3D reconstruction and feature extraction for agricultural produce grading

Abstract

This paper examines the grading of agricultural produce from multiple images using colour and texture properties. Some types of agricultural produce need to be inspected from multiple views in order to assess the entire appearance; however, using multiple images may obtain redundant data. Therefore, techniques are presented to reconstruct a 3D object, create new images without duplicated object areas and extract colour and texture features for evaluation. The performance of using multiple view images without duplicated object regions is compared with those of using only top-view images and the original multiple view images. Experiments are performed on apple and guava grading using kNN, NN, SVM and GP for classification. Performance differences from the different image sets are compared using McNemar's test and the Friedman test. It is found that the performance when using multiple view images is superior to that when using single-view images for all experiments. Employing features extracted from multiple view images without object area duplication achieves significantly higher accuracy than employing the original multiple view images for apple grading, but their performances do not differ significantly for guava inspection.