Tobacco Leaf Images Clustering using DarkNet19 and K-Means

Image clustering is an image grouping of classes without any previous labels. This article discusses Tobacco leaf image clustering. The data experiment was primary data from Pamekasan and Sumenep districts in Madura. The dataset consists of 3 clusters: healthy, curly, and hollow. Each cluster has 50 images, the total data is 150 images. To perform clustering, the system previously making feature extraction using DarkNet19. As a comparison, feature extraction was carried out using DarkNet53. DarkNet19 is a type of Convolutional Neural Network (CNN) architecture. This architecture consists of 19 Convolutional Layer, 18 Batch Normalization, 18 Leaky ReLU, 5max-pooling, 1 globalaverage-pooling, and 1 softmax. In addition to feature extraction, the system also performs dimensional reduction using Principal Component Analysis. Hereafter, clustering using K-means. The results show that the accuracy of 3 grade clustering of tobacco leaf was 0.81 until 0.93.