Object detection in images using deep neural networks and synthetic data in scenarios of partial object occlusion

This research addresses the problem of automatic object detection in images under limited-visibility conditions, where objects are partially occluded, the background is complex, and lighting and viewpoints vary widely. The proposed approach combines pretraining on a programmatically generated synthetic dataset of 18,000 images - produced using the Visualization Toolkit (VTK) library - with fine-tuning on a compact real-image dataset of 2,000 annotated photographs (500 per class). Six deep neural network architectures - Faster R-CNN ResNet-50 FPN, SSD MobileNet V3, YOLOv11n, EfficientDet-D7, DETR-DC5, and CenterNet- were evaluated across three training regimes: synthetic-only, real-only, and combined (90% synthetic / 10% real). Hybrid training yielded the most substantial improvements: YOLOv11n achieved mAP@0.5 = 0.91 and mAP@0.75 = 0.86 (Precision = 0.89, Recall = 0.90, F1 = 0.89, 82 FPS), compared to 0.79 (synthetic-only) and 0.78 (real-only), representing a gain of up to +15 percentage points in mAP@0.5. EfficientDet-D7 reached mAP@0.5 = 0.87 and mAP@0.75 = 0.81, while CenterNet achieved mAP@0.5 = 0.88 at 35 FPS. Robustness analysis under simulated occlusion demonstrated that hybrid-trained models maintain reliable detection even under severe conditions: YOLOv11n retained mAP@0.5 = 0.78 at 50% occlusion and mAP@0.5 = 0.65 at 25% object visibility, while the degradation in mAP under 75% occlusion did not exceed 20% of the baseline level. The results confirm the viability of synthetic data as a standalone pretraining resource and validate the proposed hybrid pipeline for applications in autonomous driving, video surveillance, and industrial inspection.