Implementation and performance evaluation of classifiers SVM, CNN and ANN in vineyard estimation

Grape detection, harvesting, spraying, and yield estimation are difficult activities for farmers. They take time and many workers (cost) and, moreover, are not always accurate. Therefore, harvesting with the aid of smart robots is now being explored and is currently a frontline problem in agriculture. The experiment presents an algorithm for the automatic visible recognition of grape berries and grape bunches so that they can be counted for yield estimation. Recognition is based on high-resolution images of grape berries of different sizes and colors and taken under different illumination conditions, including natural light, and different image contrasts. In this thesis, presented a grape bunch and berry recognition by employing machine learning methods; Convolution Neural Network (CNN), Artificial Neural Network (ANN) and Support-Vector-Machine-(SVM). Histogram of Oriented Gradients (HOG), The Local Binary Pattern-(LBP), and combination of both attributes have been used as feature vectors. The approach has been implemented in the freely available Iceland dataset for single berry detection and classifier train. The data contain two classes (berry and non-berry). ANN classifier with HOG+LBP (combination) features show better results in single berry recognition with an average of accuracy, precision and recall are 99.64±0.073 %, 99.70±0.094%, and 99.59±0.117% respectively. Besides on the time of detection. The proposed method for grape bunch and berry detection in field images uses a Fast Radial Symmetry Transform (FRST) as interest point detector. In the next, feature extraction and classification are computed in each interest point with a multi-scale approach. Afterwards, a DBSCAN method defines the number of grape bunches. Each cluster's spatial distribution and needs to close bunch separation for an improved bunch detection. The average accuracy of grape bunch detection 91.72±2.88% and 89.82±2.58% for berry detection. Grape bunch detection and a comparison of berry and non-berry pixels were studied, using ground-truth the labels are made manually. Four different colors of grape images, with a diversity of illuminations and acquisition protocol, were tested, with an average of the correct rate, sensitivity, and specificity are 96.97±0.95%, 97.57±0.969%, and 89.28±3.83% respectively in pixels of berry classification. Results show good performance with no need for special illumination, not require color feature used which allows recognition for white and red grapes as well as the bunch detection scheme working in varying scenarios. Keywords: Grape detection, grape bunch detection, HOG, LBP, DBSCAN, SVM, ANN and CNN