B. N. Anoop

Plant diseases significantly threaten global food security, necessitating accurate and automated diagnostic systems. This work presents a modular deep learning framework for systematic benchmarking and comparative analysis of multiple architectures using the PlantVillage dataset. The framework integrates pretrained models, including DenseNet121, MobileNetV2, VGG16, and DeiT. To enhance generalization and robustness, the proposed system incorporates dynamic architectural adaptation, dropout regularization, and 5-fold cross-validation. Experimental results demonstrate that DenseNet121 achieved the best performance with a weighted precision of 0.989, recall of 0.987, F1-score of 0.988, and an overall accuracy of 99.95%. Furthermore, the four benchmarked pretrained models (DenseNet121, MobileNetV2, VGG16, and DeiT) attained a mean cross-validation accuracy of 99.41% (standard deviation 0.815) across these four models, indicating stable and competitive classification performance across diverse plant disease categories. The proposed framework emphasizes reproducibility and extensibility, providing a strong foundation for future integration of Explainable AI (XAI) techniques such as Grad-CAM and deployment through accessible platforms like Streamlit.

Background: Accurate classification and segmentation of brain tumors from Magnetic Resonance Imaging (MRI) are essential for early diagnosis, treatment planning, and clinical decision-making. However, manual interpretation of MRI scans is time-consuming and prone to inter-observer variability. Deep learning techniques have emerged as effective solutions for automating brain tumor analysis with improved accuracy and reliability. Methods: This study proposes a unified deep learning framework that integrates ensemble-based brain tumor classification, U-Net-based tumor segmentation, and Grad-CAM explainability for automated brain MRI analysis. A soft-voting ensemble of DenseNet121 and InceptionV4 was developed for multi-class tumor classification, while a U-Net architecture was employed for tumor segmentation. The classification and segmentation models were evaluated on publicly available brain tumor MRI and BraTS 2020 datasets using standard performance metrics. Results: The proposed ensemble classification model achieved an accuracy of 98.52%, outperforming the individual classification models, while the U-Net segmentation model achieved a Dice score of 0.8423 and an Intersection over Union (IoU) of 0.7497. Grad-CAM visualisations further improved model interpretability by highlighting clinically relevant tumor regions. Conclusion: The proposed framework demonstrates the effectiveness of integrating ensemble classification, tumor segmentation, and explainable artificial intelligence into a unified diagnostic pipeline. These findings highlight the potential of deep learning-based computer-aided diagnosis systems to support accurate, reliable, and interpretable brain tumor diagnosis and clinical decision-making. The source code is publicly available on GitHub at: https://github.com/Rashmitha-R-Nayak/brain-tumor-project .

The accurate classification of underwater fish species is important for biodiver-sity monitoring, sustainable fisheries management, and aquaculture. However, underwater fish classification remains challenging due to class imbalance, varying imaging conditions, and high visual similarity among species. This study presents a comprehensive evaluation of deep learning architectures, loss functions, and ensemble strategies for underwater fish species classification. Using the Mark Daniel Lampa Kaggle Fish Dataset comprising 13,304 images across 31 species, duplicate images were removed using a combination of MD5 hashing and perceptual hashing, resulting in a cleaned dataset for model training and evaluation. EfficientNet-B0, EfficientNetV2-S, Swin-Tiny Transformer, and ConvNeXt-Tiny were trained using both class-weighted cross-entropy loss and focal loss. The experimental results indicate that class-weighted cross-entropy loss consistently provided stronger performance for individual models, whereas focal loss produced a marginal improvement when used with the soft-voting ensemble. Among the individual models, EfficientNet-B0 achieved the highest accuracy of 94.81%. Ensemble learning further improved performance, with the four-model soft-voting ensemble achieving a test accuracy of 97.21% and a macro-F1 score of 0.9678 on the cleaned Kaggle dataset. An external dataset of 310 internet-sourced fish images was additionally used to assess model generalization under domain-shift conditions, where the ensemble achieved an accuracy of 76.39% and a macro-F1 score of 0.7335. While performance decreased substantially under domain shift, the results highlight the challenges of generalizing fish-species classification models beyond the training distribution. Fish species classification represents an important component of future end-to-end underwater monitoring systems incorporating fish detection and localization modules. The study provides a benchmark for future research on underwater fish-species classification, ensemble learning, and real-world deployment. The source code is openly available on GitHub and permanently archived on Zenodo (DOI: https://doi.org/10.5281/zenodo.19315919 ).

