The proposed ACL-PEFT-LLM model combines Adaptive Curriculum Learning (ACL) and Parameters-Efficient Fine-tuning (PEFT) to overcome the shortcomings of current models that employ full fine-tuning, which is computationally expensive and impractical in many domain-specific applications. Although PEFT methods such as LoRA, Prefix Tuning, and Prompt Tuning have been designed to optimize only a few model parameters, thereby reducing memory usage and training time by a significant factor, typically lack an adaptive mechanism to adjust task difficulty during training. By comparison, ACL-PEFT-LLM dynamically adjusts the difficulty of training examples based on the model's current performance, enabling it to start with easier tasks and increase in difficulty. This will guarantee that the model learns effectively without being flooded with challenging examples during the initial stages of learning. The ACL-PEFT-LLM model is superior to the other models in both accuracy and computational efficiency. It has an F1 score of 96.2 and a high accuracy of 96.8, indicating strong task-specific performance across a wide range of datasets, including SST-2, SQuAD, and AIME. Besides, it has a high Accuracy-Efficiency Ratio (AER) of 5.12, indicating a positive trade-off between performance and resource consumption. Compared with other methods, LoRA is most efficient in terms of training time and GPU memory usage, with an accuracy of 94.0 and an F1 score of 93.7, but is marginally lower in performance. Other techniques, such as Full Fine-Tuning, Prefix Tuning, and Adapters, are either less accurate or more resource-intensive, whereas ACL-PEFT-LLM is the most efficient mechanism for domain-specific fine-tuning. ACL-PEFT-LLM offers a potent platform that facilitates efficient domain adaptation at minimal resource usage, which makes it suitable for tasks in domains such as medical, legal, and financial, where computing resources are usually scarce. This model has found a very good compromise between maximum performance and computational efficiency such that tasks that belong to a domain can be efficiently performed without overuse of resources.
Cybercriminals have become increasingly interested in the spread of critical information, particularly in interpersonal contact and mass distribution of programs and file downloads. This has heightened researchers' awareness of the rampant spread of malware and data breaches. It is anticipated that the number and intensity of malicious software will continue to rise, underscoring the imperative need for strong security architectures. This is especially critical for mobile networks and ubiquitous computing security, given the growing threat posed by hackers. The proposed project involves creating a new dataset using a strictly controlled, shared sample pool to address security threats in wireless mobile networks, leveraging dynamic analysis techniques for malware detection. The dataset aims to enhance the recognition of malicious software by leveraging methods such as encryption and obfuscation. The suggested classification algorithm is CNN-LSTM, a combination of Convolutional Neural Networks (CNNs) and the Long Short-Term Memory (LSTM) model, which excels at learning complex, sequential features. The CNN and LSTM models were tested on a dataset comprising more than 10,000 malware samples and achieved accuracies of 98% and 97%, respectively. These findings demonstrate how deep learning models can be used to enhance the security of mobile networks and provide effective protection against emerging threats in mobile and ubiquitous computing systems, in a highly beneficial way.
Breast cancer is one of the most common and deadliest diseases that women globally experience. An early and accurate diagnosis may improve the chance of treatment success and increased survival rate. Current traditional diagnostic techniques used for breast cancer like mammography, ultrasound, and MRI are largely annotated by experts, creates variability and inconsistency in the diagnosis. New developments in artificial intelligence (AI) and deep learning could offer promising solutions for touchless high-precision breast cancer diagnosis. This study proposes an Advanced Multi-View Convolutional-Recurrent Network (AMVCRN), a unique design that integrates Convolutional Neural Networks used for extracting spatial features, with Recurrent Neural Networks, using Gated Recurrent Units – GRUs, for temporal sequence modeling. The hybrid model allows greater analysis of tumor features from multi-view images collected from different modalities and angles. The intended outcomes are to improve classification accuracy and reduce errors in persecution. Ultimately, it is intended to yield and sound a decision support system for clinical radiologists.
