Abstract:Accurate identification of healthy plug seedlings is vital to intelligent greenhouse transplanting. Based on machine vision, this study presents a method to improve the accuracy and efficiency of plug seedlings' health identification, which is achieved by using image pre-processing, multilayer perceptron neural network image segmentation and connectivity domain centre coordinate calculation. To reduce the effect of invalid data on plug seedling image processing, the projection method is used to extract the image of the target region in the seedling tray where the plug seedling is situated. In addition, a multilayer perceptron neural network algorithm is utilised to segment the image of the target region into the leaves of the plug seedling and the background substrate. To solve the problem of overlapping seedling leaves, the centre coordinates of the connectivity domain are precisely calculated and matched with the corresponding holes. The health of the plug seedlings is then assessed using an area threshold. Finally, a greenhouse transplanting robot was developed and experiments were conducted to affirm the effectiveness of this approach. The experiments show that under different light intensities, the average accuracy of health identification for plug seedlings is over 96.90%. Moreover, the average transplanting success rate is 95.86% and the average number of transplanted plug seedlings per hour is 2117.65, indicating that the proposed method can accurately and quickly identify healthy plug seedlings and perform transplanting tasks, which provides guidance for intelligent transplanting in greenhouses.

Evaluation of health identification method for plug seedling transplantation robots in greenhouse environment