Apple weight estimation based on joint image optimal feature extraction and improved RBF neural network

Objective: Taking Aksu apples as an example, a joint image optimal feature extraction and improved RBF neural network learning apple weight estimation method is designed to overcome the high cost and large error of manual grading and weighing. Methods: Firstly, an apple image acquisition system was established to obtain apple foreground image information. Secondly, the optimal subset extraction strategy for apple image feature sets was designed, by transforming the process of extracting the optimal subset into an objective function optimization problem, and an improved discrete locust optimization algorithm was designed to obtain the optimal apple image feature subset. Finally, a weight estimation model for apples based on RBF neural network learning was constructed, with the optimal feature subset as network input. The locust optimization algorithm was used to optimize the configuration of RBF neural network hyperparameters, to achieve effective estimation of apple weight. Results: The proposed apple weight estimation method had higher accuracy, with an average relative error rate of 1.23% for weight estimation. Conclusion: This method can effectively achieve apple weight estimation and can also be applied to other fruits with similar axisymmetric shapes for weight estimation.