Research Articles

Sandhya G· Vol 1, No 1

This paper proposes an advanced handwritten document processing system, which is designed to enhance the handwritten text recognition by jointly using a Convolutional Neural Networks and Auto-Encoding Transformer models with BERT-based spelling correction and GECToR-based grammatical correction, and is further compared to an existing OCR-based approach for handwritten recognition. Materials and Methods: The results of this study consist of two groups: Group 1 consists of an experimental group that used the already- existing OCR system, which was tested against ten test samples, averaging about 78% accuracy in recognition. Group 2 presents the proposed CNN + Auto-Encoding Transformer model integrated with BERT and GECToR. Testing criteria included accuracy, error rate, and processing time. Sample size in both groups was determined through prior studies. It was expected to have a test power of 80%, with a significance level of 0.05 and a 95% confidence interval. Statistical analysis was done using SPSS 26.0 software; for comparing two independent samples, the appropriate statistical tool is the independent samples t-test. Result: The results show clearly that the proposed approach of utilizing the CNN Auto Encoding Transformer performs much better than a traditional approach of utilizing an Optical Character Recognition (OCR) system. The proposed approach achieves high accuracy of 93.6%, reduces errors to 0.11, and also speeds up recognizing handwritten text in less than 0.58 seconds while maintaining statistically significant improvements to confidence level of (p < 0.001). Conclusion: Overall, the combined effect of the CNN + Auto-Encoding Transformer with the use of BERT and GECToR makes the recognition of the handwritten text much more precise and even easier to read. The system is faster compared to the traditional OCR methods.

Jeevanantham M· Vol 1, No 1

The objective of this study is to design a CNN-based multimodal emotion recognition system using speech and ECG signals to improve recognition accuracy. Group 1 represents the existing speech-based emotion recognition system using MFCC features with a CNN model. Group 2 represents the proposed multimodal system combining MFCC speech features and ECG signals using CNN-based fusion. The evaluation of the system was done using Accuracy, Precision, Recall, and F1-score measures. The multimodal system's accuracy was higher, i.e., 92.8%, than the existing system, which is based on speech alone, with an accuracy of 78.5%. The results proved that the system using speech and ECG is more accurate and reliable, making it appropriate for emotion recognition in real-time applications.

KaranSingh D· Vol 1, No 1

In this study, an AI-driven investment recommendation system is designed using a Retrieval-Augmented Generation (RAG) approach for generating context aware, interpretable information. The personalized investment recommendations compared with a DRL-based approach designed in a previous study. The study proposes two models of the three. Group 1 comprises an existing DRL-based investment recommendation system that learns market decisions using reward-based market learning. Group 2 comprises the proposed RAG-based system which combines financial knowledge retrieval (market reports, news). And market history) with a transformer-based language model to generate personalized investment recommendations. The models are tested using Accuracy, Recall, Precision, F1-Score, Return Efficiency, Processing Time and Interpretability Score. The statistical tests are performed with SPSS 26. 0 and a one-sample t-test that considers the significance at 0. 05, with a 95% confidence interval. The obtained result reveals that the proposed RAG based investment recommendation system significantly outperforms the DRL-based system with attainment of 88. 63% accuracy, 87. 40% recall, 87. 40% return efficiency, processing time of 312. 50 ms and an interpretability score of 79. 85% statistically significant, p<0.01. The result suggests that the RAG-based system has higher accuracy, improved returns and improved interpretability than the DRL-based System and thus is a more effective and efficient solution for AI-based investment advisory systems.

Hariesh R· Vol 1, No 1

This work proposes a real-time object detection model that integrates the efficiency of YOLOv12 with an attention-based Transformer to address the limitations of traditional detection systems in dynamic surveillance environments. Classical models often struggle under challenging conditions such as low light, occlusion, or rapid movement. In contrast, the proposed hybrid model leverages YOLOv12’s robust feature extraction with the attention mechanism’s ability to emphasize salient spatial and contextual features. Trained on a comprehensive real-time surveillance dataset and benchmarked against a baseline YOLOv8 model with attention enhancements, the proposed system demonstrated significant improvements. It achieved a 96% detection rate, reduced average processing time from 35 ms to 20 ms, and lowered the error rate to 10%, with results statistically significant (p = 0.015). Key performance metrics—precision, recall, and F1-score—confirmed the model’s capability to accurately detect and classify multiple objects in varying environmental conditions. The attention-based Transformer further enabled dynamic focus on critical regions, enhancing localization and classification. This study underscores the potential of combining deep convolutional and attention architectures to create cost-effective, accurate, and scalable solutions for smart surveillance, paving the way for advancements in adaptive detection and AI-driven security systems.

Jeeva M· Vol 1, No 1

This research aims to build a hybrid phishing URL detection model and classify the URLs as either Legitimate or Phishing. This model is compared with other baseline models such as FNN, DNN, Wide&Deep, and TabNet. Materials and Methods: Group 1 includes base models like FNN, DNN, Wide&Deep, and TabNet that were trained on different dataset and URL features. Group 2 uses a hybrid of machine learning models like SVM, RF and DT , and is trained with 11000+ URLs. Both existing models and proposed model are compared against metric values of accuracy, recall, false positive rate, and the anti-phishing score. Result: The proposed hybrid model performed well and got an accuracy of 96.06%, a recall of 96.59%, false positive rate of 0.0446 and an anti-phishing score of 0.9158. All the results were good but there is a minor variation in APS score, this is due to the composite nature of the weighted index rather than decrease in performance efficiency. Conclusion: The results show that the proposed hybrid model is more accurate and stable in detecting the phishing URLs.