Comparison of Features for Strawberry Grading Classification with Novel Dataset

Abstract

This paper compares several image processing-based features for Support Vector Machine (SVM) to classify strawberries quality. We use binary class labels: good and damaged strawberries. We collect a new dataset taken from a webcam, consisting of 382 image data for good and 350 data for damaged-strawberries. The images are pre-processed with segmentation, resizing, and padding before their features are extracted. The features are used as inputs for SVM classifier. We evaluate each feature with cross-validation, and then we compared their accuracies. The compared features are Red Green Blue color models (RGB), Hue Saturation Value color model (HSV), RGB histogram, Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), Oriented FAST and Rotated BRIEF (ORB), and Histogram of Oriented Gradients (HOG). The experimental results showed that SURF achieved the best accuracy. The dataset is published by the authors as a free GNU General Public License so that readers can do further research, either with a combination of existing features or research their own features to produce better accuracy.