Department of Computer Science and Engineering

Deep learning techniques have demonstrated early advancements in addressing the challenges of complex Visual Speech Recognition (VSR) tasks. Nonetheless, a persistent issue arises when distinguishing characters or words with similar pronunciations, known as homophones, which results in ambiguity. Existing VSR systems also face technical constraints due to insufficient visual data for learning short-duration phonemes like “at”, “an”, “a”, and “eight”. Moreover, cutting-edge VSR techniques perform exceptionally well when interpreting overlapping speakers. However, extending these methods to unseen speakers leads to a significant performance decline due to the limited diversity in the training dataset and substantial variations in physical attributes, such as lip shape and color, across different speakers. To address the existing challenges in VSR, we propose a multi-modal approach that leverages visual and landmark information to capture complex spatio-temporal patterns for the model generalization capabilities. The model employs a multi-layered Three-Dimensional Convolutional Neural Network (3D-CNN) that extracts visual features, while a Graph Convolutional Network (GCN) captures precise landmark information for accurate lip shape localization. The extracted features are then fused for further processing using a Sequence-to-Sequence (Seq2Seq) model based on the attention mechanism. The proposed model achieved a WER of 0.53% and 8.21% for the overlap and unseen speakers category. Notably, these results surpass the performance of existing models, demonstrating remarkable accuracy for VSR on the GRID dataset in both the unseen and overlapping speaker scenarios.

Facial biometric systems are extensively applied in diverse sectors for the purposes of person authentication and verification, primarily due to the distinctive nature of individual facial characteristics. Deep learning models are typically used in face authentication to validate people with excellent recognition accuracy. However, these systems are susceptible to a variety of cyber attacks that manipulate the digital representations of real-world faces to cheat the models. In the contemporary landscape of digital identity theft, liveness detection stands as a crucial technology. The need for enhanced security prompts the demand for a resilient system that can effectively counter face spoofing attempts and prevent unauthorized access. A Binary Authentication Protocol (BAP) technique is proposed to enhance facial biometric security in combination with Visual Speech Recognition (VSR). In the proposed method, the first verification step entails face authentication. Further, the authentication protocol involves a challenge-response-based method using VSR. The proposed method achieved a word error rate of 2.7% and a word recognition rate of 97.3%, surpassing existing state-of-the art methods in VSR. The proposed scheme offers practical and effective solutions to prevent face spoofing through active liveness detection in face-based authentication systems.

Identity verification is essential in both an individual’s personal and professional life. It confirms a person’s identity for various services and establishes their legitimacy as an employee within an organization. As cybercrime evolves and becomes more sophisticated, ensuring robust, and secure personal authentication methods has become a critical challenge. Existing face-based authentication systems typically employ deep learning models for user verification. However, these systems are susceptible to various attacks, such as presentation attacks, 3D mask attacks, and adversarial attacks that exploit and deceive the models by manipulating digital representations of human faces. Although various liveness detection techniques have been proposed to combat face spoofing in face-based authentication systems. However, these systems remain vulnerable and can be exploited by sophisticated techniques. To counteract face spoofing in a face-based authentication system, we have proposed an advanced liveness detection technique using Visual Speech Recognition (VSR). The proposed VSR model is designed to integrate seamlessly with face-based authentication systems, forming a dual authentication framework for enhanced liveness detection. The VSR model decodes silently pronounced speech from video by analyzing unique, unforgeable lip motion patterns into textual representation. Although, various liveness detection techniques have been proposed to combat face spoofing in face-based authentication systems. However, these systems remain vulnerable and can be exploited by sophisticated techniques. To counteract face spoofing in a face-based authentication system, we have proposed an advanced liveness detection technique using VSR. The proposed VSR model is designed to integrate seamlessly with face-based authentication systems, forming a dual authentication framework for enhanced liveness detection. The VSR model decodes silently pronounced speech from video by analyzing unique, unforgeable lip motion patterns into textual representation. To achieve effective liveness detection using VSR, we need to enhance the accuracy of the VSR system. The proposed work employs an encoder-decoder technique to extract more robust features from lip motion. The encoder employs a three-dimensional convolution neural network (3D-CNN) combined with a fusion of bi-directional gated recurrent units and long short-term memory (BiGRU-BiLSTM) to effectively capture spatial-temporal patterns from lip movement. The decoder integrates Multi-Head Attention (MHA) with BiGRU-BiLSTM to effectively focus on relevant features and enhance contextual understanding for more accurate text prediction. The proposed VSR system achieved a word error rate (WER) of 0.79%, demonstrating a significant reduction in error rate and outperforming compared to the existing VSR models.

