INTEGRATING COMPUTER VISION AND MACHINE LEARNING MODELS FOR HANDWRITING ANALYSIS

Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4757 INTEGRATING COMPUTER VISION AND MACHINE LEARNING MODELS FOR HANDWRITING ANALYSIS NAGAVALI SAKA1, ABHINAV KUMAR2, KOLLURU SURESH BABU3, KONALA PADMAVATHI4, JAGADEESWARA RAO PALISETTI5, PADMAVATHI PANGULURI6, G LAKSHMI7* 1Department of CSE, Ramachandra college of engineering, Eluru, Andhra Pradesh, India 2Software Development Manager, Amazon, Seattle, USA 3Department of CSE,Vasireddy Venkatadri International Technological University, Andhra Pradesh, India 4Department of CSE(AI&ML), Aditya University, Surampalem, Andhra Pradesh, India 5Department of CSE(AI&ML), Lakireddy Bali Reddy College of Engineering, Andhra Pradesh, India 6Department of CSE, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, India *7Department of IT, PVP Siddhartha Institute of Technology, Vijayawada, Andhra Pradesh, India E-mail: [email protected], [email protected], [email protected], [email protected] , [email protected], [email protected] , *[email protected] ABSTRACT Handwriting has emerged as an essential characteristic in many applications, including forensic science, signature verification, and document authentication. Even though optimal character recognition (OCR) and machine learning (ML) have evolved more, the different handwriting styles combined with real-time processing issues have prevented current systems. For this reason, this study introduces a system that combines Computer Vision (CV) for preprocessing handwritten images with deep learning models CNN and LSTM for handwriting analysis. We believe integrating all these advanced techniques would improve handwriting analysis systems' accuracy and scalability. Moreover, our approach includes behavior capturing, which provides psychological insight based on the learned psychological cognition state of the writers, by using dynamic pen features such as pressure, writing speed, and stroke order. Compared to current state-ofthe-art solutions, the proposed hybrid model has improved accuracy, real-time performance, and adaptability to different handwriting styles. This improves the accuracy of handwriting recognition and establishes a new avenue for behavioral profiling, potentially impacting fields like forensic investigation, psychological evaluation, and signature verification. Finally, the study concludes with prospects related to adding more behavioral traits and optimizing real-time processing capabilities. Keywords: Handwriting analysis, Computer Vision, Convolutional Neural Networks, Recurrent Neural Networks, Signature verification, Document forensics, Deep learning 1. INTRODUCTION Handwriting analysis has critical applications, from forensic science to signature verification and document authentication to behavioral profiling. Even as access to digital data increases, handwriting is a distinctive and reliable form of personal communication that provides information on an individual’s cognitive state and personality traits. Therefore, handwriting recognition, verification, and analysis systems play a significant role in modern technology, such as e-signature verification, document forensics, and psychological profiling. Yet, traditional handwriting recognition systems face challenges due to the diversity of handwriting styles, variations in stroke speed, pressure, and writing order, which complicate accurate and scalable analysis [1][2]. The variability of human handwriting is the main obstacle in handwritten recognition. Due to various factors like writing style, emotional state, and physical condition [3][4], even trained specialists have difficulty interpreting handwritten text accurately. In addition, real-time processing and large-scale adaptability are still challenging issues even though machine learning (ML) and deep learning (DL) approaches (e.g., Convolution Neural Networks (CNN) and Recurrent Neural Networks (RNN)) have exhibited considerable progress in automated handwriting recognition [5][6]. Moreover, adding behavioral profiling to these systems refers to analyzing how the user presses the Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4758 pen, the speed, and the order in which strokes are made, which is still an unexplored area [7][8]. Previous handwriting recognition approaches were rule-based solutions or statistical models like Hidden Markov Models (HMMs) [9][10]. In contrast, contemporary solutions use more complex ML models that learn from extensive datasets covering a range of handwriting styles. As an example, convolutional neural networks (CNNs): Models widely use these networks for image extraction of features. Including CNNs in neural networks makes them experts at learning hierarchical representations of data, thus enhancing recognition efficiency [11][12]. Likewise, LSTM networks, as a variant of RNN, work well for sequential data and have been proven to benefit