Precise 2D vision solutions for estimating avocado physical characteristics

This study proceeds with a systematic review and subsequent comparison between the geometry-based and regression model-based techniques on accurately estimating the physical properties of avocados particularly mass using manually collected cross-sectional data set. The objectives of this investigation are to handle grading, weighing and packaging effectively, and cost-effectively for the agricultural and food engineering industries. Among the findings, the Frustum method consistently outperforms other approaches across all slice configurations, achieving the lowest errors with Root Mean Square Percentage Error (RMSPE) of 4.24%, Mean Absolute Percentage Error (MAPE) of 3.43%, Coefficient of Determination ([Formula: see text]) value of 97.78% and Explained Variance Score (EVS) value of 97.63% at 20 slices. This highlights its robustness and reliability for precise mass estimation, especially without requiring large datasets or complex computations. To ensure the reliability of the regression models, hyper-parameter optimization and K-fold cross-validation techniques were employed, enabling the identification of optimal model configurations and minimizing over-fitting risks. Regression-based methods, such as Ridge Regression, also exhibit strong performance, with an average RMSPE of 4.30%, MAPE of 3.52%, [Formula: see text] of 97.71%, and EVS of 97.58% across 5 folds at 15 slices, making it a competitive and stable alternative. Other regression models, such as LASSO Regression, and Elastic Net Regression, delivered strong and consistent outcomes across the evaluation metrics, followed by Linear Regression. In contrast, MLP Regressor and Gradient Boosting Regressor exhibited notable variability between folds, highlighting issues with stability and generalization. Dimensions of avocados were also assessed in this study with the error rating below 1.53% coupled with model-fit parameters of more than 99% showing that the models used had high accuracy in determining both the width and length of the avocado. The above findings provide a comparative perspective of the choice of forward models depending on the task characteristics including, availability of dataset, stability, and level of precision. Moreover, such outcomes reveal the applicability of these methodologies in implementing state-of-the-art automation technologies accordingly with a focus on robotic harvesting and grading solutions in automation of precision agriculture and modern intelligent food processing systems.