Related Experiment Videos
Benchmarking hybrid CNN and transformer backbones with graph convolution networks (GCN) for flower growth-stage
Aritra Das1, Karib Shams1, Mohammad Rifat Ahmmad Rashid2
1East West University, A, 2 Jahurul Islam Ave, Dhaka, 1212, Dhaka, Bangladesh.
Scientific Reports
|June 13, 2026
Summary
This study introduces a hybrid deep learning framework for precise flower growth stage classification, achieving 97% accuracy. The approach combines Convolutional Neural Networks (CNNs) and Transformers with Graph Convolutional Networks (GCNs) for improved plant phenotyping.
Area of Science:
- Computer Vision
- Plant Science
- Machine Learning
Background:
- Accurate flower growth stage recognition is crucial for plant phenotyping but is hindered by subtle visual differences and limited labeled data.
- Existing methods struggle with the fine-grained classification required for detailed plant analysis.
Purpose of the Study:
- To develop a robust hybrid framework for fine-grained flower growth-stage classification.
- To introduce and utilize the novel BD Flower Growth dataset for training and evaluation.
- To investigate the impact of Graph Convolutional Networks (GCNs) and graph configurations on classification performance.
Main Methods:
- A hybrid framework integrating deep backbone networks (CNNs/Transformers) with Graph Convolutional Networks (GCNs) was proposed.
- A new dataset, BD Flower Growth, comprising 3,889 images across eight species and three growth stages (early, mid, full), was created.
- Ablation studies systematically evaluated GCN depth, node resolution, and graph construction methods (grid-based vs. KNN).
Main Results:
- The hybrid CNN/Transformer + GCN framework achieved a maximum accuracy of 97% using EfficientNetV2 and DenseNet201 backbones.
- The Swin Transformer model demonstrated significant improvement, increasing accuracy from 84% to 97% after GCN integration.
- Grid-based graph construction (4- and 8-neighbour) outperformed KNN, and moderate GCN depth (3-5 layers) provided optimal performance.
Conclusions:
- Hybrid graph-based learning effectively enhances fine-grained flower growth-stage classification.
- Graph configuration plays a critical role in achieving high accuracy and stability.
- The proposed framework shows strong generalization capabilities, validated on the Oxford 102 Flower dataset.