Automatic counting of grapes from vineyard images

Harvesting, spraying, and yield estimation are difficult activities for farmers. They take time, many workers, and moreover, are not always accurate. Therefore, machines are required to ease and speed up harvesting, spraying, and yield estimation. In this study, automatic recognition of visible grape berries and bunches from Red, Green, and Blue (RGB) images acquired by a camera for harvesting, spraying machines, and yield estimation was investigated. The images of grapes of different sizes and colors were taken under divergent natural light conditions and contrasts. The freely available Iceland dataset containing white grapes and in addition, images of red white, and hybrid types of grape trees were picked and used in the study. Initially, the Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG) were extracted, individually and their combination were used as feature vectors. Next, the features obtained were categorized with Convolution Neural Network (CNN), Artificial Neural Network (ANN), and Support-Vector-Machine (SVM) separately. The samples of grape berry images in the Iceland dataset were employed to train the ANN and SVM classifiers. Finally, the grape bunches were detected by incorporating Density Based Spatial Clustering of Applications with Noise (DBSCAN) clustering method. The artificial neural network classifier with the combined features provided the best accuracy in single berry recognition. It is faster than SVM and CNN as well. The average accuracy, precision, and recall were 99.6%, 99.7%, and 99.5% respectively. The accuracies of grape berry and bunch detection from test images were obtained as 89.8% and 91.7% respectively. Results show that LPB+HOG as a feature with ANN as a classifier provide an efficient grape detection from images taken under variant natural illumination conditions.