Articles
X-ray radiography and deep learning for internal quality inspection of pear fruit
Article number
1403_16
Pages
133 – 140
Language
English
Abstract
In the fruit and vegetable industry, quality assurance is an important issue.
Besides external quality, also internal quality should be guaranteed to consumers.
Therefore, this work proposes X-ray radiography imaging, which is a fast, accurate and cheap technology which can be implemented in an inline system, and it gives information on the internal quality of objects.
Here, we focused on internal quality deterioration of pears, with browning of the fruit flesh and cavities as one of the most occurring defects.
A deep learning based model with U-Net architecture was developed for segmentation of the radiography data.
To train the model, ground-truth semantics were required which could be obtained from combining a simulation and thresholding method on X-ray computed tomography (CT) data.
Also, the radiography images themselves were simulated from the available CT data.
First, the model was trained on healthy pears and pears with both browning and cavities.
Then, this model could be further improved by using synthetic data on pears with only browning and only cavities.
From the predicted data, the kind of defect, its size and location could be immediately observed, which is beneficial compared to earlier classification models.
Besides external quality, also internal quality should be guaranteed to consumers.
Therefore, this work proposes X-ray radiography imaging, which is a fast, accurate and cheap technology which can be implemented in an inline system, and it gives information on the internal quality of objects.
Here, we focused on internal quality deterioration of pears, with browning of the fruit flesh and cavities as one of the most occurring defects.
A deep learning based model with U-Net architecture was developed for segmentation of the radiography data.
To train the model, ground-truth semantics were required which could be obtained from combining a simulation and thresholding method on X-ray computed tomography (CT) data.
Also, the radiography images themselves were simulated from the available CT data.
First, the model was trained on healthy pears and pears with both browning and cavities.
Then, this model could be further improved by using synthetic data on pears with only browning and only cavities.
From the predicted data, the kind of defect, its size and location could be immediately observed, which is beneficial compared to earlier classification models.
Publication
Authors
A. Tempelaere, H. Minh Phan, T. Van De Looverbosch, P. Verboven, B. Nicolaï
Keywords
X-ray CT, internal disorders, segmentation, deep learning, synthetic data
Online Articles (42)