Deep computer vision system and explainable artificial intelligence applied for classification of dragon fruit (Hylocereus spp.)

Abstract

Dragon fruit (Selenicereus undatus) has gained popularity in the Brazilian market, thus demanding enhanced methods for quality assessment and control. Due to its unique morphological features, evaluating its quality presents a significant challenge to the industry. Image analysis, combined with Deep Learning (DL) techniques, offers a promising solution for visually discriminating agricultural products. However, DL architecture such as residual neural networks (ResNet) and vision transformer (ViT) are challenging to train, especially for new objects, such as exotic fruit. In this work, it was compared two DL Computer Vision System (DCVS) architectures, ResNet and ViT transformer, combined with Explainable Artificial Intelligence (XAI) methods for interpretation of black-box models, such as Grad-CAM and attention maps. Machine learning methods were applied for the classification of dragon fruit in four shelf-life stages. DCVS map reveals the potential to use the morphological aspects of dragon fruit and to predict its shelf-life stage. ViT Small outperformed the other models (ViT Tiny and ResNet 18 and 50), achieving an overall accuracy of 91.0%. This approach has potential for the classification of agricultural products.