Using Computer Vision for Mango Price Estimation Based on Breed Classification and Quality Grading

Mango is a prevalent fruit. It can be found in almost every area in Thailand because it is easy to grow and has many breeds. While they share a very similar appearance, each breed of mango has a variety of flavors, both sour, sweet, and oily. Therefore, consumers need clarification on buying mangoes. Most previous works focused on breeding and grading but, to the best of our knowledge, none proposed price estimation regarding different breeds and grades of mangoes. This work proposes using computer vision to do mango breed classification, quality grading, and price estimation. Three varieties of mangoes (total of 368 images) are used—167 images of Kiewsavoey Mango, 186 images of Turmeric Mango, and 15 images of Numdokmai Mango. Images of each breed are divided into two grades, Grade B and Grade C. We use an object detection model named YOLOv5 to sort the breed and use two image classification models to tell the quality and grade of the mangoes. In the classification models, our experiments use transfer learning on three pre-trained models—ResNet18, MobileNetV2, and GoogLeNet. As a result, the ResNet18 model yields the best accuracy for grading Kiewsavoey Mango (accuracy = 87.0968%) and Turmeric Mango (accuracy = 100%).