Night time traffic monitoring is hindered by low light, motion blur, and occlusions, restricting the performance of vision-based approaches. This article introduces a multimodal CNN-RNN model that combines infrared video and stereo audio for accurate vehicle detection and classification under low-light environments. Leveraging complementary acoustic-visual cues, the approach enhances detection precision for cars, buses, trucks, and motorbikes, including off-screen and partially occluded instances. When tested on actual night-time data, the new approach makes considerable gains over baseline performance. Box-wise IoU is 0.415, a relative increase of 18.2

Holy Basil (Ocimum sanctum), commonly known as Tulsi, holds a sacred and medicinal prominence in traditional healing systems, particularly in Ayurveda. Basil is used to treat respiratory disorders, fever, diabetes, stress, and various infections due to its antimicrobial, anti- inflammatory and immunomodulatory effects. These plants are susceptible to a range of phytopathogenic threats, including fungal infections, bacterial blights, viral diseases, and arthropod infestations. These diseases weaken the plant, reducing its bioactive compounds and lead to complete crop loss. Early detection of these diseases is pivotal to reduce extensive damage. The study investigated the performance of different machine learning as well as deep learning techniques like MobileNetV2 and EfficientNetB0 with SVM, KNN and Random Forest, Transfer learning using EfficientNetB0, DensetNet121, ResNet50 etc. on a dataset of our own with 803 images which are divided into 5 classes namely Leaf spot, Downy mildew, Fusarium wilt, Fungal and Healthy which are identified as a disease which affects basil leaves. Here, we presented a hybrid model of pre-trained CNN model EfficientNetB0 with SVM performing best amongst others. EfficentNetB0 is used to extract feature implicitly unlike ML algorithms whereas SVM is used to classify images into different classes. By data augmentation, we make the dataset balanced enhancing its overall performance compared to traditional CNN. Our hybrid model i.e. EfficientNetB0 + SVM has performed reasonably well with the testing accuracy of 98.86

Cardiovascular disease continues to be the world’s number one killer, highlighting the critical necessity for precise, early diagnosis. 16 clinically significant characteristics are included in the dataset, including demographic, behavioral, and physiological factors that are important for determining the risk of long-term coronary heart disease. By using a modified sigmoid activation function that provides controlled scaling, shifting, and clipping of the input logits, the fundamental methodology focuses on improving the logistic regression model. This change is intended to stop numerical instability, like exponential overflow, which frequently occurs when training on high-magnitude or unbalanced data. By directly altering the activation function within the logistic regression framework, the innovation provides enhanced numerical stability without sacrificing interpretability, which is crucial for clinical applications. Stratified sampling was employed to train and evaluate the model, and standard performance metrics were utilized for its validation. The outcomes indicate that the proposed model improves upon baseline logistic regression (accuracy: 84.74

With the growing volume of vehicles on roads, there is a rising need for intelligent systems that can help in real-time traffic monitoring and enforcement. This paper introduces a practical and efficient solution that combines several techniques—vehicle detection, license plate recognition, and colour identification—under one, unified framework. We employ YOLOv8 for precise object detection, Haar Cascade for license plate identification, and Easy OCR to extract registration numbers from them. For vehicle colour identification, the HSV colour space is employed for enhanced classification under varying lighting. The system records video input via a Tapo C310 camera and processes the frames in real time at 30 frames per second. Each module is standalone, meaning that the system can be easily scaled and is always stable even when it is performing intensive processing. We applied our system on a dataset of 500 images and established that it recorded 98.5

The analysis of medical images is one of the key areas in the diagnosis of many conditions. However, because of the intricacy of medical images and the relatively limited size of many available datasets, traditional machine learning models are usually not as precise. Transfer learning has emerged as an approach to alleviate these problems as it relies on already existing models. This paper explores the latest advancements in transfer learning for medical image analysis, particularly in object detection. Object detection model YOLO (You Only Look Once) and its variants have gained substantial acceptance for their highly efficient architecture and improved accuracy. In this study, YOLOv8-pretrained on COCO dataset and the advanced version in YOLO family- is fine tuned for detection of leukemia cells in microscopic blood images. Model is trained and evaluated on ALL-IDB2 dataset using Google Colab. Data annotation and augmentation are used to improve the model performance. Experimental results are compared with existing state-of-the-art models, demonstrating that YOLOv8 outperform with 99.3% mean Average Precision(mAP). Our proposed model significantly improves accuracy, making real-time detection of leukemia cells more efficient.

Advanced diagnosis and successful treatment of melanoma depend on early detection, especially in light of the significant rise in melanoma cases in recent years. The suggested Active Contour Model (ACM) and machine learning (ML) techniques are used in this research to present a robust dermoscopic lesion image segmentation process for melanoma identification. By developing a strong association between forefront and background entities, the ACM effectively manages high-dimensional lesion image data, improving pixel detection performance and forming sharp edges and lesion pattern recognition. A Gaussian distribution model is used to control changes in pixel intensity. The comprehensive ISIC 2017 challenge dataset is used to assess the performance of the suggested ACM, and the PH2 dataset is used for extra validation. The testing outcomes demonstrate how well the suggested ACM segmented images in terms of segmentation accuracy, dice coefficient, and visual appeal. The abstract, however, may improve by addressing the clinical relevance of the suggested technique, outlining dataset sizes and potential biases, giving particular values for performance metrics selected, and delivering more information on the individual ML algorithms included in the ACM.