handwriting recognition by capturing time dependencies in stroke sequence [13][14]. However, these systems are still challenged by the need for real-time performance, particularly for large datasets in some applications, including e-signature verification and document analysis [15]. This study aims to capture the noted weaknesses with existing systems, proffering an integrated system that synergizes CV-based preprocessing techniques with CNNs and LSTMs for handwriting recognition, followed by behavioral profiling. This work is unique as it combines these technologies to improve recognition accuracy and offer a psychological perspective by studying handwriting evolution characteristics such as stroke pressure, writing speed, and stroke order. Our system combines behavior profiling to provide a holistic insight that standard systems intermediate do not possess, as they rely only on text fulfillment. As a result, you can expect better performance on reallife tasks where tracking and analyzing behaviors in real time are essential. Figure 1: Unified Approach to Handwriting Recognition Figure 1 depicts a comprehensive pipeline of CV techniques working with ML models to address this hybrid flow in handwriting recognition, complementing handwriting analysis for effective recognition. Tools/Techniques Involved These techniques make handwritten data clean and usable and pass more data to machine learning models such as CNN and LSTMs for recognition/sequencing. Combining Static and Dynamic Approaches for Enhanced Handwriting Analysis: By combining both static and dynamic approaches, the joint technique can improve handwriting analysis for more measurably accurate and reductively stable handwriting recognition, and the method has the potential to highlight dynamic features like stroke order, pressure, and writing speed among others. With this united approach, handwriting recognition systems have become much more computationally efficient and performance. Here, we present a hybrid method that uses CV for preprocessing handwritten images and deep learning models CNN and LSTM for analyzing handwriting. We think combining all these advanced approaches would make handwriting recognition systems more accurate and scalable. While best-handling optical character recognition (OCR) and machine learning (ML) have come a long way, there are no existing systems that could, due to the diversity of two different handwriting types (e.g. even one person can't write what they usually do) along with a real-time processing issue. To this end, we propose a system which preprocesses handwritten images using Computer Vision (CV) and analyses them using deep learning models CNN and LSTM. The rest of the paper is organized as follows: Section 2 overviews related work in handwriting recognition, signature verification, and behavioral profiling. Section 3 describes the methodology and its integration for preprocessing and recognition through CV and ML models, and Section 4 presents the results. Section 5 closes the paper and highlights future work directions. 2. RELATED WORK This has been a significant research area for decades and has made many improvements due to Machine Learning (ML) and Deep Learning (DL) techniques. Traditional approaches, including template matching and Hidden Markov Models (HMMs) [16], were often used in the early systems for handwriting recognition. However, these techniques had difficulties coping with this large variability in handwriting styles, writing pressure, and speed. Deep learning algorithms, especially CNNs, revolutionized the field by enabling automatic feature extraction without any adaptation to the style of the text, leading to an increase in accuracy in handwriting from varying styles [17][18]. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4759 Convolutional neural networks (CNNs) have proven very efficacious at learning hierarchical features from images, and they quickly became a standard component of modern handwriting recognition systems. CNNs are especially good at recognizing spatial patterns in an image, such as the edges and shapes involved in writing a number [19]. Studies have proven that CNN-based models could outperform the traditional HMM systems once the systems are trained using a large dataset with a wide variety of writing styles [20]. For example, Sundaram et al. Use of CNNs for Recognition One of the initial attempts to exploit CNNs for recognition of handwriting was presented by [21], where CNNs were used in conjunction with Long Short-Term Memory (LSTM) networks to recognize handwriting sequences, outperforming conventional methods. Yet, one of the significant difficulties with handwriting recognition is that writing is temporal. Handwriting involves. Sequential strokes where the current stroke depends on the previous one make it challenging for conventional models to accommodate changes in writing speed, stroke order, and pressure. To overcome