Today, malicious actors often use social media to send, receive, and post false, misleading, or offensive content about other people. The effects of social media bullying on its targets are similar to those of threats, gossip, and harassment in the physical world. An alarming rise in mental health issues has resulted from cyberbullying, particularly in the younger population. The effects can include suicidal thoughts and low self-esteem. Several conventional machine learning techniques have been employed to automatically identify cyberbullying on social media. Although some traditional methods of machine learning have been designed to automatically identify cyberbullying, most of them are characterized by weaknesses in feature selection and classification rates. In order to deal with such issues, a new Termite Apriori Prediction Framework (TAPF) has been proposed in this paper. The model consists of a data collection and preprocessing phase to remove noise and a hybrid feature selection phase that combines Apriori rule mining with termite optimisation to select the most discriminative features. Cyberbullying is then classified with the help of these optimized features. It has been experimentally shown that TAPF performs better than the classical models such as the Logistic Regression, Naive Bayes, Long Short-Term Memory (LSTM), and the Support Vector Machine (SVM). The proposed structure has a recall rate of 96.2, precision of 96.7, F-score of 96.3, and an overall accuracy of 96.4, which is over 2 per cent higher than the best-performing baseline (SVM). Applied to a Python environment, TAPF offers an effective and dependable mechanism of detecting cyberbullying on social media sites.
The success and the importance of industrial robotic arms in contemporary automation rely on their accuracy, velocity, and capability of functioning under varied conditions, but mechanical and electromechanical deteriorations, including bearing wear, gear backlash, joint misalignment, and imbalance of motor torque, may negatively affect performance, safety, and production unless inner problems are revealed early. The traditional methods of maintenance, such as reactive repairs and fixed-interval inspections, are inefficient and do not capture the faults in their early stages, whereas the current machine learning-based solutions typically process sensor signals separately, neglecting inter-joint kinematic constraints, and are not physically interpretable, leading to poor performance in cases of rare or hidden faults. This paper will solve these challenges by introducing a Hierarchical Temporal-Graph Physics-Guided Fusion Network (HT-GP-FusionNet) in predictive fault diagnosis of robotic arms. The framework incorporates hierarchical temporal modeling to decompose short-term and long-term dynamics in multi-sensor data consisting of accelerometers, gyroscopes, motor currents, and joint positions, and a graph neural network fully captures the inter-joint relationships and fault propagation along the kinematic chain. It uses a physics-based regularization that requires consistency with motion equations and energy conservation laws, and generates samples of faults to be generated by a generative fault augmentation model to promote few-shot generalization. The experimental results on a set of 50,000 sequences (10 sensors, five fault types) sampled at 1 kHz show that HT-GP-FusionNet outperforms CNN (89.3), LSTM (90.5), CNN-LSTM (92.8), and GNN-based (93.6) models in all metrics and has higher recall with rare faults, and can be used when the number is unbalanced. These conclusions have been proved by ablation experiments that show temporal modeling, graph reasoning, physics-guided regularization, and data augmentation make significant contributions to performance. In general, HT-GP-FusionNet offers a scalable, interpretable, and physics-consistent solution to the early and accurate predictive maintenance, hence higher reliability, safety, and operational efficiency during industrial robotic systems.
Introduction: Facial Emotion Recognition (FER) is vital to the study of human emotion, as it enables the analysis of facial expressions. Nevertheless, current FER techniques do not effectively identify emotions in people with facial wrinkles, which are common in older people, leading to incorrect classifications. This research aims to recommend a superior FER model that considers facial wrinkles as a way of improving the identification of emotions and the degree of emotion. Methodology: The hybrid methodology is introduced. Processing of input photos begins with noise removal and contrast enhancement. Facial features are then derived using the Logarithmic Function-centric Viola-Jones (LF-VJ) algorithm, which detects faces and segments facial objects. The wrinkle score and cosine similarity are computed to measure the existence of wrinkles. Entropy Weighted Secretary Bird Optimization (EWSBO) is used to do optimal feature selection and then RSRN is used to classify emotions. STFIS has to do intensity estimation to assess the intensity of identified emotions. Results: The accuracy of the proposed framework on emotion classification was 98.97 and precision, recall and F-measure values were beyond 99. The estimation of the intensity was correct and rule generation, fuzzification and defuzzification time were 958.697ms, 923.013ms and 848.998ms, respectively. It was able to efficiently deal with wrinkles, lighting effects and lessen the complexity of features and was much faster and more accurate than current methods. Conclusion: It is proven in the current study that wrinkles in the FER frameworks enhance the classification accuracy and emotion strength prediction greatly, and thus the offered approach is appropriate to be applied to real time applications, such as mental health monitoring.
