Recent Trends in Artificial Intelligence & It’s Applications (e-ISSN: 2583-4819, p-ISSN: 3107-7234)

Revolutionizing CGI and VFX with AI - Advancing Neural Rendering, Procedural Animation and Generation in Filmmaking

Abhay Kumar Mourya — 2026-06-24

This paper examines the far-reaching influence of Artificial Intelligence (AI) on Visual Effects (VFX) and Computer-Generated Imagery (CGI), with particular attention to realism, production efficiency, and the creative latitude afforded to filmmakers. AI-driven approaches — spanning Generative Adversarial Networks (GANs), Neural Radiance Fields (NeRFs), and procedural content generation — make it possible to construct photorealistic digital worlds and hyper-detailed virtual characters while keeping computational expenditure in check. Real-time rendering pipelines, strengthened by ray tracing and AI-powered denoising, allow instant on-set feedback, compressing post-production schedules. Suitless motion capture, physics-based animation, and automated compositing further liberate artists from repetitive technical tasks. By examining these technologies through both a theoretical and case-study lens, this review reveals how AI is repositioning itself not merely as a productivity tool but as a genuine creative collaborator in the filmmaking process. Persisting challenges — deepfake misuse, workforce displacement, and the black-box nature of certain neural models — are discussed alongside a forward-looking perspective on responsible AI integration.

Intelligent Real-Time Exam Surveillance and Anomaly Detection Using Deep Learning with Pose-Aware Behavioral Analysis

Anuradha M. Sandi — 2026-06-23

Maintaining academic integrity in examination halls is a key priority for educational institutions, yet traditional invigilation remains resource-intensive, subject to cognitive fatigue, and prone to oversight. While automated solutions have been proposed, they frequently rely on simple bounding box classifications that fail to capture the subtle, temporal postures associated with cheating. This paper presents an intelligent, real-time exam surveillance framework that integrates object detection, multi-object tracking, and pose-aware behavioral analysis into a unified, edge-deployable pipeline. The proposed system utilizes a fine-tuned YOLO11 Nano network to isolate students and unauthorized objects, a ByteTrack multi-object tracker to maintain persistent student identities across frames, and a secondary YOLO11-pose estimation model to extract seventeen anatomical keypoints in real time. A geometric behavior analysis engine computes head turn deflections, lateral body leaning ratios, and arm reaches based on anatomical joint vectors, passing these inputs into an exponential-decay suspicion scoring engine to filter out innocent, transient movements. Deployed as a lightweight Flask application with a real-time WebSocket dashboard, the system achieves an overall classification accuracy of 96.2 %, with an empirical precision of 89.5 %, representing a significant improvement over baseline YOLOv8 bounding-box models that suffer from excessive false alerts.

Enhanced Plant Disease Classification Using Deep Learning: An Attention-Guided Transfer Learning Approach

Rajan Kumar — 2026-07-18

Timely and accurate identification of crop diseases from leaf images remains one of the more stubborn problems in precision agriculture, largely because the visual differences between many diseases are subtle and the imaging conditions in the field are anything but controlled. This paper presents an enhanced deep learning framework for plant leaf disease classification that couples a lightweight EfficientNet-B0 backbone with a Convolutional Block Attention Module (CBAM) and a carefully tuned training pipeline. The attention module lets the network concentrate on lesion-bearing regions of the leaf rather than spreading its capacity across the whole image, while the training pipeline — combining transfer learning, an aggressive-but-realistic augmentation strategy, label smoothing, and mixup regularisation — pushes generalisation without inflating the parameter budget. Evaluation of the model is performed on the PlantVillage dataset (54,305 images across 38 disease-and-crop categories spanning 14 species), with benchmarking against six widely used convolutional architectures: VGG16, InceptionV3, ResNet50, MobileNetV2, DenseNet121, and a plain EfficientNet-B0. The proposed model reaches a test accuracy of 99.56% with a macro-averaged F1-score of 99.54%, outperforming every baseline while keeping the parameter count at roughly 5.4 million — an order of magnitude smaller than VGG16. An ablation study isolates the contribution of each component, and Grad-CAM visualisations confirm that the attention-augmented network attends to the diseased tissue rather than the background. A deliberately unflattering cross-dataset evaluation on field-condition images is also presented, where accuracy decreases to 71.8%, highlighting the challenges and implications for real-world deployment. The results suggest that modest architectural additions, applied thoughtfully, can deliver a favourable accuracy-to-cost trade-off suitable for mobile and edge diagnosis.