this, Recurrent Neural Networks (RNNs), most notably LSTMs, have been proposed for learning the data dependencies embedded in handwriting [22]. This makes LSTM networks ideal for tasks where each stroke's context matters, like sequence prediction in handwriting recognition. Combined with CNNs, these networks have successfully been leveraged to create hybrid networks that can learn handwritten text's spatial and temporal features [23][24]. Handwriting recognition is comparatively welldeveloped, but signature verification is a crucial and challenging problem. The authenticity of signatures is essential for banking, legal documents, and forensics. At first, signature verification systems concentrated on global matching techniques by analyzing the entire signature's shape and structure [25]. These approaches, however, were susceptible to variations in signature size and quality, which made them less robust at times. The latest methods employ CNNs to obtain unique features from signatures to improve performance for detecting forgeries after being distorted [26][27]. Apart from recognition and identification, handwritten forgery detection through behavior profiling has also received a lot of attention in the study over the past few years. Handwriting harbors valuable traces of the psychological and emotional state of the writer, including stress, cognitive load, and even personality traits [28]. For a long time, graphology has been used within forensic psychology to analyze written documents for behavioral signs, but the scientific basis has been debated. Recently, studies have tried to merge behavioral analysis with deep learning methods to derive features, such as pen density, stroke flow, and writing velocity [29][30], and subsequently infer psychological properties based on those features from handwritten dynamics. These findings indicate that dynamic handwriting attributes can reflect the writer's mental and emotional state during writing. However, behavioral profiling for handwriting recognition systems is still early. Current systems consider handwriting recognition and behavioral profiling as independent tasks and hardly provide a uniform framework for verifying writing habits. Some prominent exceptions are the work by Jiang et al. [31], which combined CNN and LSTM for writer identification and signature verification along with dynamic handwriting features, such as pressure and stroke order. The proposed framework showed that integrating static and dynamic features can improve handwriting recognition accuracy and behavioral profiling robustness. Current handwriting recognition systems suffer another limitation they fail to work in real-time environments, which are often essential in applications requiring real-time feedback, such as in e-signature verification or automated grading systems. Many systems have high computational costs or extended processing time when dealing with large-scale data [32]. Researchers have investigated techniques for low-latency processing, including optimizing model architecture and implementing parallel processing strategies to enable real-time handwriting recognition with minimal delay. Some methods have improved the latency by taking advantage of hardware acceleration with GPUs and TPUs to accelerate the inference of the model [33]. Although handwriting recognition and behavioral profiling have improved considerably, the systems available today still struggle with scalability, such as when segmenting multiple pieces of handwriting on the same page, identifying older, rarer forms of writing, or dealing with multilingualism [34]. The differences in writing styles, pressure while writing, and use of different languages from different temperaments make it difficult to build a universal handwriting recognition model. Additionally, many current systems lack sufficient flexibility to recognize different handwriting styles (e.g., cursive vs. printed) or quickly scale to larger datasets. We explore some of these gaps by incorporating advanced CV preprocessing processes with CNNs and LSTMs to increase handwriting recognition and enable real-time functionality. This approach Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4760 enhances recognition performance and presents behavioral profiling using dynamic handwriting features that offer cognitive insights into the writer's mental state. We, too, are built to combine all these diverse opinions our system has the added design of adaptability and permanence, ensuring scalability and applicability in real-world use cases from esignature verification to forensic document examination. The flow of the Computer Vision and machine learning models for handwriting recognition is illustrated in Figure 2. It starts with traditional methods, which struggled with complex handwriting recognition problems. The emergence of profound learning breakthroughs led to an interest in using hybrid models that take advantage