Accurate detection of small objects in low-resolution images remains a challenge in computer vision, in robotics, healthcare, and satellite imagery applications. This thesis investigates the interaction between deep learning-based super-resolution (SR) and object detection, analyzing how SR influences detection performance and how detection outcomes affect SR processing through two distinct strategies. The first strategy performs object detection after super-resolving low-resolution images. The second strategy first detects objects in low-resolution images and subsequently applies super-resolution to the detected crops. A systematic evaluation compares their effects on each other’s performance. Experimental results on a real-world dataset reveal that the best-performing detection model for identifying target objects is trained on real images, whereas the most effective SRmodel for enhancing image quality is trained on synthetic images. Although SR does not lead to a substantial improvement in overall detection accuracy, it significantly enhances the recognition of small, high-contrast objects, particularly at lower resolutions. However, in certain cases—such as detecting white objects—SR introduces unintended artifacts that negatively affect detection precision. The runtime evaluation on both synthetic and real datasets shows that applying SR before object detection significantly increases processing time. These findings highlight the importance of deep learning-based object detection using YOLO and super-resolution with Real-ESRGAN in optimizing small object detection Inlow-resolution images. While SR improves the detectability of small, high-contrast objects, its impact varies based on object characteristics, influencing refinements in detection pipelines. The use of both real-world and synthetic datasets enables a more comprehensive evaluation of detection performance across different conditions
