Semi-supervised learning of hyperspectral image segmentation applied to vine tomatoes and table grapes
Jeroen van Roy,a Niels Wouters,b Bart De Ketelaerec and Wouter Saeysd,* aKU Leuven, Department of Biosystems, MeBioS, Kasteelpark Arenberg 30, B-3001 Leuven, Belgium bKU Leuven, Department of Biosystems, MeBioS, Kasteelpark Arenberg 30, B-3001 Leuven, Belgium cKU Leuven, Department of Biosystems, MeBioS, Kasteelpark Arenberg 30, B-3001 Leuven, Belgium dKU Leuven, Department of Biosystems, MeBioS, Kasteelpark Arenberg 30, B-3001 Leuven, Belgium
Nowadays, quality inspection of fruit and vegetables is typically accomplished through visual inspection. Automation of this inspection is desirable to make it more objective. For this, hyperspectral imaging has been identified as a promising technique. When the field of view includes multiple objects, hypercubes should be segmented to assign individual pixels to different objects. Unsupervised and supervised methods have been proposed. While the latter are labour intensive as they require masking of the training images, the former are too computationally intensive for in-line use and may provide different results for different hypercubes. Therefore, a semi-supervised method is proposed to train a computationally efficient segmentation algorithm with minimal human interaction. As a first step, an unsupervised classification model is used to cluster spectra in similar groups. In the second step, a pixel selection algorithm applied to the output of the unsupervised classification is used to build a supervised model which is fast enough for in-line use. To evaluate this approach, it is applied to hypercubes of vine tomatoes and table grapes. After first derivative spectral preprocessing to remove intensity variation due to curvature and gloss effects, the unsupervised models segmented 86.11% of the vine tomato images correctly. Considering overall accuracy, sensitivity, specificity and time needed to segment one hypercube, partial least squares discriminant analysis (PLS-DA) was found to be the best choice for in-line use, when using one training image. By adding a second image, the segmentation results improved considerably, yielding an overall accuracy of 96.95% for segmentation of vine tomatoes and 98.52% for segmentation of table grapes, demonstrating the added value of the learning phase in the algorithm.