<p>Open-Vocabulary Object Detection (OVD) seeks to detect objects from both seen and unseen categories without requiring exhaustive category-level annotations. While vision-language models like CLIP enable transfer to novel classes via text embeddings, directly applying them to region-level detection tasks suffers from domain gaps and classification biases. This paper proposes a new OVD framework, Open-Vocabulary Object Detection via Prompt Learning and Dual-Branch Classification(PLDB), which addresses these limitations through two key contributions. First, we introduce a prompt learning module that optimizes a set of learnable vectors to generate task-specific textual embeddings, improving alignment with region-level visual features. Second, we design a dual-branch classification head that decouples the learning objectives for base and novel categories: one branch is trained with supervised classification on base classes, while the other leverages CLIP-based text–region similarity to generalize to novel categories. During inference, both branches jointly contribute to the prediction of all categories, and their outputs are fused via geometric averaging to balance precision and generalization. Experiments conducted on the COCO, LVIS, and Objects365 datasets demonstrate that, when compared with advanced methods, the PLDB method achieves remarkable performance in the metrics of the open-vocabulary object detection task.</p>