This study investigates the impact of various image preprocessing techniques on the performance of deep learning models for ship detection using KOMPSAT electro-optical (EO) satellite imagery. A deep learning-based object detection model was trained on annotated EO images processed under different scenarios, including changes in image resolution, brightness normalization (such as 8-bit conversion), color-representation (grayscale versus RGB), and annotation formats (bounding boxes versus rotated boxes). The evaluation focused on how these preprocessing variables affect detection accuracy, precision, and recall. The experimental results demonstrate that image preprocessing significantly influences detection outcomes. In particular, higher image resolution combined with appropriate brightness normalization and the use of rotated bounding boxes led to improved accuracy, especially in complex maritime environments where ships appear at various orientations. These findings suggest that careful selection and optimization of preprocessing steps can enhance the robustness and effectiveness of object detection models applied to EO satellite imagery. Overall, this study provides valuable insights and practical guidelines for improving deep learning-based object detection in remote sensing. The approaches tested can be adapted to different EO satellite platforms, supporting applications such as maritime surveillance, environmental monitoring, and other tasks requiring accurate detection of objects in high-resolution satellite images.



Kewords : 객체탐지, 위성영상, 영상 전처리, 해상도, 선박탐지, 밝기 정규화