Advances in computational intelligence最新文献

This study builds upon the research conducted on the personal color decision system. It employs color space logic and reduces limitations associated with capturing photos, aiming to enhance the existing personal color decision method. The objective is to obtain more reliable and objective results for personal color analysis. Our proposed approach focuses on developing a comprehensive color selection framework by leveraging personal color databases and employing decision tree methods. The findings of this research suggest that utilizing personal color analysis in image creation can assist individuals in cultivating a positive and confident image, which holds significance in interpersonal relationships and social interactions.

The concrete mixture design and mix proportioning procedure, along with its influence on the compressive strength of concrete, is a well-known problem in civil engineering that requires the execution of numerous tests. With the emergence of modern machine learning techniques, the possibility of automating this process has become a reality. However, a significant volume of data is necessary to take advantage of existing models and algorithms. Recent literature presents different datasets, each with its own unique details, for training their models. In this paper, we integrated some of these existing datasets to improve training and, consequently, the models' results. Therefore, using this new dataset, we tested various models for the prediction task. The resulting dataset comprises 2358 records with seven input variables related to the mixture design, while the output represents the compressive strength of concrete. The dataset was subjected to several pre-processing techniques, and afterward, machine learning models, such as regressions, trees, and ensembles, were used to estimate the compressive strength. Some of these methods proved satisfactory for the prediction problem, with the best models achieving a coefficient of determination (R²) above 80%. Furthermore, a website with the trained model was created, allowing professionals in the field to utilize the AI technique in their everyday problem-solving.

This study introduces a web application designed to address the challenge of ensuring correct posture and performance in weightlifting exercises, with a particular focus on fundamental bodyweight movements targeting various body parts. The problem at hand primarily concerns beginners who require guidance for accurate exercise execution. To tackle this issue, the tool leverages a live camera in conjunction with the MediaPipe and OpenCV frameworks to extract key points from the user's body. It concentrates on seven core exercise postures, using these key points to calculate numerical values and angles. Users are required to adjust their view angles to activate the tool's pose estimation functions. An algorithm, based on predefined rules that determine posture thresholds and angles between three key points, is employed to detect incorrect postures, provide real-time feedback, and track repetition counts. The completion of all required stages is necessary to count a repetition as correct. Additionally, in this study, we have expanded the algorithm to include three new exercise postures: Bent over Dumbbell Row, Seated Triceps Press, and Dumbbell Fly. We have also adapted the system to detect the lying down view, which is essential for the Dumbbell Fly posture. The results of testing this application demonstrate further development potential, particularly in enhancing the model’s framework to accommodate challenges such as high light intensity, pale skin tones, and instances when a body part is obscured by an object.

Contemporary font design is a labor-intensive process. To address this, we utilize deep learning, specifically StyleGAN2-ADA and Real-ESRGAN, for automated Thai font generation. StyleGAN2-ADA incorporates adaptive discriminator augmentation (ADA) for image synthesis. By integrating Real-ESRGAN, font quality is enhanced. Our approach produces diverse, high-resolution fonts, as demonstrated in comparative experiments. In a survey with 50 participants, StyleGAN2-ADA without augmentation proves superior in legibility and visual appeal, while StyleGAN2-ADA with augmentation excels in diversity. This research highlights the efficiency of deep learning in creating high-quality Thai fonts and has implications for automated font design advancement.

This study aims to automate the reporting process of house inspections, which enables prospective buyers to make informed decisions. Currently, the inspection report generated by an inspector involves inserting all defect images into a spreadsheet software and manually captioning each image with identified defects. To the best of our knowledge, there are no previous works or datasets that have automated this process. Therefore, this paper proposes a new image captioning dataset for house defect inspection, which is benchmarked with three deep learning-based models. Our models are based on the encoder–decoder architecture where three image encoders (i.e., VGG16, MobileNet, and InceptionV3) and one GRU-based decoder with an additive attention mechanism of Bahdanau are experimented. The experimental results indicate that, despite similar training losses in all models, VGG16 takes the least time to train a model, while MobileNet achieves the highest BLEU-1 to BLEU-4 scores of 0.866, 0.850, 0.823, and 0.728, respectively. However, InceptionV3 is suggested as the optimal model, since it outperforms the others in terms of accurate attention plots and its BLEU scores are comparable to the best scores obtained by MobileNet.

Deep learning has played an important role in many real-life applications, especially in image classification. It is often found that some domain data are highly skewed, i.e., most of the data belongs to a handful of majority classes, and the minority classes only contain small amounts of information. It is important to acknowledge that skewed class distribution poses a significant challenge to machine learning algorithms. Due to which in case of imbalanced data distribution, the majority of machine and deep learning algorithms are not effective or may fail when it is highly imbalanced. In this study, a comprehensive analysis in case of imbalanced dataset is performed by considering deep learning based well known models. In particular, the best feature extractor model is identified and the current trend of latest feature extraction model is investigated. Moreover, to determine the global scientific research on the image classification of imbalanced mushroom dataset, a bibliometric analysis is conducted from 1991 to 2022. In summary, our findings may offer researchers a quick benchmarking reference and alternative approach to assessing trends in imbalanced data distributions in image classification research.

Text generation from charts is a task that involves automatically generating natural language text descriptions of data presented in chart form. This is a useful capability for tasks such as summarizing data for presentation or providing alternative representations of data for accessibility. In this work, we propose a hybrid deep network approach for text generation from table images in an academic format. The input to the model is a table image, which is first processed using Tesseract OCR (optical character recognition) to extract the data. The data are then passed through a Transformer (i.e., T5, K2T) model to generate the final text output. We evaluate the performance of our model on a dataset of academic papers. Results show that our network is able to generate high-quality text descriptions of charts. Specifically, the average BLEU scores are 0.072355 for T5 and 0.037907 for K2T. Our results demonstrate the effectiveness of the hybrid deep network approach for text generation from table images in an academic format.

Supervised learning will be a bottleneck for developing plant disease identification since it relies on learning from massive amounts of carefully labeled images, which is costly and time-consuming. On the contrary, self-supervised learning has succeeded in various image classification tasks; however, it has not been applied broadly in the plant disease analysis process. This work, therefore, studies the effectiveness of self-supervised learning using contrastive pre-training with SimCLR for plant disease image classification. We investigated unsupervised pre-training scenarios on unlabeled plant images across multiple architectures, including supervised fine-tuning on labeled samples. In addition, we explored the label efficiency of the self-supervised approach, acquired by fine-tuning the models on various fractions of labeled images. Our results demonstrated that the performance of self-supervised learning on plant disease became comparable to that of the supervised training approach.

In this paper, we have considered a non-linear mathematical model to study the chaotic situation, arising due to slow process of recruitment, leading to an increase in unemployment. We observed the effects on recruitment process due to delay and without delay. We have also studied the stability of equilibrium points with numerical examples to compare with analytical and theoretical results.

Threat surface area for the Internet of Things is calculated as the sum of security vulnerabilities or the weakness and gaps in protection efforts for the device, operating systems, associated software applications, and the local infrastructure. This aggregates all the known and unknown threats that can potentially expose the device, logs, data, and hosted applications. By reducing the exposed elements of the device surface, the device vulnerabilities can decrease the exposed threat surface area. This research presents a new framework first to map the devices in the ecosystem, measure the potential threat surface area from the exposure indicators for each layer and then determine the threat vectors for device compromise to calculate the maturity and severity levels. The authors propose new metrics to reassess and re-calculate the maturity and severity levels. Based on the new metrics, newly exposed threat surface elements provide a new security perspective beneficial for stakeholders involved in design, implementation, and security ecosystem of smart devices.