of both. Those hybrid models utilized spatial features with CNNs (Convolutional Neural Networks) to perform feature extraction efficiently and temporal features with LSTMs (Long Short-Term Memory Networks) to catch the sequential dependencies in handwriting. These models address key challenges to improving signature verification, allow for behavioral profiling to understand writing dynamics, and enable real-time performance for practical use cases. However, the system has scalability challenges, such as efficiently adapting and processing large datasets and using diverse handwriting styles. By combining these methodologies, we can create a robust system for handwriting analysis. Figure 2: Integrating Computer Vision and Machine Learning for Handwriting Analysis Previous handwritten recognition approaches were rule-based solutions or also statistical models like HMMs. Instead, modern approaches used more complex ML models trained on large datasets containing diverse handwriting styles. The research design details include a comparative analysis of earlier studies on other domains (e.g., forensic handwriting, signature verification, and cognitive profiling). This work attempts to address the observed shortcomings of existing systems by presenting an Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4761 integrated system that combines CV-based feature extraction techniques with CNNs and LSTMs for handwritten text recognition and subsequent behavioural profiling. A new subsection was added that clearly delineated the key differences between our work and similar studies in literature. This addressed our hybrid approach, which combines computer vision and machine learning with behavioural profiling and real-time processing system capabilities. 3. PROPOSED METHODOLOGY We can conclude this segment by saying that it is all about a methodology that is related to signature verification, handwriting recognition, and behavioral profiling system verification, was designed and executed through a hybrid system of CV preprocessing, CNN, and Long Short-Term Memory networks to achieve the baseline described here. By introducing a new hybrid method, which combines dynamic stroke information with a new most-matched pattern classification algorithm, we significantly improve the accurate results in handwriting recognition and capture features of dynamic writing in a behavior-based profile. The system delivers real-time performance for esignature verification and document authentication applications. 3.1 Handwriting Data Acquisition and Preprocessing Our methodology commences with the acquisition of handwriting data. This can be scanned images of handwritten documents or real-time input from stylus devices. Due to noise, skew, and inconsistencies in handwritten images, preprocessing is essential to refine the input data and prepare it for further analysis. The preprocessing removes noise to eliminate errors like ink smears and noise in the background. Gaussian filtering techniques are applied to smooth out these noise contributors while keeping the handwriting structure intact. It is then binarized to separate the ink from the page background. This step is essential for proper feature extraction. After the binarization process, the correction process through Geometric transformation realigns the slanted handwritten text, leading to text-oriented in a regular fashion and facilitating feature extraction. Thereafter, the handwriting is divided into words, lines, and individual characters to accommodate for uneven spacing and to generate more precise input for the recognition systems. Finally, the image is resized and rectified to a standard size, ensuring consistency across different handwriting samples. 3.2 Feature Extraction with Convolutional Neural Networks (CNNs) After preprocessing the data, the next step is to represent features that retain the spatial structure of the handwritten text. The CNNs are helpful in this task because they help in learning the hierarchical patterns from image data (like edges, corners, and shapes of handwritten characters). In this study, the convolutional neural network (CNN) has several layers that extract more complex features. The convolutional layers used in the architecture can identify simple features such as line edges, strokes, and curves in the input image of the handwriting. This pooling layer reduces the feature maps' spatial dimensions, increasing computational efficiency without losing important information. The 2D feature maps are flattened after the feature extraction step into a 1D vector that can be passed into the following layers to perform the sequence modeling. 