Research and Realization of Crop Instance Segmentation Based on YOLACT

Abstract

Abstract: The invention realizes a computer vision-based system for improving crop picking accuracy and controlling automatic path-finding of robotic arm motion. Combined with the work of YOLACT (You Only Look At CoefficienTs) on Resnet-101, the traditional two-stage instance segmentation model is modified into a one-stage model, 64 robust masks are generated from the deepest layer, and the prediction coefficients are output according to the four-layer characteristic pyramid network in turn. The mask coefficients of different layers are given different weights to improve the mask accuracy. The plant is positioned, and a mask is generated to cover the crop closely pixel by pixel. The center point and contour coordinates of the crop are also outputted. Then, A* algorithm and its heuristic function are used to conduct piecewise shortest path planning using these coordinates in turn and generate corresponding G code statements to control the movement of the mechanical arm. Finally, the mechanical claw is combined with the center point and contour coordinates of the crop to pick crops.