Journal of Image and Graphics

YOLOv5 vs. YOLOv8: Performance Benchmarking in Wildfire and Smoke Detection Scenarios

Edmundo Casas — 2024-04-10

This paper provides a thorough analysis and comparison of the YOLOv5 and YOLOv8 models for wildfire and smoke detection, using the Foggia dataset for evaluation. The study examines the small (s), medium (m), and large (l) variants of each architecture and employs various metrics, including recall, precision, F1-score, and mAP@50, to assess performance. Additional considerations such as training and inference times, along with the number of epochs required for optimal recall, are also evaluated to gauge the models' real-world efficiency and effectiveness. Quantitatively, YOLOv5 models generally outperform YOLOv8, with the YOLOv5s variant achieving the highest scores across all metrics. However, visual assessments reveal that YOLOv8 models exhibit similar, and in some cases superior, capabilities, particularly in detecting dark and dense smoke. Training times favor YOLOv5 models, contributing to their efficiency, and their shorter inference times offer advantages for real-time applications. While the 'best model' variants confirm YOLOv5's numerical dominance, YOLOv8's 'best models' also display competitive performance. Future research will explore model evaluation on diverse datasets and hyperparameter optimization to further enhance performance, adaptability, and applicability in various real-world object detection scenarios.

Classification of Water Bodies Using Ensemble of U-Net and Random Forest Algorithm

Vasavi S — 2024-03-08

Water Bodies classification through remote sensing and deep learning techniques is crucial for effective management of water resources. Accurate detection, classification of water bodies helps to understand their distribution and characteristics, which can inform water usage and conservation efforts. Current methods use SVM, pixel-based methods for the classifying the water bodies. But the accuracy is low. The ensemble model which is proposed in this paper uses the U-Net and Random Forest algorithm to detect and classify water bodies. Initially, satellite images with a resolution of 0.5m are obtained. The U-Net model is used to segment the bodies of water. Contours are used to determine the shape features of the water bodies, and the Random Forest Classifier classifies the water bodies as rivers, ponds, lakes, canals, and other water bodies. Following U-Net segmentation, the acquired segments are transformed from raster to vector format. Vector data of these segments will be used to update GIS Maps. The proposed system was evaluated on the Indian dataset, specifically the urban areas of Kolkata, West Bengal and the U-Net model accuracy was 96.44%, with the Random Forest classifier accuracy being 67.01%.

Normalized-UNET segmentation for COVID-19 utilizing an encoder-decoder connection layer block

Mohammed Almukhtar — 2024-03-08

The COVID-19 epidemic has had a huge influence on human lives all around the world. The virus spread quickly and impacted millions of individuals, resulting in a large number of hospitalizations and fatalities. The epidemic has also impacted economics, education, and social connections, among other aspects of life. Coronavirus-generated CT scans have areas of interest (ROIs). The use of a modified Unet model structure to categorize the region of interest at the pixel level is a promising strategy that may increase the accuracy of detecting COVID-19 associated anomalies in CT images. The suggested method seeks to detect and isolate regions of interest in CT scans that show the existence of ground glass opacity, which is frequent in COVID-19 patients. This can assist healthcare practitioners in identifying and monitoring illness development, as well as making treatment decisions. Scale UNet is a strong U-Net design modification that can increase the performance of semantic segmentation tasks. Our model Normalized-UNET uses BN after each convolutional layer to decrease internal covariate shift, which dramatically improves the network's learning efficiency.

MRI Brain Tumor Analysis on Improved VGG-16 and Efficient NetB7 Models

Narasimha Rao Yamarthi — 2024-03-22

In recent years, MRI scanning has been the most rapidly developing field. Concerning the tumor’s size and specifics, diagnosing and classifying brain tumor’s is challenging and time-consuming for radiologists. The growth of abnormal cells in the brain is referred to as a brain tumor. A brain tumor is diagnosed in about 11,700 patients every year. It is estimated that 34% of males and 36% of females will survive five years after being diagnosed with malignant brain or different tumors. This study focuses on meningioma, pituitary, glioma and no tumors’, among the many brain tumors. Deep learning algorithms and machine learning methods were used to create an autonomous classification and segmentation system for brain tumors, significantly improving early detection. Using the Visual Geometry Group (16), parameters were set for training the model that detects brain tumors based on analysis of proposed literature solutions. Simple CNN models such as VGG16 and Efficient NetB7 perform well because they are among the highest-performing models. As a result of the study, quick, efficient, and precise decisions can be made using MRI to detect brain tumors. For this 7022 brain magnetic resonance images were used to train and test this model. According to experimental findings, the suggested differential deep-CNN model could accurately categories MRI pictures of brain tumors, including aberrant and standard images, with a 99.25% accuracy rate.

