Deep learning-based embedded system for carabao mango (Mangifera indica L.) sorting

• Main Authors: Liwag, Ryan Joshua H., Cepria, Kevin Jeff, Rapio, Anfernee, Castillo, Karlos Leo, Cabatuan, Melvin K.
• Format: text
• Published: Animo Repository 2019
• Online Access: https://animorepository.dlsu.edu.ph/faculty_research/801; https://animorepository.dlsu.edu.ph/cgi/viewcontent.cgi?article=1800&context=faculty_research
• Institution: De La Salle University

©BEIESP. This paper presents the design and development of an embedded system for ‘Carabao’ or Philippine mango sorting utilizing deep learning techniques. In particular, the proposed system initially takes as input a top view image of the mango, which is consequently rolled over to evaluate every sides. The input images were processed by Single Shot MultiBox Detector (SSD) MobileNet for mango detection and Multi-Task Learning Convolutional Neural Network (MTL-CNN) for classification/sorting ripeness and basic quality, running on an embedded computer, i.e. Raspberry Pi 3. Our dataset consisting of 2800 mango images derived from about 270 distinct mango fruits were annotated for multiple classification tasks, namely, basic quality (defective or good) and ripeness (green, semi-ripe, and ripe). The mango detection results achieved a total precision score of 0.92 and a mean average precision (mAP) of over 0.8 in the final checkpoint. The basic quality classification accuracy results were 0.98 and 0.92, respectively, for defective and good quality, while the ripeness classification for green, ripe, and semi-ripe were 1.0, 1.0, and 0.91, respectively. Overall, the results demonstrated the feasibility of our proposed embedded system for image-based Carabao mango sorting using deep learning techniques.