Segmentation of field grape bunches via an improved pyramid scene parsing network

With the continuous expansion of wine grape planting areas, the mechanization and intelligence of grape harvesting have gradually become the future development trend. In order to guide the picking robot to pick grapes more efficiently in the vineyard, this study proposed a grape bunches segmentation method based on Pyramid Scene Parsing Network (PSPNet) deep semantic segmentation network for different varieties of grapes in the natural field environments. To this end, the Convolutional Block Attention Module (CBAM) attention mechanism and the atrous convolution were first embedded in the backbone feature extraction network of the PSPNet model to improve the feature extraction capability. Meanwhile, the proposed model also improved the PSPNet semantic segmentation model by fusing multiple feature layers (with more contextual information) extracted by the backbone network. The improved PSPNet was compared against the original PSPNet on a newly collected grape image dataset, and it was shown that the improved PSPNet model had an Intersection-over-Union (IoU) and Pixel Accuracy (PA) of 87.42% and 95.73%, respectively, implying an improvement of 4.36% and 9.95% over the original PSPNet model. The improved PSPNet was also compared against the state-of-the-art DeepLab-V3+ and U-Net in terms of IoU, PA, computation efficiency and robustness, and showed promising performance. It is concluded that the improved PSPNet can quickly and accurately segment grape bunches of different varieties in the natural field environments, which provides a certain technical basis for intelligent harvesting by grape picking robots.