Automatic music transcription (AMT) is still a major challenge in music information retrieval. Polyphonic audio is a mix of several overlapping frequencies, making it challenging for the usual monophonic pitch estimators. Therefore, models like CREPE perform well on single voice audio. But when they encounter polyphonic mixes or low SNR signals, they perform poorly. This paper addresses the stated problem with a modular hybrid framework to make AMT more robust. It uses neural source separation along with a dual-branch feature-guided gating mechanism. The pipeline begins with demixing the polyphonic audio into different constituent stems vocal, bass and other instruments. This is achieved using Demucs a state-of-the-art source separation model. These stems simultaneously pass through two branches one after another, with one branch being a pitch-tracking branch that runs torchcrepe to get candidate F0 contours, while the other pulls out a voicing mask from high-resolution Mel-spectrograms and chroma-energy features. The fusion of pitch estimates from torchcrepe equipped branch with voicing mask from the other branch helps eliminate false activations like background noise. Also, since all the stems go through the above steps, ultimately, we obtain the MIDI representations of all the instruments and voices involved in the audio. This approach is tested on Annotated-VocalSet and Vocadito datasets and found that there is significant reduction on false alarms caused by noise and a boost in raw pitch accuracy resulting in refined multi-track MIDI representation.

One of the new photovoltaic solutions is the dye-sensitized solar cells (DSSCs) which have been known to be cheap to produce, easy to fabricate and can be tailored to meet desired performance by efficient design of the photoanode. The total power conversion efficiency of DSSCs is mainly controlled by the structural and physicochemical properties of the photoanode since it has a direct impact on the adsorption of dyes, electron transport, and recombination processes. To make the photoanode layer in the current work, titanium dioxide (TiO2) -zinc oxide (ZnO) one-dimensional (1D) composite nanofibers were prepared through the electrospinning method and deposited on fluorine-doped tin oxide (FTO) substrates. The power conversion efficiency (PCE) of the solar cells was 0.286% with an open-circuit voltage (Voc) of 0.61 V, short-circuit current density (Jsc) of 0.09 mA/cm 2 and a fill factor (FF) of 49. EIS analysis showed electron lifetime of 10 ms, which means that charge transport is moderately effective. The integration of ZnO in the TiO2 framework increased the speed of the electron movement and the capacity of the adsorbents to take the dye more than the pure TiO2 nanostructures. This paper has shown that it is possible to apply the electrospun TiO2 - ZnO nanofiber composites as photoanodes in DSSC using a cost-effective method that can be scaled up. The 1D nanofiber structure offers facilitated routes to dye and higher loading of dye and has potential optimization. The innovation of the work is the achievement of the bead-free introduction of TiO2-ZnO composite nanofibers into DSSC photoanodes, which enhanced the transport of electrons and minimized recombination losses that occur in the nanoparticulate system.

Brain tumors must be accurately divided in MRI to help doctors with treatment plans, procedures and keeping track of the tumor’s growth. Since brain tumors are so complex, it is very hard to accurately identify and outline the different parts of these tumors. A enhanced 3D U-Net-based framework for segmenting volumetric MRI data is presented in this study implementing the Nested Skip Connections which are dense and residual. By using an encoder-decoder network, the model can learn information about both context and space from multi-level extracted and reconstructed features. The advantage of this approach is that it works with four separate channels of volumetric data: T1, T1-contrast, T2 and FLAIR MRI. These imaging modalities enable the network to differentiate between various regions within tumors, such as peritumoral fluid, contrast-enhancing tumor areas, and necrotic regions. Valid convolution padding and using Dice loss help improve the architecture by addressing the problem of mostly one class in medical images. To refine the data and concentrate the processing, normalization, skull stripping and tumor region cropping are performed. BraTS data is used to teach and test the model by feeding the algorithm which reflects a significant improvement in creating brain tumor analysis which has revealed strong and reliable results in identifying tumors in various patient scans. The proposed framework results in better brain tumor segmentation analysis and reliable for clinicians in making decisions for ongoing treatment.