As recommender systems shift from rating-based to interaction-based models, graph neural network-based collaborative filtering models are gaining popularity due to their powerful representation of user-item interactions. However, these models may not produce good item ranking since they focus on explicit preference predictions. Further, these models do not consider side information since they only capture latent feature information of user-item interactions. This study proposes an approach to overcome these two issues by employing preference relation in the graph neural network model for collaborative filtering. Using preference relation ensures the model will generate a good ranking of items. The item side information is integrated into the model through a trainable matrix, which is crucial when the data is highly sparse. The main advantage of this approach is that the model can be generalized to any recommendation scenario where a graph neural network is used for collaborative filtering. Experimental results obtained using the recent RS datasets show that the proposed model outperformed the related baselines. © 2024 Elsevier Ltd

Infectious viruses are spread during human-to-human contact and can cause worldwide pandemics. We have witnessed worldwide disasters during the COVID-19 pandemic because of infectious viruses, and these incidents often unfold in various phases and waves. During this pandemic, so many deaths have occurred worldwide that they cannot even be counted accurately. The biggest issue that comes to the forefront is that health workers going to treat patients suffering from COVID-19 also may get infected. Many health workers have lost their lives to COVID-19 and are still losing their lives. The situation can worsen further by coinciding with other natural disasters like cyclones, earthquakes, and tsunamis. In these situations, an intelligent automated model is needed to provide contactless medical services such as ambulance facilities and primary health tests. In this paper, we explore these types of services safely with the help of an intelligent automated transportation model using a vehicular delay-tolerant network. To solve the scenario, we propose an intelligent transportation system for automated medical services to prevent healthcare workers from becoming infected during testing and collecting health data by collaborating with a delay-tolerant network of vehicles in intelligent transport systems. The proposed model automatically categorizes and filters infected patients, providing medical facilities based on their illnesses. Our mathematical evaluation and simulation results affirm the effectiveness and feasibility of the proposed model, highlighting its strength compared to existing state-of-the-art protocols. © 2024 Elsevier B.V.

In analytical modeling for biochemical pathways, precisely determining unknown parameters is paramount. Traditional methods, reliant on experimental time course data, often encounter roadblocks — limited accessibility and variable quality — that can significantly impact the algorithm's performance. In this study, we address these hurdles by unveiling a groundbreaking parameter estimation technique, Constrained Regularized Fuzzy Inferred Extended Kalman Filter (CRFIEKF). This innovative approach eliminates the need for experimental time-course measurements and capitalizes on the existing imprecise relationships among the molecules within the network. Our proposed framework integrates a Fuzzy Inference System (FIS) block to encapsulate these approximated relationships. To fine-tune the estimated parameter values, we employ Tikhonov regularization. The selection of Tikhonov regularization and Gaussian membership functions was based on the Mean Squared Error (MSE) values observed during the parameter estimation process, contrasting our results with those of previous studies. We rigorously tested the proposed approach across various pathways, from the glycolytic processes in mammalian erythrocytes and yeast cells to the intricate JAK/STAT and Ras signaling pathways. The results were impressive, showing a significant similarity (p-value < 0.001) to the outcomes of specific prior experiments. The dynamics of the biochemical networks normalized within the [0, 1] range mirrored the transient behavior (MSE < 0.5) of both in vivo and in silico results from previous studies. In conclusion, our findings highlight the effectiveness of CRFIEKF in estimating the kinetic parameter values without prior knowledge of experimental data within a biochemical pathway in the state-space model. The proposed method underscores its potential as a game-changer in biochemical pathway analysis. © 2024 Elsevier Ltd

