RAPF: A New Open-Domain Plant Segmentation Framework Integrating Perception and Reasoning

The RAPF (Reasoning-Aware Perceptual Framework) achieves reliable recognition of both known and unknown plant species through CLIP-DINOv2 feature fusion, HQ-SAM mask generation, and Dempster-Shafer evidence reasoning, providing a closed-loop perception-reasoning paradigm for open-domain visual understanding. This framework addresses the problem of overconfident misjudgment in open-set scenarios faced by traditional methods, and enhances model reliability and interpretability by combining modern foundation models with classical reasoning theories.

Wild plant recognition faces problems such as complex natural environments, unstable lighting, and large morphological differences. More importantly, open-set recognition scenarios require models to accurately classify known species and reliably identify unknown samples. Traditional deep learning models are optimized on closed training sets, lack the cognitive expression of "I don't know", and tend to make overconfident misjudgments when encountering out-of-distribution samples.

RAPF adopts a three-stage design: 1. Multimodal feature fusion: Combining CLIP semantic understanding and DINOv2 self-supervised visual representation to capture high-level semantics and fine-grained features; 2. High-quality mask generation: Using HQ-SAM to generate precise object masks, improving edge detail processing; 3. Evidence reasoning: Introducing the Dempster-Shafer mechanism to integrate multiple evidence sources, model uncertainty and unknown states. The framework also uses a closed-loop perception-reasoning structure, where perception and reasoning modules are iteratively refined to dynamically adjust observation strategies by imitating the human cognitive process.

RAPF shows advantages in multiple scenarios: 1. Ecological survey: Automatically analyze field images, label known species and mark suspicious samples, improving the efficiency of biodiversity surveys; 2. Smart agriculture: Distinguish crops from weeds, identify unknown invasive species and trigger manual inspections; 3. Educational popularization: Provide a reliable recognition backend for natural education apps, and honestly inform users when expert help is needed.

RAPF provides a reference for open-domain visual understanding, demonstrating the value of combining modern foundation models (CLIP, DINOv2, SAM) with classical AI theories. The "perception + reasoning" paradigm can be applied to open-domain problems such as wildlife recognition and medical image analysis. For trustworthy AI, explicitly modeling uncertainty can enhance model reliability and interpretability, providing directions for future research.