AI-Based Automated Defective Exhibit Identification System for Galleries

Pooja Patil — 2026-06-24

The preservation of cultural heritage exhibits such as paintings, sculptures, and historical artifacts is a critical task for museums and galleries. Over time, these exhibits may suffer from cracks, discoloration, surface erosion, or structural damage due to environmental exposure and human interaction. Traditional inspection methods rely on manual observation by experts, which is time-consuming, subjective, and often unable to detect early-stage defects. This paper presents an AI-Based Automated Defective Exhibit Identification System for Galleries, developed as an Android application using Java/XML and Firebase Realtime Database. The system allows gallery staff to upload baseline reference images of exhibits and later compare them with newly captured images using AI-based computer vision techniques. The comparison process identifies visual deviations that indicate possible defects and automatically records them with exhibit details, timestamps, and defect type. Detected issues are logged in real time, and notifications are sent to administrators for timely review and restoration tracking. Experimental evaluation shows improved detection accuracy, reduced inspection effort, and faster maintenance response. The proposed system offers a scalable, efficient, and cost-effective solution for intelligent gallery management and digital heritage preservation.

Correlation-SVM: A Multicollinearity-Aware Feature Selection Framework for SVM-Based Medical Diagnosis

Satish Kumar Kalagotla — 2026-05-22

Medical datasets often contain redundant or highly correlated features, leading to multicollinearity that adversely affects Support Vector Machine (SVM) classifiers by causing unstable decision boundaries, inflated coefficient variances, reduced interpretability, and degraded generalization performance, yet traditional feature selection methods inadequately address this issue. This paper proposes Correlation-SVM, a novel multicollinearity-aware feature selection framework that integrates Pearson correlation analysis and Variance Inflation Factor (VIF) computation within a hierarchical elimination process specifically optimized for SVM-based medical diagnosis. The framework operates in four stages: Pearson correlation analysis to identify highly correlated feature pairs; VIF computation to quantify multicollinearity severity; hierarchical feature elimination to iteratively remove redundant features while recomputing VIF after each removal; and SVM training with cross-validation evaluation. Evaluated on four benchmark medical datasets (Wisconsin Breast Cancer, PIMA Indian Diabetes, Hepatitis, and Mammographic Mass) and compared against six state-of-the-art methods (CFS, FCBF, mRMR, SVM-RFE, LASSO, and GA-SVM) using 10-fold cross-validation with five repeats, Correlation-SVM achieved 97.42% accuracy on the Wisconsin dataset with only 5 features (44.4% reduction), outperforming all comparison methods. Multicollinearity was substantially reduced, with maximum VIF decreasing from 8.3 to 2.3 (72.3% reduction) on Wisconsin, from 12.5 to 2.1 (83.2% reduction) on Hepatitis, and from 4.2 to 1.6 (61.9% reduction) on PIMA, achieving VIF values below the acceptable threshold of 2.5. The framework requires only 38.7 seconds of computational time, making it 84% faster than GA-SVM and 79% faster than SVM-RFE, thus achieving wrapper-like performance with filter-like speed. Additionally, the selected feature subsets align with established medical knowledge across all four datasets, enhancing clinical interpretability and trust. Correlation-SVM provides an effective, computationally efficient framework for multicollinearity-aware feature selection in SVM-based medical diagnosis, achieving substantial feature reduction, eliminating multicollinearity, and improving classification accuracy while maintaining interpretability.