Detecting and alerting distracted drivers is crucial to prevent traffic accidents. Although numerous studies have been proposed that use deep learning methods to detect driver distraction, most of these approaches rely on single-perspective images, which can lead to reduced accuracy due to occlusions and adverse environmental conditions. To address these limitations, we introduce MELD3, a novel approach that integrates Multi-Task Learning and Ensemble Learning to improve driver distraction detection performance. MELD3 utilizes images captured from multiple perspectives of the driver and employs multi-task learning to generate improved distraction detection results for each viewpoint. These results are then combined using the soft voting ensemble learning technique to produce a more accurate final detection outcome. The experimental results demonstrate that MELD3 achieves a high driver distraction detection accuracy of 96.22%. These findings indicate that MELD3 can significantly improve traffic safety by effectively identifying various distracting behaviors and reducing the risk of distraction-related accidents. Furthermore, its ability to generalize across different scenarios and driving conditions makes it an effective tool for real-world applications, particularly for integration into advanced driver assistance systems, thereby promoting safer driving environments.

With the advancement of information technologies in recent years, it has become common practice to analyze educational data gathered from a variety of sources. Analyzing these education data helps in improving the education system by identifying factors that affect students’ progress and assessing their performance in school. Therefore, studies on predicting students’ academic performance with high accuracy and extracting meaningful models from vast volumes of education data remain of great importance for researchers. In this study, the academic performance of students is predicted using random forest, decision tree, support vector machines, XGBoost, and logistic regression machine learning algorithms with data from Portuguese schools. In order to increase the prediction performance of the developed model, the imbalance in the dataset is eliminated with the SMOTE technique, and the most important features affecting the performance of the students are selected by using the Recursive Feature Elimination method. The results show that the XGBoost algorithm outperforms the research in the literature with an accuracy value of 97,2%.

In recent years, altering and tampering with digital images have become easier with the swift development of internet and computer technologies. Therefore, the use of image authentication in legal cases, digital forensic, and medical imaging has become of paramount importance. In this paper, as a solution to this problem, we propose an MD5 Hash-based blind reversible fragile watermarking method. We divide the input image into non-overlapped 4×4 blocks. Then, the watermark information generated from blocks using the MD5 hash function is embedded in the one-level DWT high and middle frequency sub-bands. Our experimental results demonstrate that the proposed method can provide reversibility for applications such as medical, forensic medicine, and military fields where the original image is essential. It can also provide the ability to resist various malicious attacks such as exchange of content, crop, text addition, copy-paste, rotation, content removal, and noise addition.

Nowadays, tomatoes are one of the most widely cultivated crops globally and domestically. However, insufficient nourishment during the phenological stages of tomato seedlings can negatively impact their productivity. This study examines the accuracy of VGG19, ResNet101, and MobileNetV2 models in predicting the phenological stages of tomato plants. The SVM algorithm is used to classify the features obtained using these architectures, and the performance of the resulting models is evaluated. MobileNetV2+SVM has shown significantly superior performance compared to other models, with an accuracy rate of 98.75%. The MobileNetV2+SVM’s lightweight structure and computational efficiency demonstrate potential for high-accuracy classification even in resource-constrained environments. The characteristics of this model make it well-suited for use in agricultural and robotic applications. The high accuracy rates enable the precise application of the nutrient solution to each tomato seedling’s phenological stage, boosting agricultural productivity.