Fruit recognition and yield estimation using deep learning methods

Automatic detection of kiwi fruit in orchards is a challenging task due to the similarity of the fruit and the complex background of branches and stems. In addition, the traditional manual harvesting method of kiwi fruit is highly dependent on human labor and affects the overall yield. In this thesis, the focus is on the rapid and precise detection of kiwi fruit, which is critical for yield estimation and cost reduction, in the natural environment in orchards. Two deep learning methods, Faster regional convolutional neural network (Faster R-CNN) and Mask regional convolutional neural network (Mask R-CNN), are used for kiwi fruit detection and the results are compared. In the study, firstly, an original data set of images of kiwi trees from Samsun Çarşamba Güngör farm is created. Preprocessing techniques are applied to the dataset and then detection is performed with the Faster R-CNN and Mask R-CNN methods. Using pre-trained architectures such as SqueezeNet and MobileNetV3, the average precision (mAP) value was obtained as 87,4% and 88,8%, respectively. In the second part of the study, the ResNet50-based Mask R-CNN method is processed and a higher mAP value of 98,48% is obtained. Experimental results demonstrate the feasibility and effectiveness of proposed deep learning models for real-time kiwi fruit detection in orchards.