Classification of mushrooms based on AWPF-ResNet18.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mycology

Background:

• Edible mushrooms are popular but harvesting risks accidental poisoning from toxic varieties.
• Accurate mushroom classification is crucial for both consumption safety and cultivation.

Purpose of the Study:

• To develop an advanced AI model for accurate edible mushroom classification.
• To improve upon existing ResNet18 performance in mushroom identification tasks.

Main Methods:

• Developed AWPF-ResNet18, integrating an Adaptive Window Pyramid Fusion (AWPF) module with ResNet18.
• AWPF module enables dynamic multi-scale feature fusion and refines features using a Dynamic Swin Window (DSW) module with variable window sizes.
• The model adaptively focuses on targets of varying sizes, mitigating semantic information loss during downsampling.

Main Results:

• AWPF-ResNet18 demonstrated improved performance over the original ResNet18.
• Key metrics saw increases: Accuracy (+2.5%), Macro Precision (+7.5%), Macro F1-score (+5%), and Macro Recall (+2%).
• The model outperformed current state-of-the-art classification models in various metrics.

Conclusions:

• The AWPF-ResNet18 model offers a robust and effective solution for edible mushroom classification.
• This technology provides practical value for safe mushroom identification and categorization.
• The study validates the effectiveness of the AWPF module in enhancing deep learning models for image classification.