3.3 Sequence Modeling with Long Short-Term Memory Networks (LSTMs) Since handwriting is sequential, it is essential to model temporal dependencies between strokes, letters, and words. We utilize Long Short-Term Memory (LSTM) networks, a specialized version of Recurrent Neural Networks (RNNs), to accomplish this due to their strength in modeling sequential data. The problem, however, is that LSTMs can also be very effective in cases where the order of events (e.g., strokes) must be learned to correctly classify it (and you can see by handwriting that the order of strokes matters a lot). Our system implements a three-component LSTM network. Following this, the CNN extracts feature vectors from the image, which form the input layer and flow sequentially. Because LSTM cells remember information about the previous state in the sequence, they can help remember long-distance dependencies, such as letters or words written in a specific order. The final layer outputs the predictions for each character or word, which are generated based on what the model has learned from the sequence of strokes it saw previously. 3.4 Behavioral Profiling via Handwriting Dynamics Besides the static features, i.e. recognizing the written text, we also apply behavioral profiling by analyzing the dynamic characteristics of handwriting that reflect cognitive load. Our system considers writing features in addition to static features like writing speed, angulation of pen pressure, and order of writing strokes. These measures are a time to complete each stroke, and Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4762 faster writing is usually attributed to urgency or stress, while slower writing is attributed to fatigue or slow, thoughtful writing. It tracks the force used to put pen to paper, as fluctuations in pressure may indicate emotional dispositions such as anxiety or calmness. For example, the order of strokes and smooth or jerky strokes can indicate the writer's motor coordination and cognitive load. These features are dynamically captured by pressuresensitive stylus devices that provide real-time data concerning pen pressure, stroke velocity, and timing. These datasets are then passed to a multimodal deep learning model that fuses the spatial features provided by the CNN with the temporal features given by the LSTM. The fusion of recognitional and psychological models allows the system to read the handwriting and predict psychometric and emotional conditions from the features of the writing. 3.5 Real-Time Handwriting Recognition and Analysis The proposed system works in real-time, which makes it perfect for e-signature verification and live document authentication. Several optimizations help to reach low-latency processing. We use parallel processing, taking advantage of modern Hardware Accelerators like the Graphic Processing Units and the Tensor Processing unit so that both the CNN and the LSTM are present in parallel, thus saving time for inference. Instead of processing the entire dataset in batches, handwriting data is entered and processed one by one, applying the preprocessing steps dynamically to decrease latency. As such, this method allows the system to deliver instantaneous responses when deployed in practice, making it suitable for scenarios that demand rapid and accurate handwriting recognition. 3.6 Privacy Preservation and Security The system must meet privacy constraints because handwriting data might include sensitive information, particularly in areas like signature verification. We use privacy-preserving methods like federated learning to keep this handwriting data safe during this process. It also allows model training on distributed data without storing sensitive data centrally, which helps comply with privacy regulations like GDPR. This also ensures user data is kept private and not included in the analysis. Our hypothesis is that integrating computer visionguided preprocessing with deep learning models (CNNs and LSTMs) can significantly improve handwriting recognition accuracy, scalability, and real-time analysis while also enabling behavioral profiling. 3.7 Architecture of the System Figure 3: Architecture of the Handwriting System Analysis Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4763 In Figure 3, the architecture of handwriting analysis system consists of several stages: 1. Input Layer: The system first takes a handwritten document or signature as input, represented as an image 𝑋 ∈ 𝑅××, where ℎ is the height, 𝑤 is the width, and 𝑐 is the number of channels. 2. Preprocessing: The preprocessing module applies noise reduction (Gaussian filter 𝒢), binarization 𝑇, segmentation, and normalization to clean and standardize the image. Mathematically: 𝑋filtered = 𝑋 ∗ 𝒢 (1) 𝑋binary = Binarize(𝑋filtered,𝑇) (2) 𝑋normalized =binarybinary binary binary (3) 3. CNN Feature Extraction: The CNN applies filters to extract spatial features from the preprocessed image. The convolution operation is given by: 𝑌 = σ(𝑊 ∗𝑋normalized +𝑏) (4) where 𝑊 is the filter, 𝑏 is the bias, and 𝜎 is the activation function. 