Semantic-Guided Object Detection via Visual-Language Alignment

Abstract

Object detection in complex environments is challenged by background interference, scale variations, and insufficient semantic representation. To address these issues, CVM-YOLO is proposed as a semantic-guided vision-language alignment framework built upon YOLO-World, in which a CAMVM backbone is integrated for efficient long-range dependency modeling, an MBAM module is incorporated for enhanced feature representation, a TE-CLIP module is employed for fine-grained visual-language alignment, and an MBRepVL-PAN structure is adopted for multi-scale feature fusion. Experimental results show that CVM-YOLO achieves an mAP@0.5 of 94.3% and a Recall of 88.9%, outperforming the baseline by 3.1% and 4.1%, respectively. Meanwhile, the parameter count and computational complexity are reduced by 23% and 43%. These results demonstrate that the proposed framework effectively improves detection accuracy, robustness, and efficiency through semantic-guided vision-language learning.