Prediction and visualization of moisture content in Tencha drying processes by computer vision and deep learning

Area of Science:

• Agricultural Engineering
• Food Science
• Computer Vision

Background:

• Accurate moisture content monitoring is crucial for Tencha quality.
• Current methods are subjective, destructive, or lack reproducibility.
• Objective, non-destructive methods are needed for Tencha processing.

Purpose of the Study:

• To develop a computer vision and deep learning model for non-destructive moisture content detection in Tencha.
• To compare the effectiveness of different preprocessing techniques and machine learning models.
• To visualize moisture content distribution during the Tencha drying process.

Main Methods:

• Image acquisition of Tencha samples during drying.
• Extraction and preprocessing (MinMax, Z score) of color space components.
• Development and comparison of 1D-CNN, PLS, and BP ANN models for moisture prediction.

Main Results:

• The 1D-CNN model with Z score preprocessing achieved high predictive accuracy (Rp = 0.9548).
• Spatial and temporal distributions of moisture content migration were successfully visualized.
• Computer vision and 1D-CNN demonstrated feasibility for real-time moisture prediction.

Conclusions:

• Computer vision combined with 1D-CNN offers a feasible and accurate method for Tencha moisture prediction.
• This technology provides technical support for intelligent and industrialized Tencha production.
• The findings contribute to improved quality control and process optimization in Tencha manufacturing.