Deep Orange: Mask R-CNN based Orange Detection and Segmentation

Abstract

The objective of this work is to detect individual fruits and obtain pixel-wise mask for each detected fruit in an image. To this end, we presents a deep learning approach, named Deep Orange, to detection and pixel-wise segmentation of fruits based on the state-of-the-art instance segmentation framework, Mask R-CNN. The presented approach uses multi-modal input data comprising of RGB and HSV images of the scene. The developed framework is evaluated using images obtained from an orange grove in Citra, Florida under natural lighting conditions. The performance of the algorithm is compared using RGB and RGB+HSV images. Our preliminary findings indicate that inclusion of HSV data improves the precision to 0.9753 from 0.8947, when using RGB data alone. The overall F1 score obtained using RGB+HSV is close to 0.89.