2024, Vol.29, No.12

Journal of The Korea Society of Computer and Information 2023 KCI Impact Factor : 0.65

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

In this study, we present a comparative analysis of two state-of-the-art object detection models, RT-DETR and YOLOv8, for real-time egg defect detection. As the global demand for eggs continues to rise, ensuring their quality and safety is paramount for mitigating health risks and addressing food safety concerns. The models were employed to analyze transillumination images of eggs and classify them into six distinct categories: Normal, Crack, Blood Spot, Discoloration, Foreign Body, and Deformation. A dataset comprising 11,000 egg images provided by AI-Hub was utilized, with a train-validation-test split ratio of 9:1:1 to conduct the experiments. The performance of the models was assessed based on two key metrics: detection accuracy and inference speed. For RT-DETR, two variants (r18 and r50) were trained, while YOLOv8 involved the training of three variants (nano, medium, and large). The evaluation results indicated that the RT-DETR-r50 model demonstrated superior accuracy, achieving a mean Average Precision (mAP) of 0.852 (mAP@0.5) at an IoU threshold of 0.5 and 0.851 (mAP@0.5:0.95) across multiple IoU thresholds. Conversely, the YOLOv8n model exhibited exceptional speed, with an average inference time of 0.0021 seconds per image. These findings emphasize the trade-off between accuracy and inference speed, providing critical insights into the development of efficient quality control systems for egg production facilities.

The object detection technology in the field of computer vision is utilized to detect aircraft targets using electro-optical systems. Recent advancements in object detection, driven by deep neural networks and hardware accelerators, have led to significant performance improvements, which can also be applied to aircraft target detection. However, for successful integration into weapon systems, it is essential to achieve high detection reliability and real-time processing in on-board environments. This paper presents two strategies to address these challenges. These strategies focus on improving model training by utilizing object shape information and optimizing the model with mixed precision computation. The experimental results demonstrate that the methods improve the performance of the aircraft target detection model and enhance its applicability to weapon systems.

The rapid advancement of generative AI technology has led to its increasing reliance across various industries and everyday applications. However, most generative AI solutions have focused primarily on generating text or 2D images, with relatively little attention paid to 3D model generation. This study conducts a comparative analysis of the performance of three generative AI models ChatGPT, Copilot, and Gemini in generating 3D models. The methodology involved generating scripts for use in the open-source Blender graphics tool, which operates under a GPL license, using each of the three AI models. The performance of these models was then evaluated using metrics such as Accuracy, Recall, Precision, and F1 Score. The results showed that ChatGPT and Copilot outperformed Gemini, which exhibited lower performance. This discrepancy in performance is likely due to differences in the training data related to 3D model generation. This study demonstrates that the performance of generative AI models varies significantly depending on the domain of application, highlighting the need for further development and upgrades to improve their effectiveness in specific areas like 3D modeling.