We report the green synthesis of novel magnetically separable MgFe2O4 nanoparticles using Cajanus cajan (L.) Millsp leafs via combustion method. The MgFe2O4 were characterized by powder X-ray diffraction (XRD), Fourier transform infrared spectroscopy (FTIR), scanning electron microscopy (SEM), transmission electron microscopy (TEM), vibrating sample magnetometer (VSM), and UV-diffuse reflectance (UV-DRS) spectroscopy. The crystalline structure of MgFe2O4 was confirmed via XRD studies and TEM showed that the MgFe2O4 NPs were distorted spherical particles with particle size ranging between 5 and 15 nm. UV-DRS study showed the optical band gap of MgFe2O4 NPs to be 1.8 eV. Microwave-assisted (MW) degradation of PCM-dolo drug using MgFe2O4 as catalyst was performed at different operating parameters such as time (30 min), drug concentration (PCM-dolo 50 mg/L), initial concentration of MgFe2O4 (0–110 mg/L), and microwave power (100–600 W) to obtained the degraded fragments of the drug. Experimental data was used to compute the degradation efficiency of PCM-dolo on MgFe2O4. The enhanced catalytic performance could be ascribed to the production of MW-induced active species, such as holes (h+), superoxide radicals (?O2?) and hydroxyl radicals (?OH) in the degradation process. A possible degradation mechanism and pathway was proposed. Graphical abstract: (Figure presented.) © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024.

The ability to predict diseases is critical in healthcare for early intervention and better patient outcomes. Data security and privacy significantly classified medical data from several institutions is analyzed. Cooperative model training provided by Federated Learning (FL), preserves data privacy. In this study, we offer a fusion strategy for illness prediction, combining FL with Anova and Chi-Square Feature Selection (FS) and Linear Discriminate Analysis (LDA) Feature Extraction (FE) techniques. This research aims to use FS and FE techniques to improve prediction performance while using the beneficial aspects of FL. A comprehensive analysis of the distributed data is ensured by updating aggregate models with information from all participating institutions. Through collaboration, a robust disease prediction system excels in the limited possibilities of individual datasets. We assessed the fusion strategy on the Cleveland heart disease and diabetes datasets from the UCI repository. Comparing the fusion strategy to solo FL or conventional ML techniques, the prediction performance a unique fusion methodology for disease prediction. Our proposed models, Chi-Square with LDA and Anova with LDA leveraging FL, exhibited exceptional performance on the diabetes dataset, achieving identical accuracy, precision, recall, and f1-score of 92.3%, 94.36%, 94.36, and 94.36%, respectively. Similarly, on the Cleveland heart disease dataset, these models demonstrated significant performance, achieving accuracy, precision, recall, and f1-score of 88.52%, 87.87%, 90.62, and 89.23%, respectively. The results have the potential to revolutionize disease prediction, maintain privacy, advance healthcare, and outperform state-of-the-art models

Wireless Sensor Networks (WSNs) have become pivotal in numerous applications, including environmental monitoring, precision agriculture, and disaster response. In the context of urban flood monitoring, utilizing unmanned aerial vehicles (UAVs) presents unique challenges due to the dynamic and unpredictable nature of the environment. The primary challenges involve designing strategies that maximize data collection while minimizing the Age of Information (AoI) to ensure timely and accurate decision-making. Efficient data collection is crucial to capturing all relevant information and providing a comprehensive understanding of flood dynamics. Simultaneously, reducing AoI is essential, as outdated data can lead to delayed or incorrect responses, potentially worsening the situation. Addressing these challenges is critical for the effective use of WSNs in urban flood monitoring. Initially, we formulate the problem as a mixed integer non-linear programming (MINLP) problem. Further, it is solved using a Lagrangian-based branch and bound technique by converting it into an unconstrained problem. Then, for large-scale WSN, we propose a hybrid optimization technique which combines a genetic algorithm with a particle swarm optimization technique to simultaneously maximize the data collection and reduce the AoI of the collected data with the constraint of energy consumption of the UAVs. Simulation results demonstrate that our proposed algorithm outperforms existing approaches in terms of both data collection and AoI.

Accidental falls and their consequences are critical concerns for elderly people. Fatal injuries, when delayed in treatment, can lead to severe outcomes. Fall detection systems are crucial for the timely treatment of such injuries. Although sensor-based fall detection approaches are effective, video-based approaches are more useful because they assist in analyzing the fall scene and identifying the cause of the fall. However, privacy preservation is a major concern in video-based fall detection. The proposed system introduces a privacy-preserving mechanism that masks the identified human with a silhouette. A custom dataset, including 80 activities of daily living and 70 fall activities, is introduced. An LSTM variational autoencoder architecture is designed with a gradient clipping mechanism and a smooth variant of Adaptive Moment Estimation with Stochastic Gradient Descent (AMSGrad) optimizer to enhance the accuracy of fall detection. The reconstruction error between normal and fall activities is clearly identified with the help of a dynamic threshold. This results in a system performance that achieves accuracy, precision, and sensitivity of 99%, 97%, and 99%, respectively

