Аннотация
This paper explores camera-aware data augmentation to improve object detection in digital museum collections. A dataset of 51 exhibits from the Kasteyev State Museum of Arts was created by extracting frames from videos simulating typical visitor behavior. To address the limitations of using a single smartphone camera, five types of augmentations were applied: color profile adjustment, sharpness variation, simulated noise, optical distortion, and dynamic range correction. Each augmented image received only one randomly selected transformation to maintain realism. Four datasets were prepared: two with original images (differing in size), and two with a 1:1 mix of original and augmented images. The lightweight YOLOv11n model was trained and evaluated on these sets using Google Colab with an NVIDIA A100 GPU. Results show a significant improvement in detection accuracy due to augmentation, with mAP@0.5:0.95 increasing from 0.719 to 0.977. This study demonstrates that simulating camera variation through targeted augmentation is an effective strategy for enhancing model generalization, especially in contexts with constrained imaging diversity such as cultural heritage digitization.