The Few-Shot YOLOv12 Object Detection Method Based on Dual-Path Cosine Feature Alignment

In this study, we propose a new low-shot object detection algorithm, DPCA-YOLOv12, designed to address key challenges such as high annotation costs, class imbalance, and incremental learning. By leveraging adaptive feature alignment and cosine optimization techniques, the model significantly enhances detection accuracy and recall while maintaining real-time performance, particularly under class imbalance conditions. Experimental results demonstrate the effectiveness of DPCA-YOLOv12, showing that the dual-channel architecture notably improves few-shot detection performance, which holds significant implications for both future research and practical applications.