Mushroom species classification and implementation based on improved MobileNetV3

Area of Science:

• Computer Science
• Artificial Intelligence
• Botany

Background:

• Current mushroom classification methods struggle with generalization and deployment.
• Need for robust and efficient mushroom identification systems.

Purpose of the Study:

• To systematically compare five models for mushroom species classification.
• To develop and deploy an optimized MobileNetV3 model with novel training strategies.

Main Methods:

• Comparative analysis of Transformer and Convolutional Neural Networks (CNNs).
• Implementation of transfer learning with an Adaptive Hybrid Optimizer (AHO) and dynamic cyclic learning rates.
• Training, validation, and deployment on a custom 3633-image mushroom dataset.

Main Results:

• Achieved 98.13% validation accuracy and 97.98% average test accuracy.
• Demonstrated model stability with minimal validation loss fluctuation (0.0343).
• Balanced recall and F1 scores across classes, indicating robust performance against interclass similarities.

Conclusions:

• The optimized MobileNetV3 model offers high accuracy and stability for mushroom classification.
• The model's successful deployment on PC, Android, and embedded platforms confirms practical applicability.
• Provides a valuable tool for mushroom research, wild picking, and automated sorting, supporting the mushroom industry.