Quasi Real-Time Apple Defect Segmentation Using Deep Learning.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Automated defect segmentation is crucial for apple quality control and food safety in agriculture.
• Existing methods often lack the accuracy or efficiency required for real-time applications.

Purpose of the Study:

• To develop a deep learning model for accurate and efficient automated segmentation of apple defects.
• To enhance the model's performance and applicability using data synthesis and exploring RGB image inputs.

Main Methods:

• A novel convolutional neural network (CNN) with a U-shaped architecture and targeted skip-connections was employed.
• An ad-hoc data synthesis technique was developed to augment the dataset and mitigate overfitting.
• The model was evaluated on multi-spectral apple images and compared with general-purpose segmentation architectures.

Main Results:

• The proposed model significantly outperformed existing general-purpose deep learning architectures in segmentation accuracy.
• The method demonstrated high computational efficiency, enabling real-time (GPU) and quasi-real-time (CPU) visual inspection.
• Using only RGB images yielded accuracy nearly comparable to multi-spectral images, enhancing practical applicability.

Conclusions:

• The developed deep learning approach offers a superior solution for automated apple defect segmentation in agricultural settings.
• The model's real-time capabilities and adaptability to RGB imagery make it suitable for practical, large-scale visual inspection systems.
• This research advances automated quality control in the fruit industry, ensuring better food safety and product quality.