Object detection automatically identifies and locates specific objects within images or videos, and plays a critical role in various fields. Among the many approaches, the YOLO model has significantly improved the accuracy and speed of object detection, thereby garnering attention for real-time processing. However, it exhibits limitations in detecting small or densely packed objects. This study introduces YOLO-RACE (You Only Look Once with ReAssembly and Convolutional block attEntion), an improved model based on YOLOv8 by incorporating the CARAFE upsampling technique, ResBlock, and the Convolutional Block Attention Module. YOLOv8 was selected from among various YOLO models because of its ability to deliver high accuracy in complex object detection tasks while maintaining its lightweight architecture and overall efficiency. The proposed model was evaluated using the VisDrone 2019, Pascal VOC, and SKU-110 K datasets. The experimental results demonstrated a precision of 0.419, recall of 0.321, mAP50 of 0.316, mAP50@95 of 0.18, and F1-score of 0.364 on the VisDrone 2019 dataset. On the Pascal VOC dataset, the model achieved a precision of 0.764, recall of 0.699, mAP50 of 0.773, mAP50@95 of 0.559, and F1-score of 0.73. Furthermore, on the SKU-110 K dataset, the model attained a precision of 0.904, recall of 0.841, mAP50 of 0.902, mAP50@95 of 0.575, and F1-score of 0.871. Based on these comparative results, the proposed model achieved superior performance compared with YOLOv8 and its variants, as well as with other state-of-the-art models such as YOLOv9, YOLOv10, YOLOv11, GELAN, and EfficientDet. This code is available at https://github.com/AnBLab-BAE/YOLO-RACE.git.

In the field of computer vision, convolutional neural network (CNN)-based models have demonstrated high accuracy and good generalization performance. However, in semantic segmentation, CNN-based models have a problem—the spatial and global context information is lost owing to a decrease in resolution during feature extraction. High-resolution networks (HRNets) can resolve this problem by keeping high-resolution processing layers parallel. However, information loss still occurs. Therefore, in this study, we propose an HRNet combined with an attention module to address the issue of information loss. The attention module is strategically placed immediately after each convolution to alleviate information loss by emphasizing the information retained at each stage. To achieve this, we employed a squeeze-and-excitation (SE) block as the attention module, which can seamlessly integrate into any model and enhance the performance without imposing significant parameter increases. It emphasizes the spatial and global context information by compressing and recalibrating features through global average pooling (GAP). A performance comparison between the existing HRNet model and the proposed model using various datasets show that the mean class-wise intersection over union (mIoU) and mean pixel accuracy (MeanACC) improved with the proposed model, however, there was a small increase in the number of parameters. With cityscapes dataset, MeanACC decreased by 0.1% with the proposed model compared to the baseline model, but mIoU increased by 0.5%. With the LIP dataset, the MeanACC and mIoU increased by 0.3% and 0.4%, respectively. The mIoU also decreased by 0.1% with the PASCAL Context dataset, whereas the MeanACC increased by 0.7%. Overall, the proposed model showed improved performance compared to the existing model.

Vision Transformer (ViT) has been proposed as a new image recognition method in the field of computer vision. ViT applies a Transformer structure with excellent performance in the field of natural language processing to recognize images. Unlike existing Convolutional Neural Network (CNN) models, ViT can achieve State-Of-The-Art (SOTA) image recognition without inputting Inductive Biases into the model, demonstrating that the Transformer is a useful structure in the field of computer vision. However, ViT requires large datasets such as ImageNet-21K and Joint Foto Tree (JFT) for learning. In addition, it takes a lot of time to train. Moreover, there is a problem that location information is lost by inputting images in patch units. To improve such issues, many models are being proposed. In this paper, a new model is proposed by restructuring the Convolutional neural networks Meet vision Transformers (CMT) model by applying Coordinate Attention Block, a CNN model, to improve problems of the Vision Transformer family of models. The proposed model combines Transformer, which has shown excellent performance in Long Range, and CNN, which has shown excellent performance in Local Feature, to achieve higher performance than existing models. We also compared performance of the proposed model with those of existing models with relatively small datasets such as Canadian Institute For Advanced Research-10 (CIFAR-10), Self-Taught Learning-10 (STL-10), and Tiny-ImageNet to facilitate researchers’ access to the evaluation. Despite being restructured from the smallest CMT-Tiny model, the proposed model showed better accuracy than CMT-Tiny, CMT-XS, CMT-S, and CMT-B models with CIFAR-10, STL-10, and Tiny-ImageNet datasets. The proposed model showed an accuracy of 90.21% with the CIFAR-10 dataset, higher than existing CMT models except for the CMT-S model with an accuracy of 90.6%. It had the lowest loss value of 0.3967. The proposed model is expected to be utilized as a backbone in Object Detection and Segmentation fields in the future.

