Support Vector Machine Kernel Model Calibration for Optimal Accuracy in Automatic Pineapple Slices Classification

Abstract

Sorting pineapple can be automated with use of computer vision. The unique challenge with the pineapple slices is variability of the fruit slices color, ripeness and texture due to varying environmental parameters and fruit types. The most common types of pineapple fruit are smooth Caen and MD2. Currently the pineapple industries sort the slices manually using casual workers. Before commencement of a typical production shift, there is startup shift where machine are cleaned, prepared and calibrated for the production. Fruit slices are also sampled and processed to simulated actual production. A mock sorting is done to help guide the worker for the expected sorting for the five categories i.e: fancy ¾, fancy ½, choice, broken and reject. To achieve a fully automated sorting process there is a need to calibrate machine model and capture the day to day variability of fruit color, ripeness and texture. In this paper we propose to use an analytical method to calibrate the Support Vector Machine (SVM) with Gaussian radial basis function (RBF) for optimal sigma and box constraint (C). A compelling feature of the proposed algorithm is that it does not require an optimization search, making the selection process simpler and more computationally efficient. The proposed algorithm achieves the highest accuracy when used with the Gaussian multiclass SVM, as demonstrated by experimental results on three real-world datasets.