34186

Intuitive, semi-supervised training of the segmentation of hyperspectral images

Favoritar este trabalho

Hyperspectral imaging has large potential for automated, objective quality inspection in the food industry. However, the adoption of this technology is hampered by the large investment for building segmentation and calibration algorithms by highly skilled technicians each time a new product, new cultivar or new recipe has to be handled. Therefore, most food companies still prefer to rely on subjective, visual inspection by humans as they are more flexible. To overcome these limitations and to promote the adoption of hyperspectral food quality inspection, this research focuses on the development of a semi-supervised training method which allows an operator without knowledge on multivariate statistics or image processing techniques to train the system in an intuitive way to handle a new product or cultivar.

The developed semi-supervised segmentation algorithm works in three steps. First, the hypercubes are segmented in an unsupervised way by clustering the pixel spectra in a pre-defined number of clusters corresponding to the number of objects/components set by the operator. The segmented images resulting from these unsupervised methods are then presented to the operator together with the RGB image to allow him/her to decide whether any of the proposed segmentations is acceptable and to choose the best one. In a second step, the most informative pixel spectra from each group resulting from the segmentation step are selected to define a calibration set for training a supervised classification model. In a third step, this supervised classification model is applied to a set of new samples and the output is presented to the operator for visual inspection to decide whether the segmentation output is satisfactory. If the segmentation is not satisfactory, step one is repeated on the hypercube for this sample and the supervised classification model is updated based on the pixels selected from the selected segmentation.

This method has been applied to hyperspectral data of vine tomatoes (Solanum lycopersicum). Two cultivars of vine tomatoes, Merlice and Prunus were harvested in June 2014 and August 2014, respectively. The ripeness of these tomatoes was variable, going from unripe to very ripe. They were measured using a line-scanning Visible - Near-Infrared (Vis-NIR) hyperspectral camera in reflectance mode. After the measurement of the tomatoes, the hypercubes were analysed in Matlab.

For this application, the results of this semi-supervised training method are very promising. With this approach, it was possible to teach the system very quickly to segment new, never seen products on the conveyor belt. Applied on the tomatoes, only a few hypercubes in the training phase of the algorithm were needed to segment tomatoes correctly. Both ripe and unripe tomatoes could be accurately separated from the stalk and the background. Thanks to the use of the supervised classification method, the average time needed to segment images was reduced by a factor 17,5 compared to the average time needed by the unsupervised classification methods. This makes the intuitively trained segmentation method suitable for on-line use.