Individualized e-learning systems plays a major role in the dynamic learning systems in which learners vary widely with respect to the background knowledge, the rate of learning, and the degree of engagement, as well as contextual factors like the availability of devices and the quality of the network. Nevertheless, the majority of current e-learning systems are based on the usage of either static or partially adaptive personalization models, which do not react to the changing contextual factors and the shift in learner behaviour in real-time. This drawback lowers learning performance, interaction, and knowledge retention. This paper aims at creating and testing a Context-Aware Adaptive Learning Model (CAALM) that can dynamically customize learning material and instructional approaches depending on dynamically monitored learner and environmental conditions. The proposed model combines multi-dimensional context sensing (learner performance, interaction behaviour, time-on-task, and device context) with an adaptive decision engine that manipulates content difficulty, sequencing, and presentation modality in real time. The model was deployed and tested on a real-world e-learning interaction dataset consisting of 1,200 learners, 18,000 learning sessions, and 45 contextual features divided into training (70%), validation (15%), and testing (15%) sets. It was measured in terms of performance against a non-contextual baseline and a static personalization model. The experimental outcomes indicate that the suggested method allows reaching the 17.8 % improvement in the learning gain, a 14.3 % improvement in the course completion rate, and a 21.6 % decrease in the mean response latency indicators. Statistics verification with paired t-tests proves that the process of improvements was significant (p < 0.01), and one-way ANOVA demonstrates that the performance improvement was the same in various groups of learners (F = 6.42, p < 0.05). The findings affirm that the use of real-time contextual awareness is an effective way of boosting adaptive learning. The proposed model presents a scalable and reliable model of next-generation personalised e-learning systems that could be used in the dynamic educational environment.
The social media is producing vast amounts of user-generated text, which can serve as a great indicator of initial mental health diagnosis. This paper develops a scalable, multilingual depression classifier based on classical machine learning (ML) methods and state-of-the-art, pretrained transformer-based models to overcome the weaknesses of language-specific and binary-only methods in previous studies. In a contrast to the majority of the studies, the work is a systematic exploration of bilingual and multilingual depression recognition in the context of Arabic, English, Russian, and Spanish data in a single pipeline. TF-IDF is used to represent textual information to conventional ML classifiers, such as SVM, Random Forest, Naive Bayes and AdaBoost, and transformers, such as XLM-RoBERTa and XLNet are used to train contextual semantic representations. Decades of experiments demonstrate that models using transformers always perform better in comparison to traditional models of machine learning. XLM-RoBERTa provided 94.33% accuracy, 0.94 F1-score, and 0.99 AUC, which outperforms SVM (93% accuracy) and means a lot in terms of preforming XLNet (72.36% accuracy). XLM-RoBERTa achieved 99.5% accuracy in Russian, 98% in English, 96% in Arabic, and 85.9% in Spanish in single-language tests, which shows that it is strong in various languages. The findings reveal the usefulness of pretrained multilingual transformers to identify subtle cases of depression, which offers a dependable, language-independent approach to screening early cases of digital depression in mental-health monitoring systems in the real world.
Advanced microarchitecture courses demand rich, hands-on exploration of pipelines, caches, and memory systems. Though traditional hardware labs are costly, location-bound, and difficult to scale for remote or hybrid delivery. This paper presents a cloud-hosted, multi-user virtual reality (VR) laboratory designed specifically for advanced microarchitecture education and remote teaching. The proposed platform delivers an experiment-rich environment where students collaboratively inspect, instrument, and modify microarchitectural components such as pipeline stages, cache hierarchies, and branch predictors in real time. Architecturally, the system combines a web-based cloud front-end for authentication and session management, a scalable VR services layer providing multi-user scenes and collaboration tools, and a backend to adapt microarchitecture simulators whose internal state is visualized in 3D. Pedagogically, define learning objectives around instruction-level parallelism, hazard analysis, and memory hierarchy behavior, and instantiate these through structured labs on pipeline hazards, cache performance, and branch prediction. A mixed-method evaluation in an advanced microarchitecture course contrasts a control group using traditional 2D tools with an experimental group using the VR lab over 4–6 weeks. According to quantitative findings, the VR group outperformed the control group by 24.5% in post-test scores, decisive the idea that immersive collaborative contact is a key factor in conceptual understanding. The cloud-based architecture may provide responsive multi-user experiences under practical bandwidth limits, as shown by system-level metrics like latency and frame rate. Future integration of hardware-in-the-loop and adaptive learning analytics will be informed by qualitative feedback from instructors and students that demonstrates how immersive visualization and collaborative interaction may demystify complicated microarchitectural behavior and ease remote teaching.
