the signal to some degree [26]. The Median Blur is used for noise reduction in our system. It reduces the background noise of images without affecting the original image or blurring it. The median filter is faster and provides superior outcomes in contrast to the other filters. The output of this is depicted in Figure 5.

305.png

Figure 5. Input image after performing noise reduction

3.1.4 Resizing

Maintaining a consistent size of data before feeding it to the CNN results in fewer distortions of features and patterns. We have taken image resolution in the range of 256 × 256 pixels for training CNN. There are two ways of resizing the image: one is by scaling down the image and another is by cropping the image. Though there is an option of cropping the image, it poses risks of losing some important features or patterns. Rescaling the image will distort the image features and patterns. Hence resizing or rescaling the image is more suitable for our data inputs. Figure 6 shows the output of resizing.

306.png

Figure 6. Input image after resizing

3.2 Analysis of features

The dataset includes a distinct set of handwritten patterns for each of the 5 traits of personality. Classification of these samples is based on the following features:

Baseline: A word's baseline. Baseline can be ascending, descending straight, etc.

Space: Spaces between words like normal, wide or narrow.

Slant: A writing way like moderate right/left, Extreme right/left, straight, etc.

Some samples are shown in Figure 1 based on each personality trait. The equilibrium between the ego and knowledge is shown by the baseline. The distance that the writer wishes to keep between themselves and society i.e., their boundaries is symbolized by the spaces between words. Different types of slanting information will explain the different personalities like extroverts, the openness of the person.

3.3 CNN architecture

3.3.1 Brief

A system composed of millions of artificial neurons connected to one another is called an artificial neural network (ANN). Each node is called a perceptron which computes and processes inputs and forwards the intermediate output to the next layer.

ANN comprises 3 layers, first is the input layer which is responsible for taking information as input from the environment and forwards it to the next layers which are called hidden layers. Hidden layers are sets of layers that are responsible for processing the input and extracting information from them. They process the information provided by input layers and forward that to the output layer. The output layer is responsible for giving out the output to the system. The output varies with the type of problem for which the model is built. In the case of regression problems and probabilities in the case of classification problems, CNN is a type of feed-forward Neural Network that needs and processes input in matrices-like structures like images. It is widely used in computer vision tasks, including image and speech classification, object detection, and image segmentation [27]. One of the important advantages of using a CNN model is its ability to learn features by itself. CNN comes with automatic feature extractions. Before CNN, researchers had to do manual feature engineering which used to be very time-consuming and resource intensive. But with the help of CNN feature extraction is done automatically [28].

307.png

Figure 7. CNN architecture

CNN network consists of mainly three types of layers as shown in Figure 7:

Convolutional layer: This layer is the core of the CNN models. Here input data is processed through a tiny filter, often known as a kernel (typically a 3×3 matrix) [29].

Pooling layer: The major responsibility of a pooling layer is Dimension reductions mainly known as downsampling. It reduces the number of parameters of the input. Max pooling and average pooling are frequently used pooling techniques.

Fully-connected (FC) layer: Each neuron is connected to the subsequent layer neurons. These layers serve as a classifier to the CNN model. They perform the task of classifying the inputs based on the previously learned features and their filters. One of the important aspects of CNN models is the activation functions used. They introduced non-linearity to the model which enables it to learn and represent complex relationships of the data used [30]. Various activation functions can be used such as Sigmoid, Tanh, ReLU, SoftMax, etc. Choosing the most optimal activation function depends on the type of problem the model is built for. Various factors influence the choice of activation function like non-linearity, gradient Vanishing and Exploding problem, Computational Efficiency, output range, the problem domain of the model, etc. [31].

The next important task while fitting the CNN model is to choose the optimal optimizer. An optimizer is an algorithm or technique used to adjust a machine learning model’s parameter as during training. An optimizer seeks to minimize the loss function or maximize the objective function of the model. Popular optimizers include Adam, Stochastic Gradient Descent (SGD), RMSProp, Adagrad, and others [32].

3.3.2 Proposed architecture

The proposed system consists of 4 pairs of Convolutional Layers and Max Pooling Layers followed by flattened and dense layers as shown in Figure 8. All the Convolutional Layers are with the kernel size 3×3 with ReLU as the activation function. The MaxPooling is with a pool size of 2×2.

308.png

Figure 8. Proposed model architecture

The first layer of conv2D is with input size (None, 256, 256, 32) and output shape (None, 254, 254, 32). It consists of a total of 32 kernel filters. It is followed by max_pooling2d with an output shape of (None, 127, 127, 32). The second convolutional layer conv2d_1 (Conv2D) is with the output shape of (None, 125, 125, 64). It contains a total of 64 kernel filters. It is again followed by a second max pooling layer max_pooling2d_1 with an output shape of (None, 62, 62, 64). The third convolutional layer conv2d_2 (Conv2D) is with the output shape of (None, 60, 60, 128). It contains a total of 128 kernel filters. It is then followed by the third max pooling layer max_pooling2d_2 with an output size of (None, 30, 30, 128). The last as in the fourth convolutional layer consists of 256 kernel filters with an output shape of (None, 28, 28, 256) and it is followed by the fourth max pooling layer max_pooling2d_3 with an output shape of (None, 14, 14, 256).

After this, the input is sends to the Flatten layer with an output shape of (None, 50176), which converts the two-dimensional input data from the last max pooling layer to one dimensional as the dense layers also known as fully connected layers process upon one dimensional data.

Then the flatten output is then propagated through the first dense layer with an output shape of (None, 256). Dropout layer of 0.5 is added which indicates that during training, 50% of the input units will be randomly set to 0 at each update, which aids to avoid overfitting problems.

Last layer is the final dense layer with an output shape of (None, 5) which means it has output a vector of length 5. (As our model is built to classify input into one of the 5 classes, namely, Extroversion, Agreeableness, Conscientiousness, Openness and Neuroticism.). Softmax is used as the activation function as our model is built for multi-class classification. It takes a vector as an input and gives the probability distribution of various classes involved. The sum of all the probabilities is always one. The loss calculated here is Categorical Cross entropy which is usually used in case of multi-class classification problems, where each sample can belong to one and only one class. It calculates the difference between the expected probability and the true labels. It is given by Eq. (1).

Cross Entropy(y,y)= -∑ yi log(yi)            (1)

This loss function also penalizes the model more when the predicted probabilities deviate from the true labels. During training, the goal is to decrease the loss produced by adjusting the values of the weights and bias until optimal values are not found.

The Adam optimizer is used to execute the model. It keeps track of the rapidly dropping mean of the previous gradient (u_t), which represents the mean of the first moment, and the previous squared gradient (a_t), which represents the variance of the second moment The a_t and u_t are calculated is shown in Eqs. (2) and (3) [32].

a_t = β₁a_t-1 + (1-β₁) d_t                     （2）

u_t = β₂u_t-1 + (1-β₂) d_t²                 （3）

where, β represents the decaying rate of the gradients and dt is the gradient calculated at the time t. Adam optimizer uses Gradient descent (GD) algorithm as a base during training to modify training parameters and is also useful in self-learning [33]. Adam optimizer comes with many advantages, mainly having an adaptive learning rate, and high efficiency as it merges the features of AdaGrad and RMSProp. It is robust to noisy gradients and requires very little manual tuning [34].