A dataset for fine-grained seed recognition | Scientific Data - Nature

Summary

The article evaluates advanced computer vision techniques, focusing on machine learning methods for seed identification using a dataset called LZUPSD. It explores training neural networks with different architectures, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), optimizing hyperparameters for better performance, and comparing accuracy and precision using metrics like accuracy, precision, and recall. Pretrained models on ImageNet enhance results, with accuracy significantly improved by attention models and lightweight networks offering promising results in agricultural applications. The study also visualizes network decisions using techniques like Grad-CAM, highlighting the importance of seed shape, color, and texture for accurate classification.