Human-animal conflict resolution solutions have been a keen are of interest in recent past. Animals are encountered entering human habitat due to various reasons such as their own habitat loss, climate changes, inadequate food resources, deforestation, etc. Thus, there is a necessity of a smart solution which detects when a wild animal tries to intrude into the surroundings. This paper proposed a method that uses deep learning yolov8m model on images of Tiger, Lion, Black Bear and Leopard to classify and detect them. The proposed system in the paper can detect any of the above-mentioned wild animal using CCTV cameras installed and classify them. The aim of this model is to detect presence of wild animal in human habitat in Realtime. Performance evaluation of the proposed system is done on real-time dataset containing photos of wild animals which were extracted from various videos collected from online resources. Experimental results classify animals into Lion, Leopard, Tiger and Black Bear and predict accurately up to 96.6%. Obtained testing accuracy is higher than the results given by existing models.

Waste management in urban environments has become more complex and challenging as cities have expanded, primarily due to the increased difficulty in the detection and monitoring of waste generation and disposal. Rapid urbanization leads to a rise in population density, informal settlements, and unplanned infrastructure, all of which contribute to inconsistent waste disposal practices and scattered dumping sites. Our research proposes a solution for garbage detection by integrating drone technologies and image recognition through YOLOv8 model. This methodology employs drones fitted with high-resolution cameras positioned over urban centers to capture images in real time. With GPS and GSM modules, locations are tagged with the images and transmitted to the central system enabling smart waste management. This technology accurately sends the location of trash areas. Using the YOLOv8 processor, images are received and garbage is detected and classified accurately and rapidly. The novelty of this solution is the synergy between aerial photography, YOLOv8, and specialized in-classification systems, which makes it possible to obtain high-resolution repeatable results. Experiments have shown that this system is superior to conventional waste detection system and attains 80–90

The process of understanding the underlying subjective information and opinion in a given sentence is sentiment analysis. Accurate Sentiment analysis is possible with the use of Deep-Learning techniques. We have proposed three Deep-Learning based fine-tuned models for the three-class sentiment analysis task. It is complemented with an end-to-end pipeline that includes classical approaches like stemming, lemmatization in the data preprocessing, parameter tuning, and so forth for Marathi texts analysis. The proposed models are based on ANN, IndicBERT, and BiLSTM. The “L3CubeMahaSent'’ dataset was used for both training and testing of the proposed models. The IndicBERT based model achieved a superior F1-score of 85

Military operations call for secure, imperceptible, and reliable communication systems for transmitting highly sensitive data such as geographical co-ordinates. This study proposes a novel hybrid framework combining AES-based encryption and hash-driven multi-image steganography to transmit co-ordinate data over TCP/IP networks securely. Our innovation lies in building an imperceptible, powerful yet computationally efficient and practical system for communicating sensitive location data. With the use of perceptual colour plane splicing to enhance imperceptibility against visual and statistical detection, we eliminate all traditional vulnerabilities in transmission as well as key exchange. To assess the effectiveness of the proposed framework, we evaluated the system through a practical test consisting of more than 100 experiments using images of various resolutions. The proposed model achieved a peak PSNR of 94.76 dB and an average PSNR exceeding 82 dB across all image types, ensuring imperceptibility well above the 36 dB visibility threshold. The mean square error (MSE) was consistently below 0.001, and histogram analysis confirmed visual consistency between cover and stego images. Modern steganographic attack tools (StegExpose, StegDetect) failed to detect hidden data, confirming robustness. Furthermore, the system maintained low transmission overhead (under 300 KB for 100 co-ordinates) and demonstrated resilience against passive attacks and advanced attackers targeting data leaks. Our work thus fills a critical research gap by designing a purpose-built, end-to-end secure transmission system for co-ordinate data, especially suited for real-time military scenarios requiring confidentiality, untraceability, and minimal visual and computational footprint.

Skin cancer, especially melanoma, is the most common and fatal cancer in the world, and early diagnosis is essential for enhancing patient outcomes. This paper introduces a Customized Convolutional Neural Network (CNN) for skin cancer classification structure using the HAM- 10000 dataset which has a large set of annotated dermatoscopic images. We fine tuned proposed model hyperparameters to maximize accuracy for skin cancer dataset. The structure uses several convolutional layers, batch normalization, and dropout methods to maximize feature extraction ability and ensure effective classification. Data augmentation methods of random rotation and flipping were used in combination with a stratified method of splitting the dataset into train and validation sets. The model was able to achieve competitive classification accuracy. The method has vast potential in supporting clinicians by giving consistent and fast automatic analysis of lesions, thus promoting early detection and treatment of skin cancer. Performance evaluation of the proposed model is evaluated on HAM- 10000 skin cancer dataset with 98.8% accuracy which is better than existing approaches.

