Research on Pepper External Quality Detection Based on Transfer Learning Integrated with Convolutional Neural...

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Manual pepper sorting is inefficient and subjective.
• Automated quality detection requires robust and accurate classification models.

Purpose of the Study:

• To develop an efficient and accurate automated model for pepper quality detection and classification.
• To leverage transfer learning and convolutional neural networks (CNNs) for improved performance.

Main Methods:

• Data augmentation techniques (rotation, luminance, contrast adjustments) were applied to the pepper dataset.
• A VGG 16 model was fine-tuned using transfer learning, with optimized parameters (dropout 0.3, learning rate 0.000001, batch normalization, ReLU activation).
• Comparative analysis was conducted against ResNet50, MobileNet V2, and GoogLeNet models.

Main Results:

• The fine-tuned VGG 16 model achieved a prediction precision of 98.14% and a loss rate of 0.0669.
• The VGG 16 model demonstrated superior prediction performance, faster convergence, and greater stability compared to other models.
• The model showed strong generalization and fitting capabilities.

Conclusions:

• The proposed transfer learning-based VGG 16 model is highly feasible for automated external quality classification of peppers.
• This research provides a valuable technical reference for advancing automatic pepper quality grading systems.