Enhancing Visual Perception A Deep Learning Approach to Object Detection and Segmentation in Real-World Environments

Abstract

Background: Recently, visual recognition in natural scenes becomes a bottleneck problem of computer vision, especially for object detection and segmentation tasks. The emergence of deep learning has seen architectures that can significantly outperform traditional methods for accuracy and generalization, while performance when under occlusion, clutter, or under different lighting conditions is still important.

Aim: This study focused on the precision, robustness, and inference speed for object detection and segmentation on complex real-world scenes of modern deep learning models YOLOv8, Mask DINO, and Dilate Former, and has investigated their effectiveness.

Method: A comparative experiment is conducted with benchmarks of COCO, ADE20K and Cityscapes. All models were learned on the same pipeline from hyperparameter tuning to data augmentation to real time inference tests. We evaluate metrics that includes mAP, mIoU, Panoptic Quality (PQ), latency and occlusion robustness.

Results: Dilate Former obtained state-of-the-art segmentation accuracy with a mean mIoU of 0.82 and better occlusion robustness. The performance of mask DINO was strong in detection, while that of YOLOv8 was the fastest with an inference speed of 18 ms/frame but gradually decreased in performance complex scenarios. Dilate Former also kept good cross-dataset generalization ability, and stayed precisely consistent cross-datasets.

Conclusion: Additionally, hybrid models utilizing both CNN and transformer structures have shown competitive gains for object detection and segmentation in real-world images. YOLOv8 is for speed critical purpose, but Dilate Former is for high precise demanding. Future work may consider multimodal fusion and real-time execution on edge devices.