基於任務整合與分割策略的YOLOv7全景分割系統

Abstract

本研究針對全景分割任務提出了一種新穎且高效的方法，全景分割任務旨在精確區分圖像中的所有前景與背景類別，並辨識同類別中的不同個體。現有方法在邊界預測精度不佳與重複預測問題上仍存在諸多挑戰，且許多先進方法依賴於龐大的網路架構，需大量運算資源，難以滿足實際應用需求。基於YOLOv7與FastInst的架構，本研究提出三項核心改進策略：(1) Tasks Integration，透過整合多任務的方法，解決傳統CNN-based方法邊界預測不精確問題；(2) Segmentation-based Proposal Strategy，有效避免Query-based架構中因冗餘proposals導致的重複預測；(3) Segmentation-based Intra and Counterfactual Loss，提升特徵的一致性與鑑別性，同時排除潛在誤導性特徵的影響。實驗結果表明，提出的方法顯著提升了模型的預測品質，為全景分割任務提供了一種兼具精度與效率的解決方案。

This study presents a novel and efficient framework for panoptic segmentation, a task aimed at accurately delineating all foreground and background categories in an image while distinguishing individual instances within the same category. Current approaches face persistent challenges, including imprecise boundary predictions and redundant proposals resulting in duplicate predictions. Moreover, many state-of-the-art methods rely on resource-intensive network architectures, making them less practical for real-world applications. Building on the architectures of YOLOv7 and FastInst, this research introduces three core advancements: (1) Tasks Integration, which unifies multi-task learning to address boundary prediction inaccuracies inherent to traditional CNN-based methods; (2) Segmentation-based Proposal Strategy, which effectively mitigates duplicate predictions by addressing redundancy in Query-based architectures; and (3) Segmentation-based Intra and Counterfactual Loss, which enhances feature consistency and discriminability while suppressing the influence of misleading features. Experimental evaluations demonstrate that the proposed methodology achieves substantial improvements in prediction quality, offering a robust and efficient solution for panoptic segmentation tasks.