Low-resolution image acquisition remains a fundamental bottleneck in real-world vision systems, including surveillance, UAV monitoring, autonomous navigation, and medical imaging, where degraded inputs significantly impair object detection and semantic annotation accuracy. While modern deep learning-based super-resolution (SR) techniques have achieved high reconstruction fidelity in terms of PSNR and SSIM, most existing approaches optimize primarily for perceptual quality without explicitly validating improvements in downstream detection performance. Consequently, a gap persists between low-level image enhancement and high-level semantic reliability. To bridge this gap, we propose EMAF-Net, an Edge-Aware MAXIM Attention Fusion Network that integrates edge-aware SRCNN enrichment, MAXIM-based denoising, lightweight squeeze-and-excitation attention, and multi-scale feature fusion within an adaptive training framework. The architecture is explicitly designed to preserve structural boundaries and semantic consistency required for robust object detection. Unlike conventional SR pipelines, EMAF-Net evaluates enhancement quality not only through reconstruction metrics but also through annotation performance using YOLOv8 on manually curated object labels. Experimental results on the DIV2K dataset demonstrate that EMAF-Net achieves a peak PSNR of 44.10 dB and SSIM of 0.9965, while improving detection performance on super-resolved images to 0.79 precision, 0.69 recall, and 0.68 mIoU. The proposed framework consistently enhances detection reliability under degraded conditions, demonstrating that task-aware super-resolution can effectively align perceptual reconstruction with semantic objectives. These findings position EMAF-Net as a unified enhancement-annotation framework suitable for deployment in real-time vision systems where both image fidelity and detection robustness are critical.

The exponential growth of digital imagery across online and printed media has intensified the need for reliable authorship verification systems. Ensuring the integrity of embedded watermarks becomes especially challenging under print-and-capture (P&C) conditions, where halftoning, optical blur, illumination variation, and sensor noise jointly distort the embedded signal. To address these challenges, this work presents AURA-Net, a deep learning-based watermarking framework capable of embedding and reliably recovering a fixed 88-bit watermark after full physical reproduction. The system incorporates a dual-head U-Net architecture that jointly predicts the embedded bitstream and estimates a presence confidence score. A confidence threshold ( $\tau = 0.76$ ), together with a bounded retry-decoding mechanism, ensures robust recovery even under difficult lighting and capture conditions. In addition, we introduce a mathematical formulation of the P&C process and an analytical derivation of AURA-Net’s robustness, including a linearized P&C degradation model, latent-space margin separation, likelihood-based presence detection, and cumulative reliability under retry decoding. The model is trained using progressive clean masking and photometric distortions (JPEG compression, Gaussian blur, histogram variation, and synthetic noise) to improve generalization. Experiments on 200 physically captured samples demonstrate that AURA-Net achieves a mean bitwise accuracy of 99.8% in good illumination ( $\mu = 0.801$ ) and maintains a confidence of $\mu = 0.745$ in low-light settings, where recovery is stabilized through retry decoding. Under purely digital attacks, the model yields 40.91-44.32% bitwise accuracy across JPEG compression, blur, salt-and-pepper noise, and speckle noise, reflecting a practical trade-off due to its optimization for physical robustness. High imperceptibility is consistently maintained, with an SSIM of 0.997 and negligible grayscale histogram deviation. AURA-Net executes embedding and extraction in 0.003 s and 0.002 s, respectively, enabling real-time deployment. Overall, the proposed framework demonstrates reliable watermark preservation across both digital and physical domains, establishing a reproducible and analytically grounded foundation for post-print authorship verification.

In recent years, panoptic segmentation has emerged as an important image‐segmentation paradigm, leading to the development of many new algorithms and techniques in the area of computer vision. The segmentation technique not only facilitates the segmentation of individual objects but also makes it possible to detect the regions of the background. Panoptic segmentation has opened new ways and methods for object detection and has diverse applications across various fields, including video surveillance, autonomous driving, medical imaging, crowd counting, and so on. Panoptic segmentation, an integration of instance and semantic segmentation, facilitates adaptability in various domains where precise object delineation is important for system performance and decision‐making. This review paper provides a detailed overview of panoptic segmentation that outlines its history, performance measures, architectures, and their benchmarks with several datasets as well as the latest developments in this area. At the end, the article discusses the challenges that this domain is encountering and the possible extent of available opportunities for further advancements.