Dragon fruit weight estimation based on machine vision and machine learning

Objective: In order to solve the problem of manual weighting of dragon fruit, including time-consuming, laborious and expensive, an automated weight estimation method based on machine vision and machine learning was proposed in this research. Methods: Firstly, 106 dragon fruits were weighed, recorded and photographed, and images of dragon fruits were constructed. Secondly, binary images were obtained after denoising and segmentation. Moreover, the three features of pixel area, major axis pixel length and minor axis pixel length of dragon fruits were extracted on the basis of binary images. The three features of each image and their corresponding weights were combined into a set of data, which was divided into training set and test set according to the ratio of 7∶3. Finally, the training set was input into the Gradient Boosting, Random Forest, K-Neighbors and Artificial Neural Networks machine-learning models for training, and the test sets were used for model evaluation. Results: The evaluation index of the Artificial Neural network performed well compared with other models, with R2 of 0.986 and RMSE of 13.091. Conclusion: The experimental result demonstrates that the method proposed in this research can accomplish the weight estimation of dragon fruit effectively, and meet the weight estimation requirements of dragon fruit.