Combining Conditional Generative Adversarial Networks and YOLOv4 for Mango Classification

Mango is one of the most important exporting products in Taiwan and it has provided Taiwan with numerous economic benefits. To maximize the value, mangos are usually classified into grade A, grade B, and grade C before they are sent to exporting process, domestic sale, or fruit punching factory, respectively. Since the freshness of mangos is limited by time, it is important to accelerate the classification process. To improve the classification performance, this research proposes the Irwin mango classification system (IMCS) by combining Conditional Generative Adversarial Network (CGAN) and YOLOv4. CGAN is applied to help data expansion when the amount of data is insufficient. YOLOv4 is applied to classify the grade of mango and to detect the location of mango. To show our approach is better than other methods, we compare our model with YOLOv3 SPP, YOLOv4, ResNet, AlexNet, VGG 16, DenseNet, ShuffleNet, Fusion model, and the best outcome in AI CUP. The experiments show that our CGAN plus YOLOv4 (CGAN+YOLOv4) model has the best outcome with 85% precision for training, 82% precision for testing, and 90.07% for mAP. The weighted average recall (WAR) is 84.83% which is better than the best outcome of AI CUP 2020.