Early and accurate diagnosis of Acute Lymphoblastic Leukemia (ALL) is essential to treat it successfully and improve patient outcomes. In this paper, a modified VGG-based hybrid architecture that combines an adapted K-means segmentation algorithm is proposed to classify leukocytes in blood smear images using this framework, and the researcher is able to classify blood cells based on the similarities between them. The segmentation step transforms images to the HSV color space and dynamically recalculates cluster centers using both color and morphological constraints, which ensure accurate leukocyte isolation in situations of overlapping cells and variability of staining. The fragmented areas are then subjected to a hybrid VGG network with residual skip-connection and attention modules to allow the extraction of discriminative features of the nuclear and cytoplasmic structures robustly. The framework was tested on 260 high-resolution images in the ALL-IDB dataset and the performance measured by metric values of accuracy, precision, recall, F1-score, specificity, and the Matthews Correlation Coefficient (MCC). The model achieved 98.5% accuracy, 96.5% precision, 97.0% recall, 98.5% F1-score, 97.9% specificity, and an MCC of 0.96, which is much higher than those of the state-of-the-art models ResNet-50, DenseNet121, and DDRNet. The results of cross-validation revealed low variance (±1.2%), indicating strong generalization across folds. Also, the mean processing time of 16.5 ms per image indicates that it can be used in real time. These findings indicate that advanced segmentation, hierarchical feature learning, and attention can be associated not only with better classification results but also with adequate reproducibility and reliability, and can be considered a clinically applicable tool for automated real-time detection of ALL in blood smear analysis.
The study presents a new hybrid deep learning framework that combines the Multi-Function Differential Evolution (MF-DE) global optimization engine and the Morphological-OTSU analysis to improve the role of image segmentation and pattern recognition in this study, named as LungNet. The MF-DE engine has been utilized to optimize the hyperparameters of LSTM (Long Short-Term Memory). The morphological-OTSU method is adopted to perform effective image thresholding to help in the extraction of important features of images to enable proper classification. The proposed model is evaluated using the benchmark datasets, i.e., the LIDC-IDRI lung CT dataset in medical imaging and the NSCLC Radiogenomics dataset in lung cancer research. The MF-DE optimization ensured that the LSTM models were much more accurate, with the hyperparameters being optimized towards the optimal performance of the model. On LIDC-IDRI, the hybrid mode achieved a higher segmentation accuracy of 93.2% as compared to conventional methods that had an average segmentation accuracy of about 86.5%. Equally, the object recognition performance on the NSCLC Radiogenomics dataset achieved an mAP (mean Average Precision) of 48.0% that is 7.5% better than the baseline CNN models. The Hybrid Model had a classification accuracy on the LIDC-IDRI dataset of 95.0%, and on the NSCLC Radiogenomics dataset, the classification accuracy was 94.5%. This paper shows that the suggested hybrid design is better in global optimization and image analysis and provides significant improvements in real-time image processing tasks. The findings reveal a significant change in computational efficiency and accuracy of segmentation, especially with challenging and noisy data sets. The hybrid paradigm can be used in a very diverse range of applications, such as in medical image processing, industrial automation, and security surveillance systems.