Musical Note Position and Duration Recognition Model in Optical Music Recognition using Convolutional Neural Network

Douglas Rakasiwi Nugroho — 2024-02-06

This research focuses on developing a comprehensive Optical Music Recognition (OMR) system that allows users to input music sheet image and obtain a sequence of position and duration of all detected musical note in the image. This research divides the position and duration classification models, both of which utilize Convolutional Neural Networks (CNN). The evaluation of the models yielded high performance, with the position recognition model achieving 97.88% accuracy, 97.92% precision, 97.88% recall, and 97.90% F1-score. Similarly, the duration recognition model achieved 99.23% accuracy, 99.24% precision, 99.23% recall, and 99.24% F1-score. This study makes use of template matching to conduct additional experiment on the model that has been made and found that if the dataset did not have enough position variations of musical note dot symbol over the image for each images class, it will affect the performance of the recognition model in the real case although the evaluation of the model shows a high performance over the testing dataset.

Convolution Neural Network Approach for Early Identiﬁcation of Patchouli Leaf Disease in Indonesia

Rustam — 2024-04-10

Indonesia is the largest supplier of patchouli oil in the world market, contributing 80-90%. Most patchouli oil products are exported in the perfume, cosmetics, pharmaceutical, antiseptic, aromatherapy, and insecticide industries. The emergence of patchouli leaf disease significantly reduced the production of wet, dry, oil, and patchouli alcohol. Therefore, selecting patchouli cuttings (seedlings) that are entirely healthy and disease-free is very important to prevent disease transmission from one area to another. In addition, the selection of disease-free seeds is also essential to prevent the use of diseased patchouli plant propagation. So far, the early identification of patchouli plant health is carried out through visual observations by experts using antiviral serum tested in the laboratory. However, this testing process is expensive. The presence of an artificial intelligence system that can identify the presence of diseases in patchouli leaves is the answer to this problem. This study proposes a system for early identification of whether a patchouli leaf is diseased or healthy. Identification was carried out based on patchouli leaf images using a convolutional neural networks (CNN) approach. We proposed a CNN model using three convolution layers, a dense layer, and a dropout layer. We compare the proposed model with well-known models, namely EfficientNetB0, AlexNet, InceptionV3, MobileV2, and VGG16. The results show that the proposed model outperformed five well-known models as a comparison. The proposed model has an accuracy of 0.95. It has been confirmed by predicting the new and different testing data.

6D Hyperchaotic Encryption Model for Ensuring Security to 3D Printed Models and Medical Images

Siju John — 2024-04-10

A hyperchaotic system is a type of dynamical system characterized by exhibiting more than one positive Lyapunov exponent, which indicates strong sensitivity to initial conditions. Hyperchaotic systems have more degrees of freedom and exhibit even more complex and intricate dynamics compared to standard chaotic systems. The security of the encryption scheme depends on the complexity of the hyperchaotic system and the randomness of the secret key. Image encryption using hyperchaotic systems is a powerful technique for protecting the confidentiality and integrity of digital images, and it has applications in a wide range of fields, including military, medical, and commercial imaging. In this work, we implemented a 6D hyperchaotic system for the encryption of the 3D printed model and medical images. The performance evaluation by metrics reveals the robustness of the encryption model in ensuring security.

