On the Efficacy of Handcrafted and Deep Features for Seed Image Classification.

Area of Science:

• Plant Science
• Computer Vision
• Machine Learning

Background:

• Seed analysis offers insights into plant evolution, agriculture history, domestication, and ancient diets.
• Computer vision applications are increasingly vital in agriculture and plant sciences.
• Accurate seed classification is crucial for understanding plant diversity and history.

Purpose of the Study:

• To conduct a comprehensive comparison of various features for multiclass seed classification.
• To evaluate the effectiveness of traditional machine learning classifiers versus deep learning models (CNNs) for seed identification.
• To assess the performance of handcrafted features against deep features extracted by CNNs.

Main Methods:

• Utilized two public plant seed datasets for family/species classification.
• Optimized five traditional machine learning classifiers with seven categories of handcrafted features.
• Fine-tuned established Convolutional Neural Networks (CNNs) and the novel SeedNet model.

Main Results:

• CNN features proved suitable and representative for multiclass seed classification scenarios.
• SeedNet achieved a mean F-measure of at least 96%.
• Ensemble strategy with handcrafted features reached a mean F-measure of at least 90.93%, offering a computationally efficient alternative.

Conclusions:

• Deep features from CNNs are effective for seed classification tasks.
• Handcrafted features, particularly with ensemble methods, provide a competitive and efficient alternative.
• The findings represent a significant step towards automated seed recognition and classification systems.