The Design and Evaluation of an Orange-Fruit Detection Model in a Dynamic Environment Using a Convolutional Neural Network

Abstract

Agricultural robots play a crucial role in ensuring the sustainability of agriculture. Fruit detection is an essential part of orange-harvesting robot design. Ripe oranges need to be detected accurately in an orchard so they can be successfully picked. Accurate fruit detection in the orchard is significantly hindered by the challenges posed by the illumination and occlusion of fruit. Hence, it is important to detect fruit in a dynamic environment based on real-time data. This paper proposes a deep-learning convolutional neural network model for orange-fruit detection using a universal real-time dataset, specifically designed to detect oranges in a complex dynamic environment. Data were annotated and a dataset was prepared. A Keras sequential convolutional neural network model was prepared with a convolutional layer-activation function, maximum pooling, and fully connected layers. The model was trained using the dataset then validated by the test data. The model was then assessed using the image acquired from the orchard using Kinect RGB-D camera. The model was then run and its performance evaluated. The proposed CNN model shows an accuracy of 93.8%, precision of 98%, recall of 94.8%, and F1 score of 96.5%. The accuracy was mainly affected by the occlusion of oranges and leaves in the orchard’s trees. Varying illumination was another factor affecting the results. Overall, the orange-detection model presents good results and can effectively identify oranges in a complex real-time environment, like an orchard.