Super-Rays Grouping Scheme and Novel Coding Architecture for Computational Time Reduction of Graph-Based Light Field Coding

Bach Nguyen Gia — 2024-03-22

Graph-based Light Field coding using the concept of super-rays is powerful to exploit signal redundancy along irregular shapes and achieves good energy compaction, compared to rectangular block -based approaches. However, its main limitation lies in the high time complexity for eigendecomposition of each super-ray local graph, a high number of which can be found in a Light Field when segmented into super-rays. This paper examines a grouping scheme for super-rays in order to reduce the number of eigendecomposition times, and proposes a novel coding architecture to handle the signal residual data arising for each super-ray group, as a trade off to achieve lower computation time. Experimental results have shown to reduce a considerable amount of decoding time for Light Field scenes, despite having a slight increase in the coding bitrates when compared with the original non-grouping super-ray based approach. The proposal also remains to have competitive performance in Rate Distortion in comparison to HEVC-based and JPEG Pleno -based methods.

Blind Steganalysis Method using Image Spectral Density and Differential Histogram Correlative Power Spectral Density

Hafedh Ali Shabat — 2024-01-04

Recent research has demonstrated the success of employing neural networks for the purpose of detecting image tampering. Nevertheless, the utilization of reference-free steganalysis has become increasingly popular as a result of the challenges associated with obtaining an annotated dataset. This dataset is crucial for the classification process using neural networks, which aims to detect and identify instances of tampering. This paper introduces a robust approach to blind steganalysis, utilizing image spectral density and differential histogram correlative power spectral density. The proposed method employed two distinct forms of image data, namely a gray-scale image and true-color image data. The results indicate that the proposed methodology successfully achieved the anticipated outcomes in identifying manipulated images as evidenced by its successful application on the two distinct datasets.

Contribution to an Advanced Clinical Aided tool dedicated to explore ASPECTS Score of Ischemic Stroke

Haifa — 2024-02-15

The Alberta Stroke Program Early CT Score (ASPECTS) is a simple and reliable systematic method used to quantify and explore acute ischemic stroke. It was initially developed to standardize the assessment of the early ischemic changes' extent within the Middle Cerebral Artery (MCA). The ASPECTS assessment has been increasingly incorporated into treatment decision-making and has been used in several randomized clinical trials for endovascular treatment decision-making. The e-ASPECTS software is a tool for the automated use of ASPECTS.

The purpose of this paper is twofold: the first objective is to present an advanced clinical tool that would be useful by neuro-physicians to carefully facilitate the extraction of ASPECTS regions of interest (ROIs) and improve medical diagnosis. Firstly, our proposed method starts with a preprocessing step containing skull bone stripping of CT stroke image, then the ASPECTS regions of the image are segmented and extracted using edge detection technique and thresholding process. The second objective is to propose an automated semi-quantitative method based on non-contrast computed tomography (NCCT), to provide a reference for neuro-physicians in the diagnosis and evaluation of acute ischemic stroke.

The proposed automated ASPECTS method presents an independent predictor for clinical practice and ischemic core judgment and treatment selection.

Localization of Defects on PCBs with Diverse Acquisitions

Sainath Chaithanya Aravalli — 2024-02-16

The printed circuit board (PCB) houses various IC constituents through smaller bond pads, lines, tracks, and a restricted pitch. In manufacturing, abnormalities or defects often arise at any of the advancements of drilling, etching, stripping, etc., which affect the performance and functionality of the same. Hence, PCB inspection is necessary to check the characteristics of the layout in accordance with the design specifications. The defects that surface in any form may mostly fall into soldering pads and copper balancing, and locating the same through human intervention is time-consuming and error-prone. Hence, automated optical inspection (AOI) of PCBs is appropriate. Template matching is widely accepted in image analysis to score similarities, figure out the differences, and recognize and locate them using specific mathematical algorithms. However, template matching is not tolerant of variations in viewpoint, background noise, and non-rigid deformations. The present work of inspection was attained with two units of PCB, which are alike in design but hold different abnormalities each, adding to the fact that one PCB acquired normality and the other registered tilt. The developed model not only discovers the cumulative angle of deviation the PCB object undergoes but also succumbs to a scheme that extracts and spots the location of dissimilarities in known catalogs without bounding boxes in accordance with the description of the co-model. It also reduces the cycle time of inspection and is striking in comparison with the traditional system of assessments.