With the rapid growth of Internet technology and social networks, the generation of text-based information on the web is increased. To ease the Natural Language Processing (NLP) tasks, analyzing the sentiments behind the provided input text is highly important. To effectively analyze the polarities of sentiments (positive, negative and neutral), categorizing the aspects in the text is an essential task. Several existing studies have attempted to accurately classify aspects based on sentiments in text inputs. However, the existing methods attained limited performance because of reduced aspect coverage, inefficiency in handling ambiguous language, inappropriate feature extraction, lack of contextual understanding and overfitting issues. Thus, the proposed study intends to develop an effective word embedding scheme with a novel hybrid deep learning technique for performing aspect-based sentimental analysis in a social media text. Initially, the collected raw input text data are pre-processed to reduce the undesirable data by initiating tokenization, stemming, lemmatization, duplicate removal, stop words removal, empty sets removal and empty rows removal. The required information from the pre-processed text is extracted using three varied word-level embedding methods: Scored-Lexicon based Word2Vec, Glove modelling and Extended Bidirectional Encoder Representation from Transformers (E-BERT). After extracting sufficient features, the aspects are analyzed, and the exact sentimental polarities are classified through a novel Positional-Attention-based Bidirectional Deep Stacked AutoEncoder (PA_BiDSAE) model. In this proposed classification, the BiLSTM network is hybridized with a deep stacked autoencoder (DSAE) model to categorize sentiment. The experimental analysis is done by using Python software, and the proposed model is simulated with three publicly available datasets: SemEval Challenge 2014 (Restaurant), SemEval Challenge 2014 (Laptop) and SemEval Challenge 2015 (Restaurant). The performance analysis proves that the proposed hybrid deep learning model obtains improved classification performance in accuracy, precision, recall, specificity, F1 score and kappa measure.

The ring oscillator (RO) is widely used to address different hardware security issues. For example, the RO-based physical unclonable function (PUF) generates a secure and reliable key for the cryptographic application, and the RO-based aging sensor is used for the efficient detection of recycled ICs. In this paper, a CMOS inverter with two voltage control signals is used to design a configurable RO (CRO). With its control signals, the proposed CRO can both accelerate and lower the impact of aging on the oscillation frequency. This vital feature of the proposed CRO makes it suitable for use in PUFs and RO-based sensors. The performance of both the proposed modified architecture, i.e., CRO PUF and CRO sensor, is evaluated in 90 nm CMOS technology. The aging tolerant feature of the proposed CRO enhances the reliability of CRO PUF. Similarly, the aging acceleration property of CRO improves the rate of detection of recycled ICs. Finally, both the proposed architectures are area and power-efficient compared to standard architectures

Globally, cancer stands as the second leading cause of mortality. Various strategies have been proposed to address this issue, with a strong emphasis on utilizing gene expression data to enhance cancer detection methods. However, challenges arise due to the high dimensionality, limited sample size relative to its dimensions, and the inherent redundancy and noise in many genes. Consequently, it is advisable to employ a subset of genes rather than the entire set for classifying gene expression data. This research introduces a model that incorporates Ranked-based Filter (RF) techniques for extracting significant features and employs Extreme Learning Machine (ELM) for data classification. The computational cost of using RF technique over high dimensional data is low. However extraction of significant genes using one or two stage of reduction is not effective. Thus, a 4-stage feature reduction strategy is applied. The reduced data is then utilized for classification using few variants of ELM model and activation function. Subsequently, a two-stage grading approach is implemented to determine the most suitable classifier for data classification. This analysis is conducted over four microarray gene expression data using four activation function with seven learning based classifiers, from which it is shown that II-ELM classifier outperforms in terms of performance matrix and ROC graph

In recent years, the healthcare data system has expanded rapidly, allowing for the identification of important health trends and facilitating targeted preventative care. Heart disease remains a leading cause of death in developed countries, often leading to consequential outcomes such as dementia, which can be mitigated through early detection and treatment of cardiovascular issues. Continued research into preventing strokes and heart attacks is crucial. Utilizing the wealth of healthcare data related to cardiac ailments, a two-stage medical data classification and prediction model is proposed in this study. Initially, Binary Grey Wolf Optimization (BGWO) is used to cluster features, with the grouped information then utilized as input for the prediction model. An innovative 6-layered deep convolutional neural network (6LDCNNet) is designed for the classification of cardiac conditions. Hyper-parameter tuning for 6LDCNNet is achieved through an improved optimization method. The resulting model demonstrates promising performance on both the Cleveland dataset, achieving a convergence of 96% for assessing severity, and the echocardiography imaging dataset, with an impressive 98% convergence. This approach has the potential to aid physicians in diagnosing the severity of cardiac diseases, facilitating early interventions that can significantly reduce mortality associated with cardiovascular conditions

