Utilization of Artificial Neural Networks to classify sunflower seed damage

Artificial Neural Networks (ANN's) are recognized to be effective for use in pattern recognition and data classification problems. This work was aimed to classify Xray images of sunflower seeds as to their level of damage. Therefore, the sunflower seeds were radiographed and the generated images were categorized into full seeds, seeds with slight damages or deformed seeds. These images were resized in order to standardize their dimensions and decrease the number of entries to the ANN. After this, the equalization of the images was performed, aiming at improving the contrast and thus accentuating the imperfections present in the seeds. A feed-forward topology ANN and a hidden layer was used and the set of images was randomly divided, reserving part of the images for validation and testing of the trained ANN. After a thousand training sessions with each configuration, the average accuracy of the neural network was 74.5% using the three classes, from 77.4% for full seeds versus seeds with slight damage, 96,0% for full seeds versus deformed seeds and 86.5% for seeds with slight damage versus deformed seeds. Therefore, the approach used can be applied to the classification of seeds in an automated way, since it obtained good indexes of correctness.