4. LSTM Sequential Modeling: LSTMs analyze the sequential dependencies in handwriting. The LSTM equations are: 𝑖= σ(𝑊𝑥+𝑈ℎ +𝑏) (5) 𝑓= σ𝑊𝑥+𝑈ℎ +𝑏 (6) 𝑐= 𝑓⋅𝑐 +𝑖⋅tanh(𝑊𝑥+ 𝑈ℎ +𝑏) (7) ℎ= 𝑜⋅tanh(𝑐) (8) where 𝑖, 𝑓, and 𝑜 are input, forget, and output gates, respectively, and 𝑐 is the memory cell at time step 𝑡. 5. Behavioral Profiling: Writing dynamics such as speed 𝑣 and pressure 𝑝 are analyzed for psychological profiling. Speed is calculated as: 𝑣 =   (9) where 𝑑 is the distance between two strokes and Δ𝑡 is the time difference. 6. Output Layer: The system outputs recognized text 𝑇, signature verification score 𝑆verify, writer identification probability 𝑝(𝑤|𝑋), and psychological profiling. The signature similarity is calculated using cosine similarity: Similarity𝑆input,𝑆ref = input⋅ref |input||ref| (10) Writer identification uses: 𝑝(𝑤|𝑋)=, ∑󰇡,󰇢   (11) where 𝑓(𝑤,𝑋)measures similarity between writer 𝑤 and handwriting 𝑋. 7. Real-Time Processing: The system is optimized for low-latency processing. The total processing time is: 𝑇total = 𝑇preprocessing +𝑇CNN +𝑇LSTM + 𝑇output (12) ensuring that feedback is provided in realtime. 4. RESULTS The experimental results for the proposed handwriting recognition system in this article using computer vision-based preprocessing for text line segmentation in subsections and stages, Convolutional Neural Networks (CNNs) in feature extraction, and Long Short-Term Memory (LSTM) networks for modeling of sequences are presented in this section. Compared with state-of-the-art solutions, we further analyze performance in applications such as handwriting recognition, signature-based biometric verification, and behavioral profiling. 4.1 Handwriting Recognition Performance The IAM Handwriting Database was used to evaluate the proposed system, which is characterized by various handwriting styles associated with different writers. It is recognized accurately with 95.4% precision, outperforming traditional HMM-based and acoustic model-based systems, which usually work around 90–92% accuracy. As shown here, The CNN-LSTM hybrid architecture effectively captures the spatial and temporal characteristics of handwriting, resulting in improved accuracy in recognizing complex writing styles and sequences. Along with accuracy, the system's precision and recall were assessed. The precision of 94.8% means that when the system predicts a character/word, it is most likely correct; the number of false positives is small. This means that with a recall rate of 95.1%, this method successfully retrieves most of the relevant words or characters, leading to fewer false negatives. Overall, these results demonstrate that our model balances precision and recall, which is essential for reliable performance in practice. Moreover, an F1 score of 94.9% further emphasizes the system's strength in tackling different types of Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4764 handwriting, including cursives and printed text. The performance enhancement can be directly tied to the implementation of CNNs for feature extraction, which makes it possible to learn essential handwriting patterns and characters from raw image inputs combined with LSTMs that spill over the sequential dependencies among the strokes. 4.2 Signature Verification Performance The system was further tested using the GPDS-960 Signature Dataset, which consists of genuine signatures and forgeries from 960 users. This signature verification task is essential in fields like document verification, banking, etc. Our proposed system reached an accuracy of 97.8% on the signature classification task, surpassing many other existing models, primarily those that integrated only the static signature shape similarity. The precision for signature verification was 98.1%, meaning that whenever the model predicts a signature is genuine, it is almost always right. With a recall rate of 97.5%, the numbers show that the system can spot nearly every genuine signature while not incorrectly identifying forgeries as accurate. The F1 score for this task is 97.8%, meaning the system can recognize genuine and forged signatures consistently. The Area Under the Curve (AUC) value of 0.991 also signifies the model's online application effectiveness of authentic and forged signature identification, which outperformed the traditional form signature verification model AUC by around 0.95. These results further validate our hybrid CNNLSTM approach's signature verification effectiveness, mainly when signatures are distorted, vary in size, or have pressure changes. Moreover, the behavioral profiling module adds to the system's overall strength by examining the variability of dynamic features (e.g., stroke velocity, stroke pressure) that are often neglected by conventional approaches. 