Journal of Artificial Neural Networks and Learning System (p-ISSN: 3049-0758, e-ISSN: 3048-6629)

Deep Learning-driven Computer Vision for Reliable Autonomous Driving: Perception, Sensor Fusion, and Future Directions

Suraj R. Nalawade, Tapase H. O., Madhura Yadav — Thu, 17 Sep 2026 00:00:00 +0000

Computer vision is a central perception technology for autonomous driving because it converts visual observations of the road environment into information that can be used for navigation and decision support. A vehicle must recognize road users, estimate their locations, identify lanes and traffic signs, interpret traffic lights, and understand the surrounding scene while operating under strict latency and reliability constraints. This study presents a structured review of deep learning-driven computer vision techniques for autonomous driving, with emphasis on object detection, lane and road-boundary detection, semantic and instance segmentation, traffic-sign recognition, tracking, depth estimation, and three-dimensional perception. It also examines the role of convolutional neural networks, one-stage and two-stage object detectors, point-cloud learning, and sensor-fusion strategies that combine cameras, LiDAR, and radar. Rather than treating a perception model as an isolated classifier, the study considers the complete perception pipeline, including data preparation, inference, temporal tracking, uncertainty, computational deployment, and validation. The results synthesize the capabilities and limitations of the reviewed approaches and show why accuracy alone is insufficient for safety-critical autonomous driving. Particular attention is given to adverse weather, illumination changes, occlusion, dataset bias, domain shift, computational cost, and the need for interpretable and verifiable outputs. Future perspectives include edge AI, transformer-based perception, improved 3D scene understanding, simulation-based testing, explainable AI, continual adaptation, and multimodal fusion. The review concludes that dependable autonomous driving requires an integrated perception stack in which algorithmic accuracy, real-time execution, robustness, and system-level validation are considered together.

Consistency-Driven Self-Supervised Feature Learning for Limited-Sample PolSAR Image Classification

Saboor Saniya, S. Sehar Tasneem, S. Siddarth, Sadanand, Gnanamani H. — Mon, 28 Sep 2026 00:00:00 +0000

Polarimetric Synthetic Aperture Radar (PolSAR) images provide valuable information for identifying and monitoring different types of land cover. However, deep learning-based PolSAR classification generally requires large amounts of labelled training data, while obtaining reliable annotations is time-consuming and costly. Self-supervised learning offers an alternative by learning useful representations from unlabelled data before downstream classification. This study implements and systematically evaluates the Self-Supervised Learning with Multibranch Consistency (SSL-MBC) framework, originally introduced for few-shot PolSAR image classification. The framework exploits complementary PolSAR representations through multiple branches and employs consistency learning to obtain robust feature representations from unlabelled samples. The pretrained representation is subsequently transferred to a few-shot classification task using a limited number of labelled samples. The framework is evaluated on the Flevoland, San Francisco, and Oberpfaffenhofen PolSAR datasets. In addition to describing the architecture and training procedure in greater technical detail, this study presents dataset-specific experimental results and analyzes the contribution of multibranch representation learning under limited-label conditions. The revised experimental protocol aims to improve reproducibility and provide a clearer understanding of the effectiveness and limitations of SSL-MBC for few-shot PolSAR image classification. The results further support systematic assessment across datasets and label regimes and practical few-shot classification scenarios.

Natural Language Processing in Vedic Science: A Comprehensive Review and AI-based Framework for Knowledge Extraction

Rajesh Ramnaresh Yadav — Tue, 29 Sep 2026 00:00:00 +0000

The Vedas represent one of the oldest repositories of human knowledge, encompassing diverse domains such as philosophy, medicine, astronomy, mathematics, linguistics, ethics, and spirituality. Composed in Vedic Sanskrit and preserved primarily through oral traditions, these texts possess complex grammatical structures, rich morphology, and profound semantic depth that make computational interpretation a challenging task. Recent advances in artificial intelligence (AI) and Natural Language Processing (NLP) have opened new avenues for the systematic analysis, preservation, and interpretation of ancient textual resources. Modern NLP techniques, including machine learning, deep learning, transformer-based language models, semantic parsing, dependency analysis, and knowledge graph construction, provide powerful tools for extracting structured knowledge from unstructured Sanskrit texts. This paper presents a comprehensive review of NLP applications in Vedic science and examines the evolution of computational approaches for Sanskrit language processing. It discusses the linguistic characteristics of Vedic Sanskrit, existing computational resources, and the major challenges associated with digitization, morphological analysis, semantic interpretation, and multilingual translation. Furthermore, the paper proposes a conceptual AI-driven framework integrating corpus creation, preprocessing, linguistic analysis, semantic modeling, ontology construction, and intelligent knowledge retrieval. The proposed framework aims to preserve the authenticity of Vedic literature while improving accessibility for researchers, educators, and interdisciplinary scholars. The study concludes that the convergence of NLP and Vedic science has the potential to revolutionize digital humanities by enabling semantic search, intelligent question-answering, automated annotation, and knowledge discovery from ancient Indian scriptures.

SkinGuard 1: A Robust Multi-Class Deep Learning-Based Skin Cancer Detection Using a ResNet50-Based Model

Thu, 08 Oct 2026 00:00:00 +0000

Skin cancer is an increasingly prevalent health concern, and early detection is essential for effective treatment and improved patient outcomes. Traditional diagnosis largely depends on dermatologists’ visual interpretation of skin lesions, which can be subjective and may result in variations in diagnosis or delays in treatment. This study presents an automated skin cancer detection system based on ResNet-50, a deep convolutional neural network architecture designed for image classification. The model was trained and evaluated using a dataset obtained from the International Skin Imaging Collaboration (ISIC). To improve accessibility and practical usability, the trained model was integrated into a user-friendly web application called SkinGuard, which enables users to upload skin lesion images and receive real-time predictions. Experimental results demonstrate a multi-class accuracy of 88%, precision of 91%, recall of 87%, and F1-score of 89%. These findings indicate that the proposed system has potential to support dermatologists in early skin cancer detection.

Biomedical Signal-to-Image Representations for Artificial Intelligence: A Review

Donthagani Srihas, Aditya Sarin, Samala Thanay Kumar, Sourav Suman, Ashish Jain — Fri, 09 Oct 2026 00:00:00 +0000

Biomedical signals such as Electrocardiogram (ECG), Electroencephalogram (EEG), Surface Electromyogram (sEMG), Electromyogram (EMG), Electromyotomogram (EMT), and Photoplethysmography (PPG) are among the most commonly used physiological signals. They provide useful information for diagnosis and monitoring, rehabilitation, and wearable healthcare. However, these signals are not easily analyzed directly because physiological recordings are often non-stationary, nonlinear, noisy, and variable across different subjects and recording instances. This has led to the development of a large number of approaches that transform one-dimensional signals into two-dimensional representations, which can be processed using image-based artificial intelligence. This review discusses spectrograms, scalograms, EMD and Hilbert–Huang representations, recurrence plots, Gramian Angular Fields, and direct image encodings. Their suitability is evaluated with respect to the information they provide and the requirements of downstream learning tasks. The review also compares CNNs, transfer learning, hybrid CNN–recurrent neural networks, and LSTMs, as well as attention networks, Transformers, Vision Transformers, and self-supervised learning. Key drawbacks are transformation sensitivity, domain shifts, inconsistent evaluation protocols, computational expense, and poor interpretability. The review thus focuses on the co-selection of signal representation and AI architecture. Future research should look into adaptive, data-efficient multi-modal explainable edge-oriented architectures and explore potential paths for further research.