Deep learning technologies have significantly advanced the field of single-image super-resolution (SISR), yet existing methods often prioritize peak signal-to-noise ratio (PSNR) over visual quality and realism. In this study, we propose NeXtSRGAN, which integrates a ConvNeXt-based discriminator to overcome these limitations and achieve more realistic and high-quality super-resolution (SR) images. NeXtSRGAN enhances image realism through its novel discriminator structure and generator residual block design, leveraging a relativistic discriminator and residual scaling. Experimental results on benchmark datasets demonstrate that NeXtSRGAN surpasses existing methods, including enhanced SRGAN (ESRGAN), with an average PSNR improvement of over 1.58 dB and an SSIM enhancement of over 0.05. Notably, NeXtSRGAN exhibits exceptional performance in facial image SR, as confirmed by the KID-F metric. By focusing on perceptual quality rather than solely PSNR, NeXtSRGAN sets a new standard for image restoration and holds promise for applications in various domains, such as video surveillance, medical imaging, and satellite photos. This code is available at https://github.com/PomKlementieff/NeXtSRGAN.

This paper presents a crowd-density simulation system developed using the Unity engine to predict and prevent crowd-crush accidents. Crowd-crush accidents can occur in various environments where large crowds gather, often leading to significant casualties. Effective tools for risk assessment, training, and decision-making are crucial in preventing such accidents. The proposed simulator aims to provide a realistic virtual environment that accurately models crowd behavior and dynamics. By analyzing crowd density, flow patterns, and bottleneck situations, the simulator enables users to identify potential hazards, evaluate evacuation strategies, and develop proactive crowd management measures. The system incorporates advanced features such as real-time crowd visualization, interactive scenario creation, and data analytics. Validation experiments demonstrate the effectiveness of the simulator in predicting crowd-crush risks and supporting decision-making processes. The proposed solution enhances crowd safety and helps prevent accidents in various public-gathering scenarios.

In the field of natural language processing, artificial intelligence (AI) technology has been utilized to solve various problems, such as text classification, similarity measurement, chatbots, machine translation, and machine reading comprehension. Significant advancements have been made in complex and rule-intensive natural language processing through deep learning, in which machines directly learn patterns. Machine reading comprehension, a natural language processing task, involves machines understanding questions and paragraphs to find answers within paragraphs. In 2019, bidirectional encoder representations from transformers (BERT) and a robust optimized BERT pretraining approach (RoBERTa) were introduced. They were then optimized for pretraining and fine-tuning, resulting in significant advancements. RoBERTa outperformed BERT in terms of training speed and performance by increasing the pretraining data and batch sizes, employing dynamic masking, and eliminating the next sentence prediction task. In RoBERTa, machine reading comprehension involves simultaneously inputting questions and paragraphs. However, this simultaneous input method suffers from the attention separate representation (ASP) problem, in which the attention distribution between the question and the paragraph spreads widely across keywords. This study proposed two methods to address the ASP problem. The existing input format, question-paragraph, was changed to three independent inputs: question, paragraph, and concatenated RoBERTa outputs. The concatenated matrix was then transformed into two matrices, and a machine reading comprehension algorithm using co-attention was proposed. An ablation study was conducted to evaluate and analyze the model’s performance, comprehension, and design efficiency. According to the experimental results, the proposed method improved the EM by 0.9% and F1 by 1.0% compared to the existing methods. Consequently, the learning performance was enhanced through attention concentration and co-attention, and it could demonstrate much better performance compared to the existing models.