The centralized nature of software?defined networks (SDN) makes them a suitable choice for vehicular networks. This enables numerous vehicles to communicate within an SD?vehicular network (SDVN) through vehicle?to?vehicle (V2V) and with road?side units (RSUs) via vehicle?to?infrastructure (V2I) connections. The increased traffic volume necessitates robust security solutions, particularly for Sybil attacks. Here, the attacker aims to undermine network trust by gaining unauthorized access or manipulating network communication. While traditional cryptography?based security methods are effective, their encryption and decryption processes may cause excess delays in vehicular scenarios. Previous studies have suggested machine learning (ML) like AI?driven approaches for Sybil attack detection in vehicular networks. However, the primary drawbacks are high detection time and feature engineering of network data. To overcome these issues, we propose a two?phase detection framework, in which the first phase utilizes cosine similarity and weighting factors to identify attack misbehavior in vehicles. These metrics contribute to the calculation of effective node trust (ENT), which helps in further attack detection. In the second phase, deep learning (DL) models such as CNN and LSTM are employed for further granular classification of misbehaving vehicles into normal, fault, or Sybil attack vehicles. Due to the time series nature of vehicle data, CNN and LSTM are used. The methodology deployed at the controller provides a comprehensive analysis, offering a single? to multi?stage classification scheme. The classifier identifies six distinct vehicle types associated with such attacks. The proposed schemes demonstrate superior accuracy with an average of 94.49% to 99.94%, surpassing the performance of existing methods

The study examines how well the routing system based on logic performs using both simulations and real life situations. It shows that the system is effective in improving emergency patient transportation and reducing response times. By using membership functions and fuzzy inference the system can. Direct patients to the suitable healthcare facility. It handles imprecise and unclear inputs better through variables and fuzzy rules resulting in accurate and responsive routing decisions.By incorporating logic this approach takes into account the uncertainties and complexities in emergency scenarios, such as location, medical condition severity and real time traffic conditions.Routing emergency patients is crucial for optimizing healthcare services to ensure efficient care delivery. The study assesses how well the routing system based on logic performs through simulations and real world situations demonstrating its effectiveness, in optimizing emergency patient transportation while minimizing response times.The Patient Routing Application is an innovative solution designed to streamline the process of providing emergency medical assistance to individuals based on their health conditions and geographical location. Leveraging a combination of fuzzy logic, geospatial data, and real-time mapping, the system evaluates a patient’s vital signs and recommends the nearest and most accessible hospitals.The software uses logic to evaluate how critical a patients condition is,Take into account factors, like heart rate, blood pressure and temparature. The fuzzy logic system utilizes predefined rules to categorize the severity and prescribe appropriate medical responses. Patient data, including personal information and medical history, is collected through a user-friendly graphical interface

Hyperspectral unmixing (HU) is crucial for extracting material information from hyperspectral images (HSI) obtained through remote sensing. Although linear unmixing methods are widely used due to their simplicity, they only address linear mixing effects. Nonlinear mixing models, while more complex, often focus solely on the nonlinear aspects affecting individual pixels. However, in practice, light reflected from materials within a pixel experiences linear and nonlinear interactions, necessitating a hybrid mixing model (HMM) that leverages spatial and spectral information. This work proposes a novel deep learning-based autoencoder (AE) with dual-stream decoders to enhance spectral unmixing. Our approach employs multitask learning (MTL) to process spatial and spectral information concurrently. Specifically, one decoder stream performs linear unmixing from HSI patches, while the other stream utilizes fully connected layers to capture and model the nonlinear interactions within the data. By integrating linear and nonlinear information, our method improves the accuracy of unmixing the mixed spectrum. We validate the effectiveness of our architecture on three real-world HSI datasets and compare its performance against various baseline methods. Experimental results consistently demonstrate that our approach outperforms existing methods, as evidenced by superior spectral angle distance (SAD) and mean squared error (MSE) metrics