Knee osteoarthritis (KOA) is a widespread degenerative joint disease that affects millions globally. While manual interpretation of X-ray images remains common in clinical practice, it is time-consuming and prone to subjective errors. In this study, we propose a hybrid automated diagnostic model for KOA detection using feature extraction from Convolutional Neural Networks (CNNs) combined with traditional feature selection methods. The hybrid Feature Selection (FS) strategy includes Mutual Information (MI), Linear Discriminant Analysis (LDA),Analysis of Variance(ANOVA), and K-means clustering. In addition, advanced deep learning architectures such as ResNet and EfficientNet were incorporated for benchmarking purposes to evaluate the effectiveness of the proposed hybrid model. To enhance the robustness of the proposed model , advance deep learning architectures were also considered for comparative analysis. The model received training and validated using a publicly available X-ray dataset containing 3,815 images obtained from the Kaggle repository. After preprocessing and balancing the dataset, we extracted deep features via a retrained CNN and selected the most informative features using the hybrid (FS) strategy. The proposed model achieved a prediction accuracy of (98%), The novelty of this work lies in integrating deep learning with multiple statistical feature selection techniques within a unified hybrid framework for early KOA detection accuracy and reducing diagnostic workload but also enhances model generalizability and robustness. The study also acknowledges limitations such as dataset imbalance and lack of external validation.

In this work, we developed a mesoporous silica COK-19 based drug delivery system for the sustained and controlled drug delivery of ceftizoxime sodium. Hydrothermal method was employed for the synthesis of mesoporous silica COK-19 matrix and by using post impregnation method drug was incorporated into COK-19. This study demonstrates high drug loading efficiency while maintaining the structure and surface morphology of both the drug and mesoporous silica. The evaluation of incorporation of ceftizoxime sodium within mesoporous COK-19 matrix was done through UV-Visible spectroscopy and drug entrapment efficiency of 79.92% was achieved. A sustained, controlled release of 72.9% was seen when the drug’s in vitro release profile was evaluated in physiologically similar conditions (pH 6.8). This regulated release feature suggests reduced dose frequency and improved patient compliance. Ceftizoxime COK-19 system provides a promising basis for the development of innovative drug delivery system designed to effectively treat chronic and resistant bacterial infections.

Alzheimer disease (AD) is a progressive neurodegenerative disorder that is marked by an irreversible cognitive loss, and it represents a significant burden in the healthcare systems of the world. Structural Magnetic Resonance Imaging (MRI) can be very valuable in terms of anatomical data to detect the presence of a disease-related brain change; but feature redundancy and slight inter-class variation tend to restrict the performance of traditional deep learning models. To overcome such problems, this paper suggests Dynamic Grey Wolf Optimization (DGWO)-powered deep feature fusion framework to classify Alzheimer disease in three classes based on the ADNI MRI data. In the framework, contrast enhancement is implemented by Contrast Limited Adaptive Histogram Equalization (CLAHE), and deep feature extraction is used to obtain complementary multi-scale representations by EfficientNet and InceptionV3. The features extracted are combined and optimized with DGWO, which dynamically balances the exploration and exploitation of redundancy in features and reduces the separability of classes. The selected subset of features is optimized using a better Multi-Layer Perceptron (MLP). As demonstrated by the experimental results, the proposed DGWO-based framework reaches a classification accuracy of 98.61%, with precision, recall, and F1-score all exceeding 98.5%, consistently outperforming GA, GWO, and WOA with notable performance improvements across all evaluation metrics. The ablation analysis confirms the effectiveness of feature fusion and DGWO optimization, while the reported p-values (< 0.05) indicate that the performance improvements are statistically significant. These findings validate the usefulness of dynamic feature optimization to the correct and clinically significant classification of Alzheimer diseases.

Plants are essential to life on Earth, and the diversity of plant species continues to expand. Accurate plant species identification is critical for a wide range of individuals, including foresters, farmers, environmentalists, educators, and urban gardeners. This is especially important in rooftop gardening where the choice of species and space utilization is the most important. Non-expert identification may however be difficult and time-consuming in traditional plant identification because it may need a thorough knowledge of botanical features. This is of particular concern to the urban gardeners who lack formal training in botany. The recent developments in machine learning and computer vision provide promising solutions to the automation of plant species identification to improve the availability of urban agriculture and improve decision-making. Even with these advances, it is problematic to create an all-encompassing system that will be able to detect a sufficiently wide range of plant species. In this research, we suggest a plant species recognition system based on deep learning technology. The framework comprises of four steps; image acquisition, preprocessing, feature extraction and classification. The model was trained on a dataset of 80,000 images of 14 plant species, transfer learning on a ResNet9 architecture, pre-trained on 64,000 images and tested on 16,000. Our model scores around 92 out of the 100 and it can be further optimized to score higher. This piece of work is meant to guide people in the selection and management of plants with specific reference to rooftop gardening and urban agriculture.

