Kiwi Fruit Detection with Deep Learning Methods

Abstract

The automatic detection of kiwifruit in orchards is a challenging task due to the similarity between the fruit and the complex backgrounds formed by branches and stems. Moreover, the traditional method of hand-picking kiwifruit heavily relies on human labor and affects the overall yield. This study focuses on the fast and accurate detection of kiwifruit in natural orchard environments, which is crucial for yield estimation and cost reduction. Two deep learning methods, Faster Region-based Convolutional Neural Network (Faster R-CNN) and Mask Region-based Convolutional Neural Network (Mask RCNN), are utilized for kiwifruit detection, and their results are compared. The study begins with obtaining images of kiwi trees from the Güngör farm in Samsun Çarşamba and creating an original dataset. Preprocessing techniques are applied to improve the dataset, followed by detection using the Faster RCNN method. Different pre-trained architectures like SqueezeNet and MobileNetV3 are used, achieving average precision (mAP) values of 87.4% and 88.8%, respectively. In the second part of the study, kiwifruit images are processed using the ResNet50-based Mask R-CNN method, which achieves a higher mAP value of 98.48%. The experimental results demonstrate the applicability and effectiveness of the proposed deep learning models for real-time kiwifruit detection in orchards. Accurate kiwifruit detection allows farmers to optimize yield prediction, reduce costs, and improve productivity. The application of Faster R-CNN and Mask R-CNN in this study showcases their potential for enhancing the efficiency and accuracy of kiwifruit detection in orchard environments.