Affiliations: [a]
Department of Computer Science, Amrita School of Computing, Mysuru Campus, Amrita Vishwa Vidyapeetham, India
| [b]
Department of Computer Science, Yuvaraja college, Mysuru, India
Correspondence:
[*]
Corresponding author. N. Shobha Rani, Department of Computer Science, Amrita School of Computing, Mysuru Campus, Amrita Vishwa Vidyapeetham, Karnataka 570026, India. E-mail: [email protected].
Abstract: One of the most essential factors in classifying and qualitatively evaluating mangoes for various industrial uses is weight. To meet grading requirements during industrial processing, this paper presents an orientation-independent weight estimation method for the mango cultivar “Alphonso.” In this study, size and geometry are considered as key variables in estimating weight. Based on the visual fruit geometry, generalized hand-crafted local and global features, and conventional features are calculated and subjected to the proposed feature selection methodology for optimal feature identification. The optimal features are employed in regression analysis to estimate the predicted weight. Four regression models –MLR, Linear SVR, RBF SVR, and polynomial SVR—are used during the experimental trials. A self-collected mango database with two orientations per sample is obtained using a CCD camera. Three different weight estimation techniques are used in the analysis concerning orientation 1, orientation 2, and combining both orientations. The SVR RBF kernel yields a higher correlation between predicted and actual weights, and experiments demonstrate that orientation 1 is symmetric to orientation 2. By exhibiting a correlation coefficient of R2 = 0.99 with SVR-RBF for weight estimation using both orientations as well as individual orientations, it is observed that the correlation between predicted and estimated weights is nearly identical
