Integrating deep learning with non-destructive thermal imaging for precision guava ripeness determination

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Accurate guava ripeness assessment is crucial for minimizing post-harvest losses and ensuring fruit quality.
• Visual assessment of guava ripeness can be unreliable due to subtle changes in certain varieties.
• Non-destructive methods are needed to overcome the limitations of traditional ripeness evaluation.

Purpose of the Study:

• To develop and evaluate a non-destructive method for determining guava ripeness using thermal imaging.
• To compare the performance of five deep learning models for guava ripeness classification.
• To provide a scalable solution for efficient fruit production and supply chain management.

Main Methods:

• Thermal images of guavas at various ripeness stages were captured.
• Image data underwent pre-processing.
• Five deep learning models (AlexNet, Inception-v3, GoogLeNet, ResNet-50, VGGNet-16) were trained and evaluated.

Main Results:

• VGGNet-16 achieved superior performance in guava ripeness assessment.
• The VGGNet-16 model demonstrated high average precision (0.92), sensitivity (0.93), specificity (0.96), F1-score (0.92), and accuracy (0.92).
• The model achieved these results within a training time of 484 seconds.

Conclusions:

• Thermal imaging combined with deep learning provides a scalable and non-destructive approach for guava ripeness determination.
• This method contributes to reducing agricultural waste and improving efficiency in the fruit supply chain.
• The findings support environmentally friendly practices in agriculture.