期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2017
卷号：95
期号：21
页码：5756
出版社：Journal of Theoretical and Applied
摘要：The texture features would be important part when we conduct image classification. Local Binary Pattern (LBP) is one of feature extraction method that has most improvements by many researchers. Weighted Rotation- and Scale-invariant LBP (WRSI-LBP) is one of improvement versions. It uses minimum magnitude of local differences as an adaptive weight (WRSI-LBP-min) to adjust the contribution of LBP code in histogram calculation. The motivation is minimum magnitude gives minimum distortion to change LBP code in histogram calculation. In the classification of mango leaves case, the texture characteristic of mango leaves is highly difficult to be differed directly. So, for high accuracy detection, system requires texture feature with strength discrimination character, robust to illumination change, not sensitive to scaling and rotation. To achieve the goal, we propose average and maximum of magnitude of local differences as an adaptive weight of WRSI-LBP (WRSI-LBP-avg and WRSI-LBP-max). This scheme can be used to generate texture features for classification of mango leaves and general classification cases. The motivation of average weight is to cover all local different magnitude, because each LBP code generated would has unique neighbors pattern. The motivation of maximum is it gives maximum distortion to change LBP code, but it gives highest local different magnitude. We use Support Vector Machine (SVM) and K-Nearest Neighbor (K-NN) as classification methods. We use 240 images for performance evaluation, contains three varieties: Gadung, Jiwo and Manalagi. The K-Fold Cross Validation and Leave-One-Out are used as validation method. From the experiments show that WRSI-LBP-avg and WRSI-LBP-max achieve the highest accuracy compare to WRSI-LBP-min, LBP, Center Symmetric LBP (CS-LBP) and Dominant Rotated Local Binary Pattern (DRLBP). SVM achieve accuracy 75.21% with 16 bins, while K-NN achieve accuracy 79.17% with 256 bins. For uniform pattern, we apply experiments to WRSI-LBP-min, WRSI-LBP-avg, and WRSI-LBP-max. The highest accuracy is also achieved by WRSI-LBP-avg and WRSI-LBP-max.
