Improved Leaf Segmentation Method for UNet Feature Encoding

Abstract

Leaf segmentation provides more insight into horizontal characteristics such as leaf area, number and pressure. Plant leaves overlap under different environmental conditions such as light changes, wind blowing, growing place. It makes the leaf segmentation task more complicated. To address the above issues, this paper proposes a novel, simple and efficient segmentation algorithm GrcsUnet. It uses the GRCSblock module to improve the feature extraction section based on the UNet architecture. GRCSblock integrates the ideas of Resnet, GoogLeNet, channel attention and spatial attention. Firstly, it changes the Resnet residual connection method. Secondly, it draws on the parallel idea of GoogLeNet and adopts multiple downsampling methods to reduce the loss of semantic information. Finally,it embeds channel attention and spatial attention separately into channel stitching and module output, assigning different weights to each channel and spatial position. Experiments on KOMATSUNA and MSU-PID datasets show that the segmentation performance of GrcsUnet is better than that of advanced UNET, ResUNet and UNet++.