Features construction for starfruit quality inspection

Abstract

Up to present, the starfruit quality inspection process is performed manually. Manual inspection will cause inconsistency in quality due to human subjective nature, slow processing and labor intensive. Hence, this thesis presents automation process development for the starfruit quality inspection in terms of techniques and algorithms design based on image processing. Basically, there are three main processes of the starfruit quality inspection discussed in this thesis, which are the maturity index classification, skin defect estimation and shape defect estimation. Throughout these processes, new features constructed based on colors and shape are proposed. In maturity index classification, a two-color feature, M, is proposed to differentiate six maturity indices of the starfruit. With the two-color feature, one third of computational data is reduced compared to the typical 3-color features. For skin defect estimation process, a new gray level co-occurrence matrix (GLCM) statistical feature is introduced. This feature has the ability to segment skin defect areas on non-homogenous in illumination and color of the starfruit image. As the GLCM consumes high computation, this thesis proposed a new algorithm based on Haar wavelet that reduces computational burden. Lastly, a shape-based feature is constructed for the shape defect estimation process where a modification of Melkmen convex hull algorithm is designed in order to construct the feature. Experimental results prove that these features are able to convey the three main processes of the starfruit quality inspection process where high accuracies were achieved; 93.33% for the maturity index classification feature, 82% for the skin defect estimation feature and 96% for the shape defect estimation feature.