A Study on Optimum Feature Selection for Recognition of Starking Apple

Abstract

Pattern recognition problems require selection of feature subset in the image process. In order to increase the performance and to reduce the fails of classifier, it is needed to construct a sensitive feature selection. In this study, firstly apple images were taken from different angles of each apple via developed computer vision system. These images were then segmented and 25 features of the pattern were extracted by colors, texture and statistical. Feature selection methods which are SFS (Sequential Forward Selection), SBS (Sequential Backward Selection), SFFS (Sequential Floating Forward Selection), SBFS (Sequential Backward Floating Selection) were compared to specify the feature subset which has the maximum discriminatory. To perform this, a database used which has 726 patterns of 100 apples. According to experimental results, feature subset with 11 elements was created with SBFS method which is more successful than the other methods with % 83,1 accuracy rate. This method provides more successful labeling of patterns (defect, apple calyx, apple stem and healthy tissue). 