DPENet: A Dual Prototype-Enhanced Network for Few-Shot Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot object detection methods struggle with representative prototype learning.
• Existing approaches are sensitive to background noise and object complexity.
• Current methods fail to effectively use semantic information for prototype generation.

Purpose of the Study:

• To propose a Dual Prototype-Enhancement Network (DPENet) for optimizing object detection prototypes.
• To enhance support feature representation and prototype discriminability.
• To address limitations in existing meta-learning based few-shot object detection.

Main Methods:

• Developed an Object Enhancement Module (OEM) using dynamic hypergraph construction and hypergraph convolution.
• Introduced a Semantic Fusion Perception Module (SFPM) for discriminative prototype generation.
• Integrated weighted intra-class prototypes with text-based semantic embeddings.

Main Results:

• DPENet significantly improves support feature representation by capturing high-order semantic interactions.
• The Object Enhancement Module suppresses background noise and highlights salient object features.
• DPENet demonstrates superior performance compared to existing methods on PASCAL VOC and MS COCO datasets.

Conclusions:

• DPENet effectively optimizes prototypes for few-shot object detection.
• The proposed modules enhance feature representation and prototype discriminability.
• DPENet offers a promising advancement in meta-learning for object detection.