Efficient Hybrid Models for Multiclass Plant Recognition: From Mask-Guided Focusing to Mobile Deployment Optimization

Abstract

Plant identification in natural environments is often hindered by complex background clutter, causing accuracy degradation in lightweight models suitable for mobile deployment. We propose a sequential hybrid framework using a mask-guided focusing strategy to isolate specimens and enhance classification accuracy. By integrating segmentation modules such as DeepLabv3, SegFormer-B4, and Mask2Former with classification backbones, our system concentrates on discriminative morphological features while suppressing environmental noise. Evaluation on BigPlants-100, a new curated dataset of 100 Vietnamese plant species, shows that our approach consistently outperforms standalone baselines. Specifically, the combination of ResNet-50 with DeepLabv3 achieved a 92.7% Macro F1-score, marking a 1.0% improvement at 512×512 resolution. Furthermore, hardware-aware optimizations via TensorRT and ONNX Runtime ensure high-speed mobile inference without compromising accuracy. Our findings demonstrate that mask-guided focusing effectively bridges the gap between efficiency and accuracy, providing a robust solution for real-time biodiversity monitoring.