Enhancing Visual Perception A Deep Learning Approach to Object Detection and Segmentation in Real-World Environments
DOI:
https://doi.org/10.66021/Keywords:
Object Detection, Image Segmentation, Yolov8, Mask Dino, Dilated Convolutions, Vision Transformers, Mean Average Precision, Intersection Over Union, Real-Time Inference, Deep Neural NetworksAbstract
Background: Object detection and segmentation in unconstrained natural scenes remain among the most persistent challenges in computer vision. Despite significant advances in deep learning, performance under real-world conditions involving partial occlusion, cluttered backgrounds, and inconsistent illumination continues to demand further investigation.
Aim: This study systematically evaluates and benchmarks three state-of-the-art deep learning architectures, YOLOv8, Mask DINO, and Dilate Former with respect to detection precision, segmentation quality, computational efficiency, and robustness against visual degradation in complex real-world environments.
Method: A structured comparative experimental design was employed across three benchmark datasets: COCO, ADE20K, and Cityscapes, with 5,000 randomly sampled images from each. All models were trained under an identical pipeline incorporating data augmentation, model-specific hyperparameter tuning, and optimization via SGD or Adam algorithms. Evaluation covered five dimensions: mean Average Precision (mAP), mean Intersection over Union (mIoU), Panoptic Quality (PQ), per-frame inference latency, and occlusion robustness.
Results: Dilate Former outperformed competing models in segmentation, achieving an average mIoU of 0.83, PQ of 0.81, with less than 3% occlusion degradation and ±3% cross-dataset consistency. Mask DINO led in detection with the highest average mAP of 0.75 and mIoU of 0.77. YOLOv8 demonstrated the fastest inference at 18ms per frame, while recording an average mAP of 0.72 and mIoU of 0.66, with approximately 7% occlusion sensitivity.
Conclusion: Hybrid CNN-transformer architectures offer significant advantages for real-world visual perception tasks. Dilate Former suits precision-critical applications, while YOLOv8 is ideal for real-time deployments. Future work should explore multimodal fusion and edge-optimized lightweight models.
