Intelligent Methods In Engineering Sciences

Detection of Emergency Words with Automatic Image Based Lip Reading Method

Beyza Ülkümen — 2024-03-27

Lip reading automation can play a crucial role in ensuring or enhancing security at noisy and large-scale events such as concerts, rallies, public meetings, and more by detecting emergency keywords. In this study, the aim is to automatically detect emergency words from the lip movements of a person using images extracted from silent video frames. To achieve this goal, an original dataset consisting of silent video images in which 8 emergency words were spoken by different 14 speakers was used. The lip regions of the images obtained from the videos in the dataset were labeled through relevant region detection. Labeled data were then evaluated using the SSD (Single Shot MultiBox Detector) deep learning method. Subsequently, subsets of labeled data with 8, 6, and 4 classes were created. The SSD algorithm was evaluated separately for each of these subsets. During the training of the SSD algorithm, weight initialization methods such as 'he,' 'glorot,' and 'narrow-normal' were used, and their performances were compared. Additionally, the SSD algorithm was trained with two different values of the maxepochs parameter, which were 20 and 30, respectively. According to the results, the lowest accuracy value was found for the 8-class subset, with an accuracy of 42% obtained using 20 epochs of training and the 'narrow-normal' weight initialization method. The highest accuracy value was achieved for the 4-class subset, with an accuracy of 76% obtained using the 30 epochs of training and the 'glorot' weight initialization method.

Emotion Detection with Pre-Trained Language Models BERT and ELECTRA Analysis of Turkish Data

Abdulkadir TEPECIK — 2024-03-27

Developments in artificial intelligence have led to positive developments in many fields. Sentiment analysis, one of these areas, has become more applicable with the models and architectures developed. In this study, emotion detection and emotion analysis were performed on the transcribed data of Turkish voice recordings. In the emotion detection phase, after the emotional states (positive, negative, neutral) of the data were detected with BERT and ELECTRA models, which are transformer-based structures, machine learning algorithms were used in the accuracy analysis of these emotional states and the Google Colaboratory platform was used in the application phase. Naive Bayes, Random Forest, Support Vector Machine and Logistic Regression algorithms were used in the accuracy analysis. As a result of the study, both Naive Bayes and Logistic Regression algorithms achieved the best accuracy rate in emotion detection with the BERT model with a rate of 70%. In emotion detection with the ELECTRA model, both Random Forest and Logistic Regression algorithms achieved the best accuracy rate of 72%. BERT and ELECTRA methods are used to provide a better understanding of understanding and classification of emotional content in Turkish texts and contribute to the development of sentiment analysis-based applications. In addition, two Turkish emotion data sets were obtained, and by using more than one method in emotion analysis, our study has been a unique study in the field, allowing the analysis of the study to be done more effectively.

Automatic Classification and Detection of Faulty Packaging using Deep Learning Algorithms: A Study for Industrial Applications

Irem CATAL — 2024-03-28

In order to market the product to the consumer with the correct methods and to increase the reliability and sustainability of the brand in many stages from the production stage to the launch of the product in the national and international environment, to prevent faulty problems that may be encountered, the project will classify the packages with computer vision within the framework of deep learning algorithms and detect faulty packages. Studies have been carried out in this direction with the aim of saving labor and time, reducing the margin of error and increasing efficiency. In the study, a total of 3000 images, 1000 from each class, were used in three classes of fruit juice boxes called "Flawless", "Pressed" and "Stained" to ensure the image distribution ratio according to classes. In the study, training and testing of the model was carried out using the YoloV8 object detection algorithm. In addition, in order to make comparisons, SqueezeNet and IncepptionV3 classification models were trained and tested using images. Values of 99.5% for mAP50 and 97.9% for mAP50-95 were obtained from the YoloV8 model. 100% classification success was achieved from the SqueezeNet model and 99.9% classification success was achieved from the InceptionV3 model. The performance results obtained from the tests of the models were analyzed and evaluated, and then real-time testing was carried out. The accuracy of the study was evaluated by taking real-time images of the juice boxes moving on the conveyor with a camera. It is thought that the system created as a result of the study can be used in the industrial field.

Classification of Diseases in Tomato Leaves Using Deep Learning Methods

Muslüme Beyza YILDIZ — 2024-03-29

The automatic detection of diseases in tomato leaves significantly contributes to tomato production and enables farmers to manage these issues more effectively. Tomatoes are a crucial commercial crop for local markets and exports, representing a significant agricultural sector in our country. Diseases affecting tomato leaves directly influence tomato yield and quality, making early detection and intervention paramount. Our study aims to address tomato losses due to leaf diseases using computer technology. Recently, Convolutional Neural Networks (CNN) have been employed in various fields including agriculture, military, robotics, and medicine for classification, object detection, and segmentation tasks. The integration of computer vision and image processing with deep learning architectures has led to notable advancements in these areas, offering solutions with higher accuracy and reducing human error. In our research, a dataset was created using images of tomato leaf diseases selected from Kaggle. Algorithms such as k-Nearest Neighbors (KNN), Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Neural Networks (NN) were applied using Orange, a data visualization and analysis software. Moreover, a custom algorithm developed in Python demonstrated the highest accuracy. while the highest classification accuracy of classification made with machine learning algorithms was 95.6%, the classification accuracy was achieved about 96% with the developed deep learning model. This system was integrated into an Amazon Web Services (AWS) Lambda function, subsequently utilized in a mobile application developed using Flutter for the UI and Dart for backend, ensuring connectivity with the Lambda function.

Evaluation of Machine Learning and Deep Learning Approaches for Automatic Detection of Eye Diseases

İbrahim KAYA — 2023-03-29

There are many ocular diseases present in the world. These diseases may arise from factors such as genetic predisposition, environmental influences, and aging. In recent years, advancements in technology have facilitated the detection of ocular pathologies through machine learning techniques. Machine learning models can serve as decision support mechanisms in diagnostic scenarios. In this study, the aim is to detect ocular diseases using machine learning and deep learning techniques. To enhance the results obtained from classification with 4,217 images in the study, 705 images were added to the glaucoma class and 370 images were added to the Diabetic Retinopathy class. The supplemented dataset with additional images comprises a total of four classes. One class represents the control group and is labeled as "normal." The remaining three classes represent disease categories: Diabetic Retinopathy, Cataract, and Glaucoma. To extract deep features from the images, a pre-trained InceptionV3 model was utilized, resulting in 2048 features extracted. These extracted features were then classified using Neural Network (NN), Logistic Regression (LR), k-Nearest Neighbors (k-NN), and Random Forest (RF) machine learning models. Before the dataset supplemented with additional images, the classification accuracies of the machine learning models were as NN 89.2%, LR 87.3%, k-NN 81.2%, and Random Forest 76.9%. Upon examining the classification accuracies after dataset supplemented with additional images, the following improvements were observed: NN 90.9% with a 1.7% increase, LR 90.2% with a 2.9% increase, k-NN 84.6% with a 3.4% increase, and Random Forest 82% with a 5.1% increase. Performance evaluation was conducted using recall, precision, and F-1 score metrics. Additionally, the learning performance of the machine learning models was assessed through Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC) values.