4.3 Behavioral Profiling and Cognitive Trait Estimation Apart from handwriting recognition and signature verification, our system has a behavioral profiling module that studies the dynamics of writing to derive the writer's cognitive status. When in action, the part of the system that actionably presents behavioral profiles estimates a person's stress, cognitive load, emotional state, etc. Based on their writing speed, pressure, and stroke order, APPROXIMATE with 5% deviation personality profiles 88.3% of the time. This performance demonstrates that the system can consistently assess writing behavior, providing valuable insight into the writer's psychological state. The speed at which you write may signal stress or anxiety, with increased writing pressure correlating with those feelings, while lower writing speeds signal fatigue or careful consideration, for example. The model's ability to capture these dynamic aspects indicates its potential for use in forensic psychology, mental health assessments, and other scenarios where a writer's psychological state is fascinating. The accuracy in identifying the writer was 91.7%, which showed that the system can differentiate between writers by analyzing features unique to handwriting, including order of strokes, pressure, and rhythm. These findings indicate that our framework applies in settings such as author identification in history document analysis or forensic investigation, where there is a need to identify whether a script sample comes from a suspected author. 4.4 Real-Time Performance Our real-time processing is crucial for e-signature verification and live document authentication. The real-time experiments showed that the processing speed of each character could be as low as 45 ms on average. The low latency is essential for real-time applications, where information feedback must be fast. In e-signature verification, for instance, users want to know whether the signature is authentic in real-time. By implementing parallel processing techniques and optimized deep learning models, our system can process large volumes of data with low latency while retaining the detailed structural information necessary for accurate analysis. 4.5 Comparison with State-of-the-Art Solutions We then compared our system's results to those of other state-of-the-art handwriting recognition and signature verification systems to analyze its performance. The performance evaluation of the proposed system, as shown in Table 1, demonstrates its superiority in signature verification and processing time in real-time comparisons against several well-known models. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4765 Table 1: Performance Comparison with Existing Systems Model/Appro ach IAM Databa se Accura cy GPDS Signatu re Accura cy AUC (Signatur e Verificati on) Proposed System 95.4% 97.8% 0.991 CTC + RNN 93.2% 94.6% 0.95 CNN + LSTM 94.1% 95.5% 0.96 CNN for Signature N/A 96.1% 0.98 Table 1 indicates that the proposed system provides better handwriting recognition accuracy and signature verification rates than other systems. Moreover, it obtains 87% AUC for signature verification, a measure that indicates forgery detection difficulty, which is substantially higher than others. Figure :4 Handwriting Recognition Accuracy Comparison Distributed Generations based on vehicle-to-grid technology (V2G): Car parking is an extensive endeavor, but it is also a traditional way of hoarding energy. Figure 4 depicts the handwriting recognition results using four models: the proposed system, CTC + RNN, CNN + LSTM, and CNN for Signature. The proposed system achieved the highest accuracy, 95.4%, outperforming all other models. CTC + RNN obtained 93.2%, the lowest, while CNN + LSTM and CNN for Signature reached 94.1% and 94.5%, respectively. This shows the proposed system's performance compared to existing approaches that deal with different styles of handwriting recognition. Figure :5 Signature Verification Accuracy Comparison The comparison of accuracy for signature verification among the four models is shown in Figure 5. The proposed system has also outperformed the other models in the results, with an accuracy of 97.8%. The results show the CTC + RNN model achieves 94.6 %accuracy, the CNN + LSTM model gives a slightly better result of (95.5), and the CNN for Signature gives the best result of 96.1% accuracy. The evaluation results demonstrate that the proposed system outperforms existing systems regarding signature verification, making it a promising solution for real-world applications such as document authentication and banking. Figure :6 AUC for Signature Verification Comparison Figure 6 depicts the model's performance in discriminating between actual and forged signatures. The system proposed achieves the best value AUC of 0.991, which means excellent performance in signature verification. The CTC + RNN, CNN + LSTM, and CNN for Signature models achieved lower AUC values of 0.95, 0.96, and 0.98, respectively. A greater AUC suggests a superior classification function, and this figure highlights the proposed system's competency to detect valid signatures with a limited risk of forgery successfully. In addition, we showed how our improvement had outperformed the current state-of-the-art systems in