Detection and Instance Segmentation of Grape Clusters in Orchard Environments Using an Improved Mask R-CNN Model

Abstract

Accurately segmenting grape clusters and detecting grape varieties in orchards is beneficial for orchard staff to accurately understand the distribution, yield, growth information, and efficient mechanical harvesting of different grapes. However, factors, such as lighting changes, grape overlap, branch and leaf occlusion, similarity in fruit and background colors, as well as the high similarity between some different grape varieties, bring tremendous difficulties in the identification and segmentation of different varieties of grape clusters. To resolve these difficulties, this study proposed an improved Mask R-CNN model by assembling an efficient channel attention (ECA) module into the residual layer of the backbone network and a dual attention network (DANet) into the mask branch. The experimental results showed that the improved Mask R-CNN model can accurately segment clusters of eight grape varieties under various conditions. The bbox_mAP and mask_mAP on the test set were 0.905 and 0.821, respectively. The results were 1.4% and 1.5% higher than the original Mask R-CNN model, respectively. The effectiveness of the ECA module and DANet module on other instance segmentation models was explored as comparison, which provided a certain ideological reference for model improvement and optimization. The results of the improved Mask R-CNN model in this study were superior to other classic instance segmentation models. It indicated that the improved model could effectively, rapidly, and accurately segment grape clusters and detect grape varieties in orchards. This study provides technical support for orchard staff and grape-picking robots to pick grapes intelligently.