To develop and test a foundation-level multimodal ophthalmic model that learns common cross-modal representations for automated classification of eye diseases (normal, diabetic retinopathy, glaucoma, cataract) using the Kaggle Eye Diseases classification fundus image dataset and textual descriptors. The proposed framework is based on a modality-agnostic vision encoder, initialized via transfer learning and trained on 4 categories of 4217 color fundus images, and a lightweight text encoder fed by textual tokens for label and description levels. The features of the fundus and text embeddings are matched in a shared latent space via image-text contrastive objectives, leveraging recent multimodal ophthalmic foundation models such as EyeCLIP and Eye Found. In this single space, a classification leader is used to perform multi-class disease prediction, enabling image-only and image-and-text inference. The proposed multi modal model with extensive data augmentation reaches test accuracy of 95, on the same Kaggle dataset, which is comparable or a little higher than current Efficient NetB3 based and transformer ensemble baselines, which report a test accuracy of 95. The model has a high macro averaged precision, recall, and F1 scores in all four classes, and significantly less confusion between cataract and glaucoma than the single modal CNN and transformer baselines. Experiments of ablation demonstrate that either the removal of the text arm or the contrastive alignment goal deteriorates performance and class balance which confirms the advantage of learning both cross modal representation unanimously as supported by previous multimodal ophthalmic experiments.Multi-modal ophthalmic style, based on a foundation style and trained on a single public Kaggle fundus dataset, can acquire unified cross-modal representations that result in robust classification of eye disease across multiple classes. It can be extended to other imaging modalities (e.g., OCT) and to more detailed clinical text, which aligns with the direction of large multimodal foundation models in ophthalmology. This is why the given approach can be considered an effective starting point of scalable, real world ophthalmic AI systems.
Online collaborative learning communities are susceptible to the problems of lack of peer interaction, skewed participation, and poor knowledge sharing, which adversely affect learning outcomes. The recent developments in artificial intelligence (AI) provide the opportunity to solve these problems with the help of adaptive learning analytics and intelligent collaboration support. The paper suggests an AI-based Collaborative Learning Model (AI-CLM) that is aimed at improving peer interactions and sharing of knowledge during an online study. The suggested model combines the analysis of the interaction of the learners, peer grouping using AI, adaptive feedback, and knowledge recommendation in one framework. It utilizes a systematic algorithm and mathematically formulated evaluation measures to guarantee reproducibility and stringent evaluation. An online collaborative learning course of eight weeks has been evaluated using an experimental mode and showed that the proposed AI-CLM not only performs significantly higher than the traditional collaborative learning methods. Precisely, the degrees of participation boosted since the Peer Interaction Index grew by 12.4 to 21.8, the Knowledge Sharing Score rose by 68.2 to 84.6, and the Participation Balance Index rose by 0.54 to 0.81, which is more equitable learner participation. Moreover, the average learning performance increased to 85.3 percent as compared to 72.5 percent, and the Collaboration Effectiveness Score also increased to 0.86 compared to 0.63. These findings can be used to conclude that AI-based adaptive collaboration support can significantly enhance social interaction and academic performance during online learning. These results can be relevant to the development of scalable, smart online learning systems that facilitate efficient collaboration and interaction with a learner.
The educational metaverse will be improved greatly with the fast development of 6G networks, which offer high-speed bandwidth and low latency, and the ability to support many devices at once. The paper provides a critical assessment of the self-managed resource management approaches that are vital in ensuring technological reliability and quality of learning in the 6G-enabled metaverse. Among the major threats are the maintenance of network stability, the availability of resources, and minimizing latency, which is crucial when dealing with real-time interactions in virtual learning environments. Interactive technologies like Virtual Reality (VR) and Augmented Reality (AR) are also useful to complement learning because they provide hands-on experiential learning, which is central to increasing student engagement and ease of access to the content. As noted in the review, some of the AI-based solutions, including reinforcement learning and federated learning, can be used to optimize the allocation of resources in real time, thereby facilitating the smooth functionality of the metaverse platforms. Although current strategies are promising, technological bottlenecks, such as network congestion, interference, and security issues, are still a major challenge. The future directions of research involve finding out the sustainability practices, enhancing interoperability among the metaverse platforms, and finding solutions to the question of data privacy. Network engineering, AI, and education science require interdisciplinary efforts to overcome these difficulties and guarantee the successful implementation of a stable and quality 6G-powered educational metaverse. The study will help to offer valuable information on creating efficient, scalable, and secure systems to support further educational metaverses. The review of 51 studies has shown that AI strategies can decrease latency by up to 50%.
