Transformer-based Seed Segmentation and Germination Prediction

This paper puts forwards a Transformer model for seed germination and seed segmentation prediction. The model addresses the vents between traditional approaches and precision agriculture approaches. As the traditional approaches for seed evaluation often based on simple and subjective image analysis or visual inspection, it leads to an ineffective solution. The proposed framework embraces a hybrid approach by combining a high accuracy Vision Transformer concept for seed segmentation with traditional learning based classifiers. Classifiers like Support Vector Machines and XGBoost for germination predictions based on extracted morphological features. The proposed model has been examined on the LZUPSD dataset having 4,496 high resolution images of different seed types. The experimental results of the proposed model outperform existing CNN-based and single Transformer-based baselines in segmentation accuracy and predictive consistency. Further, the proposed research work improves seed quality automation for farmers. It also offers many seed evaluation methods to scale up their performance by embedding it. The proposed work highlights the transformation based seed generation mechanism and the morphological intuition in agriculture by the domain experts.