Classification and weighing of sweet lime (Citrus limetta) for packaging using computer vision system

Weight is widely used as an important measure to study the physiology and agronomy for monitoring the fruit growth, grading, and packaging. The development of a computer vision system to measure the sweet lime fruit weight by relating the weight with its physical attributes is economically efficient than the mechanical online load cell used in the fruit sorting machines. In the present work, firstly a classification tree is developed using classification and regression tree algorithm to classify the fruits based on size. The average accuracy, sensitivity, specificity, and F score achieved are 98.16%, 94.01%, 98.51%, and 94.85% respectively. Secondly, parametric and non- parametric models are developed for predicting the weight of these classified fruits. A non-parametric model is developed using feed forward artificial neural network (FFANN) with error back propagation. The best topology is found among the fifty different FFANN configurations formed by varying the count of neurons in the hidden layer. Two parametric models are also developed using an approach of dimensional analysis (DA), and normal regression (NR). If the volume and the weight of the fruit have high correlation; then the bulk density of the fruit is fairly constant. This is the hypothesis used for developing the DA model. A lower value of mean square relative error and the remarkable value of Nash–Sutcliffe coefficient of efficiency indicate the superiority and the robustness of the proposed NR model in estimating the weight of the sweet lime fruits. Furthermore, an estimation uncertainty Theil_UII value which demonstrates the effectiveness and the credibility of the model’s estimation ability is used for performance evaluation.