Meta-YOLOv8: multi-scale few-shot object detection for Chinese medicinal decoction pieces
摘要
Recently, deep learning has been widely used in the field of Chinese Medicinal Decoction Pieces (CMDP) detection to address the limitations of inefficiency and subjectivity of professionals. However, deep learning networks typically require a large amount of data for training, and it is difficult to collect a large amount of labeled data for thousands of CMDP. This study aims to tackle the challenges posed by limited data availability in the context of CMDP detection. We propose Meta-YOLOv8, a novel few-shot object detection network based on YOLOv8. To effectively integrate YOLOv8 with meta-learning, we introduce three key modules: (i) Multi-Scale Class Feature Extraction Module (CFEM), (ii) Heterogeneous Graph Convolutional Networks (HGCN), and (iii) Multi-Scale Classification Auxiliary Module (CAM). The CFEM extracts high-quality class features at each scale through Task Alignment Learning. The HGCN facilitates context information enhancement and cross-sample information sharing via intra-scale subgraphs and learns intrinsic connections across scales using inter-scale subgraphs to create scale-specific class prototypes. Finally, the CAM leverages these scale-specific class prototypes to support multi-scale classification. We tested the proposed Meta-YOLOv8 model on two common image datasets, VOC and COCO, as well as a self-built CMDP image dataset. Compared to other few-shot object detection approaches, our proposed model achieves competitive results. Compared to YOLOv8, our method improves the mean average precision (