This article presents a new model, ALL-Net, for the detection of acute lymphoblastic leukemia (ALL) using a custom convolutional neural network (CNN) architecture and explainable Artificial Intelligence (XAI). A dataset consisting of 3,256 peripheral blood smear (PBS) images belonging to four classes—benign (hematogones), and the other three Early B, Pre-B, and Pro-B, which are subtypes of ALL, are utilized for training and evaluation. The ALL-Net CNN is initially designed and trained on the PBS image dataset, achieving an impressive test accuracy of 97.85%. However, data augmentation techniques are applied to augment the benign class and address the class imbalance challenge. The augmented dataset is then used to retrain the ALL-Net, resulting in a notable improvement in test accuracy, reaching 99.32%. Along with accuracy, we have considered other evaluation metrics and the results illustrate the potential of ALLNet with an average precision of 99.35%, recall of 99.33%, and F1 score of 99.58%. Additionally, XAI techniques, specifically the Local Interpretable Model-Agnostic Explanations (LIME) algorithm is employed to interpret the model’s predictions, providing insights into the decision-making process of our ALL-Net CNN. These findings highlight the effectiveness of CNNs in accurately detecting ALL from PBS images and emphasize the importance of addressing data imbalance issues through appropriate preprocessing techniques at the same time demonstrating the usage of XAI in solving the black box approach of the deep learning models. The proposed ALL-Net outperformed EfficientNet, MobileNetV3, VGG-19, Xception, InceptionV3, ResNet50V2, VGG-16, and NASNetLarge except for DenseNet201 with a slight variation of 0.5%. Nevertheless, our ALL-Net model is much less complex than DenseNet201, allowing it to provide faster results. This highlights the need for a more customized and streamlined model, such as ALL-Net, specifically designed for ALL classification. The entire source code of our proposed CNN is publicly available at https://github.com/Abhiram014/ALL-Net-Detection-of-ALL-using-CNN-and-XAI.

Skin cancer involves abnormal growth of skin cells, typically caused by ultraviolet radiation exposure. Timely and accurate detection is essential to mitigate significant health risks and ensure effective treatment. This paper proposes a nanoantenna to enhance diagnostic and therapeutic capabilities for skin cancer detection. These antennas, emitting electromagnetic waves in the terahertz band (0.1–10 THz), improve integration for miniaturized wireless systems and serve as a foundation for the Internet of Medical Things (IoMT). We design a miniaturized, metamaterial-inspired gold-patch axe-shaped nanoantenna (), implemented in CST Studio Software. The antenna resonates at 1.152 THz, with a very low return loss (dB), a gain of 2.42 dBi, and a bandwidth of 40 GHz. The proposed antenna can be used as a sensor, considering the S11 spectra as a key parameter to differentiate between normal and cancerous skin (i.e., basal cell carcinoma). The simulation demonstrates significant and quantifiable differences between normal and cancerous skin and also highlights the proposed antenna’s suitability for applications such as radar systems, satellite communications, and advanced measurement technologies.

Pneumonia is a global health concern, especially in underserved regions. Traditional diagnostic methods, relying on costly chest X-rays, suffer from interpretation variances. To overcome these challenges, we’ve developed an advanced computer-assisted evaluation system utilising deep transfer learning techniques. Our approach aims to improve diagnostic accuracy and accessibility, particularly in resource-limited settings, offering a promising solution to enhance pneumonia diagnosis globally. Using a large dataset gathered from Kaggle, our novel approach uses convolutional neural network (CNN) models such as VGG16, ResNet-50, and InceptionNet-v5 to autonomously detect pneumonia in chest X-ray pictures. Deep transfer learning helps our models overcome data scarcity limits, allowing them to accurately recognise relevant visual attributes and patterns. Our methodology employs an ensemble approach, combining the strengths of each CNN model. We introduce a groundbreaking strategy for calculating optimal weights based on key evaluation metric such as accuracy. Evaluation using RSNA dataset shows remarkable accuracy: 92% with the CNN model and 84% with ResNet-50, promising improved diagnosis, especially in resource-constrained settings, potentially saving lives. This innovative approach represents a significant leap forward in pneumonia diagnosis, offering a scalable and reliable solution for healthcare providers globally. To sum up, our computer-aided diagnosis method offers a state-of-the-art approach to addressing the difficulties involved in diagnosing pneumonia. By combining ensemble modeling with deep transfer learning methods, we have created a very useful tool for correctly detecting pneumonia in chest X-ray pictures. This technology offers enormous promise for enhancing healthcare delivery and outcomes globally with additional development and use.