Fruit classification using attention-based MobileNetV2 for industrial applications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for fruit classification offer high accuracy but are computationally expensive due to heavy-weight architectures.
• The need for efficient models with reduced storage and training costs is critical for practical applications in the fruit industry.

Purpose of the Study:

• To develop a lightweight deep learning model for fruit classification that maintains high accuracy while reducing computational resource requirements.
• To explore the integration of MobileNetV2 and an attention module for enhanced feature extraction and semantic information capture in fruit images.

Main Methods:

• A novel lightweight deep learning architecture was proposed, utilizing a pre-trained MobileNetV2 model for initial feature extraction.
• An attention module was incorporated to focus on salient semantic information within the extracted features.
• Convolutional and attention features were fused, followed by fully connected and softmax layers for classification, leveraging a transfer learning approach.

Main Results:

• The proposed lightweight model demonstrated superior classification accuracy compared to four recent deep learning methods on three public fruit datasets.
• The model achieved this performance with a significantly smaller number of trainable parameters, indicating greater efficiency.
• Evaluation confirmed the effectiveness of combining MobileNetV2 with an attention mechanism for fruit classification tasks.

Conclusions:

• The developed lightweight deep learning model offers a promising solution for automatic fruit identification and classification.
• Its efficiency and accuracy make it suitable for adoption in fruit growing, retailing, and processing industries.
• This research addresses the need for resource-efficient AI solutions in agricultural technology.