The increasing growth in the population demands an urge for a well adequate & optimized model to manage this mass gathering. The paper presents a novel prototype model for crowd counting using the two CNN models that are SD-CNN and Cascade CNN and outline the complete working and comparison of three most optimal types of CNN models for crowd counting that is SD-CNN, Cascade CNN, and Switch CNN. The paper showcases the prototype model where the Cascade CNN takes input images from the SD-CNN to get an accurate result works for high-level prior and density estimation stage. Performance is compared with various CNN models. An exhaustive insight is provided for crowd analysis using machine learning and deep learning approaches. A comparative analysis is represented for multiple types of CNN and Deep Learning models for crowd analysis tasks and lists various datasets available for crowd analysis. Experimental evaluation demonstrates that SD-CNN and Cascade CNN achieve lower error rates compared to other CNN architectures for dense crowd counting tasks. Specifically, the results show that SD-CNN achieves a Mean Square Error (MSE) of 345.6, while Cascade CNN achieves the lowest Mean Absolute Error (MAE) of 322.8 on dense crowd datasets. Furthermore, CSRNet shows superior performance for crowd density estimation when compared with MCNN and Switch CNN on the ShanghaiTech dataset, achieving an MAE of 68.2 and an MSE of 115. The findings indicate that the proposed hybrid framework can effectively improve crowd counting accuracy while reducing computational complexity.

Intrusion detection on encrypted network traffic has already proven to be more difficult with the popularity of encryption protocols like TLS and HTTPS that limit the inspection of packet payloads. Although encryption grants privacy and security of data, it equally facilitates criminal activities to be hidden in the normal flow of communication. The current intrusion detectors have a high dependency on a payload-based analysis or coarse-grained statistics and therefore cannot be utilized effectively in encrypted settings. In order to overcome these weaknesses, the present paper suggests a privacy-sensitive deep learning segmentation architecture, called SecDL-SegNet, to conduct intrusion detection in encrypted traffic. The given model uses flow-level statistical properties, metadata associated with packets, and temporal patterns of communication, without the need to decipher the payloads. An architecture that is based on segmentation is presented to divide the traffic into temporal windows so that a fine-grained analysis and localization of anomalies can be performed. The framework combines convolutional encoding, temporal modeling and computing segment-level anomaly score to improve the detection accuracy and interpretability. The proposed approach is proven to be effective as extensive experiments on the encrypted traffic benchmark datasets have been done. SecDL-SegNet has an average accuracy of 97.17, precision of 96.34, recall of 96.68 and F1-score of 96.59, which is stronger than the traditional machine learning and baseline deep learning predictors. Also, the model has low false positive of 2.18% and can be run in close real-time, with an average inference time of 5.32 ms per flow. The findings indicate that SecDL-SegNet developed a scalable, powerful, and protection conscious network intrusion detection system that works in the next generation encrypted networks.

Over the recent years, Earth’s biosphere is being increasingly contaminated by numerous toxic substances because of rapid production and utilization of hazardous chemicals in industries and agriculture. Industrial pollutants include polychlorinated biphenyls (PCBs), polybrominated diphenyl ethers (PBDEs), short‑chain chlorinated paraffins (SCCPs), and perfluorooctane sulfonic acid (PFOS). Agricultural sources contribute organochlorine pesticides such as aldrin, chlordane, dichloro diphenyl trichloroethane (DDT), dieldrin, endrin, heptachlor, mirex, toxaphene and hexachlorocyclohexane. These contaminants spread through air, water, and soil, inducing carcinogenic impacts, birth defects, immune disorders and reproductive disorders, higher vulnerability towards diseases and impaired cognitive development. Persistent Organic Pollutants (POPs) are specifically a cause distress because of them being carbon‑based, highly toxic, and extremely persistent, enabling long‑range transport via environmental media. Their widespread distribution and ecotoxicological properties pose serious threats to humans, wildlife, and marine life. This review examines POP classification, sources, impacts, and treatment approaches, underlining their significance as a severe environmental and human health concern.