Article Preview
TopIntroduction
This paper contains an extended and redeveloped version of our presentation for the 25th Conference of Open Innovations Association FRUCT’25 (AL-Abri, Khriji, Ammari, & Awadalla, 2019).
Oman is one of the countries that depends on dates as a source of income. It has more than 250 varieties of dates having different colors, shapes, sizes, and texture (Ghnimi, Umer, Karim & Kamal-Eldin, 2017); (Bhargava & Bansal, 2018); (Janecek, Gansterer, Demel & Ecker, 2008). Manual classification of dates into different classes needs meticulous and hard effort. Automating these tasks using a dedicated embedded computer vision system will help classify the multiple varieties of dates fast and accurately. It will also improve the date’s production quality of the country (Haidar, Dong & Mavridis, 2012). However, automating the classification tasks is complex and challenging because of the irregular sample features within the same type of classes (Mizushima & Lu, 2013); (Manickavasagan, Al-Yahyai, & Khriji, 2014); (Naik & Patel, 2017) and of the similarities that can be found between samples from different classes (Hameed, Chai, & Rassau, 2018).
Many previous works implemented multiple artificial intelligence techniques to automate dates classification using multiple types of features. In 2012, 15 features are used to automate the classification of seven different categories of dates (Haidar, Dong & Mavridis, 2012). In this system, Nearest Neighbor, Linear Discriminant Analysis (LDA), and Artificial Neural Network (ANN) classifiers are tested for comparison purposes. Date classification using ANN studied in (Alrajeh & Alzohairy, 2012) achieved a maximum classification accuracy of 91.1%. In 2014, texture, color and shape features are extracted from the dates images (Muhammad, 2014). In this work, Fisher discrimination Ratio (FDR) is used for dimensionality reduction of features and Support Vector Machine (SVM) was selected for classification. An automatic date classifier is also developed in (Manickavasagan et al., 2017). This system uses histogram and texture features and implements LDA and ANN classifiers. In 2018, an automated system that identifies different date fruit maturity status and classifies their categories is developed. This system extracts color, size and skin texture features from images, counts the number of dates, classify them into different classes and identify the defects (Najeeb & Safar, 2018).
All these previous studies attempted to automate dates classifications using different artificial intelligence techniques and multiple types of color, shape, size, and texture features. However, none of these studies has considered embedded systems prototyping for real time classification. In this paper, we examine HW/SW Co-design for real time classification of Khalas, Khunaizi, Fardh, Qash, Naghal, and Maan dates fruit varieties in Oman. These varieties are among the most delicious and sweet types of dates. In this study, the effect of color, shape, size, and texture features to classify the different varieties of dates is studied and the performance of multiple ANN, SVM, and K-Nearest Neighbor (KNN) classifiers are compared. Embedded systems prototyping of the highest performant classifier is then considered aiming at a minimum of 10 fps real time classification performance.