Pineapple Slices Classification Using a Hybrid Feature Extraction Technique and Multiclass SVM

Abstract

Automatic fruit quality inspection is important to improve quality and reduce production costs. Fruit grading and sorting is primarily visual. Real-time fruit inspection and grading requires high-quality photos with easily recognized features and fast software and hardware to process the images. Since forever, pineapples have been graded using a labour intensive and expensive method. Due to high labour costs and inconsistent in manual methods, industries have turned to automation. Visual sorting of pineapple slices is challenging to automate in food processing. The research intends to build a computer vision based algorithm for sorting and automating pineapple slices based on texture, colour and shape. This research designs a unique classification approach based on Max-Wins-Voting SVM employing Gaussian Radial Basis kernel for real-time automatic implementation. First, digital photos of pineapple slices are obtained at factory floor and their backgrounds are removed using Otsu segmentation. Then, features (colour, colour moment, texture, and shape) of each image are extracted using a hybrid method to create a feature space. Third, Principal Component Analysis reduces feature space dimensions. Finally, multi-class SVMs (Max-Wins-Voting SVM, Directed Acyclic Graph SVM and Winner-Takes-All SVM) are built. The accuracy and calculation time of SVMs with different kernels (Linear, Gaussian Radial Basis, and Homogeneous Polynomial) are compared. SVMs are trained with the reduced feature space vector using 5-fold stratified cross validation. Max-Wins-Voting SVM utilizing Gaussian Radial Basis kernel was found reliable and robust, as shown in result. The approach provides fast, accurate real-time automation of pineapple slice sorting achieving over 90% accuracy. The method classifies as good as human operator with added advantage of high quality standards and reduced production costs.