The prediction of heart disease with several comorbid conditions (hypertension, diabetes, arrhythmia, and obesity) makes the prediction complex due to nonlinear interactions between conditions and heterogeneous risk patterns. As a solution to these research issues, this paper suggests an Adaptive Convolutional Neural Network (CNN) architecture to predict multi-comorbid heart disease on a unified U.S. Heart Disease dataset (n = 1,025) that is comprised of a union of Cleveland, Hungarian, Switzerland, and VA repositories. Out of the 76 attributes, 24 clinically validated features were then picked up based on correlation ranking, SHAP-based interpretability analysis, and cardiologist validation. These characteristics were converted to structured representations of images through a feature-to-image encoding plan that facilitated learning of patterns that are deep both in space and relationally. In the proposed CNN architecture, layer-driven optimization, regularization through dropout, adaptive learning-rate scheduling, and reproducibility control are included to guarantee consistent generalization in a wide range of comorbidity groups. SHAP-based feature attribution generates clinical and understandable explanations about model predictions, achieving clinical interpretability. There is strong performance on experimental evaluation with 92% accuracy, 89% precision, 90% recall, an F1-score of 0.89, and an ROC-AUC of 0.94, and high sensitivity and specificity that is applicable in a clinical decision-support setting. The proposed framework can provide scalable, interpretable, and clinically reviewed methods of automated multi-comorbid cardiac risk assessment to aid in the integration of AI transparency into healthcare system components and clinical implementation.
Environmental monitoring systems for remote areas and areas with limited resources are constantly generating large volumes of time-series sensor data, resulting in excessive energy consumption, storage overhead and scalability issues. Traditional static and event-driven sampling techniques are likely to yield redundant information, reducing operational sustainability and efficiency. To address the above-mentioned limitations, the present study proposes a novel framework, Adaptive Time-Based Data Reduction with Blockchain (ATDR-BC), which combines dynamic timestamp-based data sampling, Isolation Forest-based anomaly detection, and Merkle Tree-enabled decentralised storage. The proposed framework is an intelligent way to adjust sampling intervals based on environmental stability, without compromising critical anomaly detection or tamper-resistant data integrity in a hybrid on-chain/off-chain blockchain model. Experimental evaluation based on actual environmental sensor data shows that ATDR-BC can achieve an average data reduction rate of 45-60%, and the maximum data reduction rate is 70% under stable conditions. The approach leads to 35-50% reduction in energy consumption compared to conventional sampling methods with high reconstruction accuracy (96.59%) and RMSE values that are smaller than 5%. These results confirm that ATDR-BC provides a good balance between energy efficiency, data reliability and storage optimisation and can be used for applications involving long-term deployment in remote and resource-limited applications such as environmental monitoring.
This study examines the utilization of a Deep Convolutional Neural Network (DCNN) for identifying defects in fruit via digital image processing. The primary challenges encountered in the agricultural and fruit distribution sectors are the inefficiency and variability of the manual inspection process, necessitating the development of an accurate and dependable automated system. This study conducted an experiment comparing the GoogleNet baseline model with the Proposed Method, a modified version of GoogleNet that incorporates transfer learning and final layer optimization to enhance classification performance. The fruit image dataset underwent pre-processing and data augmentation to enhance data variability and improve the model's generalization capability. The data was then divided into three categories, namely training, validation, and test, with 80 % of the data set allocated to training, 10 % to validation, and 10 % to test. The two models performed well in the course of training, although the Proposed Method had a greater advantage. The proposed method achieved a recall, precision, and F1-score of 0.99 and an accuracy of 0.9980. GoogleNet recorded the value of recall, precision, and F1-score as 0.98 and accuracy of 0.9870, respectively. These findings indicate that transfer learning and layer modification of the GoogleNet structure can be used to help fruit defect detection processes achieve improved results. The suggested approach, which is highly precise, may be applied in automated inspection systems within the area of the fruit industry to make the products superior and the processes more productive. This paper also reveals that it is possible to use DCNN and appropriate optimizations so that processing fruit images can be a good idea.