The significance of plants as a primary energy source for humanity cannot be overstated. However, the susceptibility of plant leaves to diseases at any point between sowing and harvesting poses a significant threat to crop production and market value. Detecting these leaf diseases is pivotal in agriculture, but conventional methods demand substantial manpower, time, and an in-depth understanding of plant diseases. In response to these challenges, machine learning along with image processing emerges as a promising solution for disease detection in plant leaves. Here, we proposed customized deep learning model to thoroughly analyze potato plant leaf data, categorizing it into predefined sets. The classification process considers structural features and properties such as color, intensity, and dimensions of the plant leaves. Leveraging machine learning enables a more streamlined and accurate identification of diseases across various plant species. The proposed model offers a comprehensive overview of potato plant diseases and explores customized deep learning classification techniques applied in disease identification. By synthesizing these approaches, the study contributes to the development of automated systems for potato plant disease detection, addressing the limitations of traditional methodologies. This study stands out for its innovative combination of deep learning method with image processing to tackle the complexities of identifying diseases in potato plant leaves. The proposed model improved the overall detection process, offering a more efficient and dependable solution compared to traditional techniques. The proposed model using CNN gives an accuracy of 99.6

Diabetic retinopathy is one of the common causes of blindness with diabetes. Early diagnosis is important to prevent irreversible vision loss. Conventional methods for diagnosing diabetic retinopathy are often based on manual examination of retinal images, which can be time-consuming and subject to human error. The integration of machine-based automated diagnostic systems offers a promising solution to this challenge. Machine-based automated diagnosis of diabetic retinopathy can prevent vision loss with early detection and treatment. In this study, we investigated the performance of different transfer learning models-DenseNet, EfficientNet, VggNet, and ResNet-on a large dataset called Diabetic retinopathy from Kaggle, consisting of 35,108 retinal images in 5 classes. Out of which 28086 samples were used for training purpose and 7,022 samples for validation testing. While previous research has explored machine learning for retinopathy diagnosis, our research uniquely combines modern transfer learning models and evaluates the effectiveness of specific processing methods with Ben Graham's processing methods. This combination distinguishes us from existing methods by contributing to a significant increase inAaccuracy. In particular the accuracy of the proposed approach is 97.7%, Aour tests show that the diagnostic accuracy increases by about 4-5% when using Ben Graham preprocessing. The results of our research may help develop more accurate and efficient automated systems for diagnosing diabetic retinopathy, thereby improving patient outcomes.

This research paper presents the implementation of two image restoration techniques, namely AHE filter and CLAHE filter, along with an Enhancer using Python for medical imaging applications. Aim of paper is to improve the quality of degraded medical images by reducing noise, blur, and other imaging artifacts, leading to enhanced diagnostic quality. We discuss the advantages and limitations of these techniques and present empirical evidence of their effectiveness in restoring medical images (Hai et al. J Vis Commun Image Representation 90, 1047–3203, 2023). Furthermore, we outline the key challenges and open research problems in the field and propose potential future research directions. Finally, we present real-world examples of the application of image restoration techniques in medical imaging, such as X-rays, and discuss the potential impact of these techniques on clinical diagnosis and treatment. The AHE and CLAHE filters were evaluated on a Pneumonia Classification model, resulting in an accuracy increase from 80.93

There have been significant advancements in object detection in recent years, notably with the emergence of the YOLO approach. In this paper, we proposed a thorough comparison between YOLO-NAS and YOLOv8 specifically for detecting vehicle license plates based on the analysis of experimental results. Real-time vehicle license plate detection can have a wide range of applications like crime prevention, surveillance, traffic-pattern analysis, automated toll collection, etc. Our evaluation uses a diverse dataset capturing vehicle images from various angles and levels of occlusion to ensure a comprehensive assessment of the models' capabilities. The metrics considered include precision, recall, and mAP, offering insights into detection accuracy and false positive rates. The results reveal nuanced performance differences between YOLO-NAS and YOLOv8 in vehicle license plate detection tasks. YOLO-NAS demonstrates superior accuracy in specific scenarios due to its optimized architecture. Conversely, YOLOv8 shows notable efficiency gains, especially regarding inference speed, making it a compelling choice for real-time applications. Here, we contribute valuable insights to the ongoing discussion on object detection methodologies, assisting practitioners in making informed decisions when selecting a model for vehicle license plate detection applications. The findings underscore the importance of considering trade-offs between accuracy and computational efficiency in specific use cases

XAI is critical for establishing trust and enabling the appropriate development of machine learning models. By offering transparency into how these models make judgements, XAI enables researchers and users to uncover potential biases, admit limits, and eventually enhance the fairness and dependability of AI systems. In this paper, we demonstrates two techniques, LIME and SHAP, used to improve the interpretability of machine learning models. Assessing Explainable AI (XAI) approaches is critical in searching for transparent and interpretable artificial intelligence (AI) models. Explainable AI (XAI) approaches are designed to provide insight into how complex models make decisions. This paper thoroughly analyzes two prominent XAI methods: Shapley Additive explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). This study aims to understand the decision made by a machine learning model and how the model came to that decision. We discuss the approaches and framework of both LIME and SHAP and assess their behavior in predicting the model's outcome.