Volume 12 | Issue 12 | December 2024

 Abstract

The goal of this paper is to identify flower varieties in conjunction with Kaggle Flowers dataset employing Convolutional Neural Networks (CNNs). The dataset contains images of five classes of flowers being, daisy, dandelion, rose, sunflower, and tulip. A classifier based on CNN was developed based on 4323 images and was able to attain a classification accuracy rate of 99.09%. The framework displays an efficient interface where users can upload images of flowers with hopes of obtaining an accurate classification. The model performance was improved by implementing data augmentation and multi-level design of CNN enabling the model to accommodate images variability. This application is useful in areas such as botany, gardening and retail sector because it helps people to use images in identifying and selecting their opportunities faster and more accurately. The project demonstrates the efficiency of CNN networks as applied in image classification and their future use into practice.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Flower,CNN

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Drowsy driving is one of the major causes of road accident and fatalities. This paper proposes a low-cost, a real-time, driver drowsiness detection system through webcam based behavioral analysis and technologically modified software. Facial features are defined to evaluate anthropometric measures such as eye aspect ratio (EAR) and Mouth Creating Ratio (MAR). Adaptive thresholding is used to identify prolonged eye closure or yawning where drowsiness is likely to set in. It deploys a learning of classification Support Vector Machines (SVM) that detects and recognizes images for later use offline achieving 95.58% sensitivity and 100% specificity. Such a system is non-obtrusive and cost effective thereby reducing the risk of drowsy driving. Future work should involve embedding into car systems and evaluation based on the actual driving experience.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

MAR, Drowsy, SVM

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Children are increasingly becoming vulnerable to cyber harassment and predatory behavior on social media. Hence, this project aims to enhance the online safety of children by building a system which uses machine learning algorithms to detect and combat online harassment. The developed system integrates various supervised learning algorithms namely, support vector machine, random forest, naive bayes, k nearest neighbors and decision tree. Upon analyzing user content, the algorithm seeks possible abuse potential regarding posts and messages. Numerous harassing and non-harassing texts are included in the dataset which is used to create algorithms for prediction of such actions in real time. The system will first send alerts to a designated authority within the cyber cell every time, austere patterns are observed. This way, no time is wasted in the intervention. The system further enables providing a quicker and reasonable approach to tackle the issues that come up with young people by ensuring their safety as much as possible.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

K-Nearest Neighbors, SVM, Decision Tree

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The goal of this paper is to boost the resolution of images using Convolutional Neural Networks (CNN) along with the auto-encoder layers. This method trains the CNN model with pairs of low-resolution and high-resolution images so that it can learn to improve the pixel values of low-quality images. The pixel values of low-intensity pixels are replaced with high-intensity pixels and a high-resolution image is generated. The intention of the system is to provide effective enhanced images that can be used in a variety of fields including medicine, satellite, and surveillance images. We propose our own dataset and code to emphasize the focus on the novelty while ensuring the efficacy. Having conducted a number of tests, the project brings forth enhancement in the quality of the reconstructed image which depicts the superiority of auto-encoders and CNNs as promising tools for further developments in super-resolution imaging technology.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

CNN, Auto-encoder

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The focus of this work relates to the identification and tampering detection of forensic scanners through deep neural production techniques. Tools built on convolution neural networks (CNNs) and the CASIA dataset are employed to determine which type of scanner was used to create an image and locate portions that were edited. Among others, processes involve transforming the photographs into error level analysis (ELA) images in order to accentuate the inconsistencies and training a CNN in a procedure where the CNN architecture is optimized. The system achieves promising levels of performance and is thus appropriate for distinguishing between images that are or are not altered. As demonstrated by the series of experiments, the model was able to withstand different circumstances and enabled accuracy of over 82% during validation. This work has a considerable contribution to automated media forensics since it solves the problem of scanner identification and digital manipulation detection in an effective and scalable manner.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

CNN, ELA, SCANNER

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

This paper presents the design of a multiscale convolutional neural network (MCNN) for the task of human behavior recognition, with the aim of increasing accuracy and reducing complexity of the model. Previous models such as CNN2D and LSTM time series modeling relied on a mean global average but neglected other spatial and depth features resulting in inaccuracies. The MCNN model is based on the concept of space-time interaction and depth-separable convolution modules inserted into a CNN3D model in which both the spatial and temporal information are enhanced. The system was trained and evaluated on the UCI HAR dataset which has six activity labels that were recorded using smartphone's sensors. The evaluation showed that the new model offered a 94% accuracy rate while improving learning complexity. An extended hybrid model comprising of a combination of convolutional neural networks, Gated recurrent units, and bidirectional algorithms produced an accuracy of 96% while using less parameters thus showing tremendous effectiveness and capability for real world tasks.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

MCNN, GRU,CNN2D

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Protection of data in digital communication is important to avoid any tampering especially for information of sensitive nature like financial or personal details. The branch of mathematics known as cryptography is concerned with the processes of encryption and decryption which allows secure communication across open and untrustworthy channels. In this project, auto-associative neural networks based on Hebb's rule are proposed to advance cryptographic processes. The system issues keys for encryption, teaches neural networks to encode-decode data streams, and measures the effectiveness of the system in terms of accuracy and time factors. To prevent data from the growing range of threats, the frameworks formulates protection with neural network flexibility. The system also provides the user with a graphical display so that recording and playback is performed automatically. Such method indicates the prospect of use of neural networks for the improvement of ordinary codes and ciphers, and focuses a number of challenges emerging from the contemporary information security.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Cryptography, encryption, decryption

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

All traditional economic control has been nullified due to the existence of counterfeit currency, everywhere across the globe. Counterfeit detection using Decision Tree, Random Forest, SVM, Naives Bayes, and other approaches have been used by researchers in the recent past to test fake banknotes. A survey of UCI repository showed that some features were missing and a certain amount of data was pre-processed before model training. A number of splits in the training and testing samples were created and the algorithms were assessed and trained on parameters such as accuracy, precision, recall and F1 score. The outcomes have shown the strength of specific models such as Random Forest that almost achieved satisfactory classification accuracy and can be used in real-life applications within banking kiosks and ATMs. The research further highlights with facts, the need for automated systems which can curb financial fraud and the methodologies required to put in place an effective counterfeit detection machine learning system.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Security concerns have arisen in today's society as a result of using facial recognition systems that use deep learning model due to some feature being artificially manipulated. This paper revisits the question of how image manipulation in the form of social media filters for a facial feature or a facial expression impacts the recognition performance of models such as ResNet, MobileNet and InceptionV3. It was found that the recognition accuracy drops from 75% on unaltered images to 50% on altered images, thereby validating the observed weakness across the three architectures. Furthermore, a more advanced version VGG16 demonstrated to be more effective with 85% accuracy on altered images. This paper investigated these problems and proposed GRADCAM for heatmap analysis demonstrating why certain images with known resistive properties cannot be used to repulse such attacks, while also discussing possible countermeasures for protecting biometric systems.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

VGG16, MobileNet, biometric

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Biomedical imaging forgeries pose a challenge to social networks and forensics. It is becoming harder to tell these images apart from the real ones. This project employs a set of machine learning approaches, that CNN uses to solve the problem. A network called LBPNET is built with local binary patterns (LBP) for texture feature extraction. The parameters extracted through the LBP are used to train the CNN which learns the difference between real and fake face images. Advanced image preprocessing and training are integrated into the model in order to ensure high performance in a situation whereby real-time decision making has to occur. This work averts the threat that image forgeries bring to different sectors.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

CNN, LBP, images

  License

Creative Commons Attribution 4.0 and The Open Definition