The adoption of automated image-based classification systems was motivated by the fact that traditional manual approaches are subjective, time-consuming, and often erroneous. This study details the process of illness detection in pigeon pea leaves using the Mendeley Data Pigeon Pea Leaf Illness Dataset, which is accessible to the public. Its ability to distinguish between objects is substantially enhanced by the proposed model, which integrates an optimised convolutional neural network (CNN) architecture with feature extraction layers and attention-based augmentation. To use TensorFlow to run tests on a system with an NVIDIA GPU to see how the suggested model compares to baseline models and typical designs like VGG16, ResNet50, DenseNet121, and then MobileNetV2. When it comes to pigeon pea leaf disease identification, this paper employs a modified CNN to circumvent shortcomings of traditional deep learning constructions like VGG16, ResNet50, and DenseNet121. Common issues with traditional CNNs include their high processing cost, their inability to handle real-world differences in lighting, occlusion, and background noise, their sluggish inference speed, and the likelihood of overfitting on small agricultural datasets. Improved feature extraction, regularisation, and lightweight optimisation are all part of the suggested Modified CNN's architectural overhaul, which fixes these problems and makes the network more efficient and accurate. Quantitative evaluations employing illustrate the proposed model's superior performance, attaining an overall accuracy of 97.86%, exceeding prior work by 5.4%. The results show that the suggested deep learning-based method can be a useful and scalable way to automatically find Pigeon Pea diseases. This will help precision agriculture and smart crop monitoring systems.
Diabetic Retinopathy (DR) is a major cause of vision loss in the world hence the relevance of automated and precise detection systems. This paper introduces a superior method of automated detection of DR on the basis of transfer learning (TL) and ensemble models (ET). The technique uses the APTOS-2019 dataset, which uses five categories of DR severity, namely: No DR, Mild, Moderate, Severe, and Proliferative DR. The research employs two trained deep learning networks, Inception V3 and Xception, that are fine-tuned in order to obtain discriminative features of retinal images. In order to enhance the overall performance of classification, an ensemble approach, with the average and weighted voting is used, to utilize the outputs provided by both models. The performance measures applied to assess the proposed framework include standard performance measures, including accuracy, F1-score, scalability, and patient outcomes. The improvements are significant and the accuracy, F1-score and patient outcomes ratio are 97.94%, 98.41 % and 97.52 % respectively. Also, the scalability ratio is 97.63% and the rate of early diagnosis is 96.84 %, which indicate that the model can be considered robust and efficient and can be used to work with larger data volumes and provide the possibility of timely DR identification. The approach provides a valid, accurate, and generalizable solution to the early diagnosis and treatment planning of DR. High transfer learning methods and ensemble models also guarantee the high level of classification, which is why this approach can be effectively deployed to clinical practice in practice. In addition, the suggested system increases the level of healthcare accessibility as it offers automated solutions to resource-limited settings.
The rapid evolution of Generative Artificial Intelligence (AI) has significantly increased the prevalence of high-fidelity deepfakes, posing a severe threat to digital security and social media integrity. Existing detection frameworks often struggle with generalization and the identification of multi-scale spatio-temporal artifacts. This research addresses these challenges by proposing a robust detection system utilizing the Deep Inception V4 architecture integrated with Predictive Representation Learning. By leveraging multi-scale convolution kernels and specialized reduction blocks, the proposed model extracts both microscopic pixel-level inconsistencies and macroscopic structural anomalies commonly found in manipulated videos. The methodology was rigorously evaluated using the large-scale Deepfake Detection Challenge (DFDC) dataset. Experimental results demonstrate that the Inception V4 model achieves a Training Accuracy of 94.87% and a Validation Accuracy of 92.74%, representing a substantial improvement over baseline CNN (82.00%) and VGG-19 (80.00%) architectures. Statistical validation using a two-sample t-test yielded a p-value < 0.001, confirming the significance of these performance gains. Furthermore, the model achieved a Receiver Operating Characteristic (ROC-AUC) of 0.96, indicating high precision and recall in identifying 2,308 true positive cases within the test subset. These findings conclude that multi-scale feature extraction, combined with self-supervised pre-training, provides a superior defense against generative threats. Future work will investigate the integration of lightweight hybrid transformers to enhance real-time detection on mobile platforms while maintaining high classification sensitivity.
