Physical Property Analysis of Sweet Potatoes Using Computer Vision

This paper demonstrates the performance of using computer vision for sweet potato grading. Different qualities of sweet potatoes can lead to different prices. Better-grade fruits can be sold at a higher price. Therefore, they should be graded before they are transferred to sell. 471 pictures of sweet potatoes are employed for the experiment. The experimental sweet potatoes are divided into four groups. The first group contains good quality with required shapes. The second group also includes good quality roots but undesired shapes. Experimental samples of the third group have no defects, but they are too small, short or thin shaped. Moreover, for the last group sweet potatoes have defects. The samples are inspected from their physical properties including shape, color and texture features. Top-view pictures are captured and used for feature extraction. Various 184 physical properties are extracted. As using a large number of features may cause high computational cost, so important extracted features are selected. The effective features are used for classification based on k-nearest neighbour and neural network theories. The experiments comprise of twenty sub-experiments. About half the numbers of samples are randomly chosen for training sets, the remaining samples being employed for test sets. The experimental results show that in the average of the k-nearest neighbor and neural network classifiers achieve 97.14% and 96.46% accuracy respectively. Nevertheless, the difference of the classifier performances is insignificant that is proved by the paired sample t-test.