A novel and accurate chess pattern for automated texture classification

In this study, a novel chess based local image descriptor is presented for textural image recognition. The proposed descriptor is inspired by chess game and the main objective of it is to extract distinctive textural features using chess game rules. Patterns of the proposed method are created by using the movements of the knight, rook and bishop chessmen and six feature images are constructed using the proposed chess-based textural image descriptor. Therefore, this method is called as chess pattern (chess-pat) consisting of 4 phases. These four phases are block division, binary features calculation using chess patterns, histogram extraction, feature reduction with maximum pooling and classification. In the first phase, the image is divided into 5 x5 overlapping blocks. To extract the features, the proposed chess patterns are used. In the proposed chess-pat, 6 varied patterns are used based on chessmen moves and 6 feature images are created based on these patterns. Then, the histograms of these images are extracted and they are combined to create feature set of 1536 dimensions (D). In addition, maximum pooling is used to reduce this feature set with 256D and two versions of the chess-pat are obtained during the feature extraction. K-nearest neighbor (KNN), support vector machine (SVM) and linear discriminant analysis (LDA) and are utilized for classification. To evaluate the performance of the proposed chess-pat, Outex TC 00013, Outex TC 00001, Outex TC 00000, Kylberg and 2D Hela texture datasets are used. We have obtained the best accuracy rates of 75.5%, 100.0%, 99.7%, 88.9% and 100.0% for 2D Hela, Outex TC00000, Outex TC00001, Outex TC00013, and Kylberg respectively. Also, the proposed chess-pat achieved 100.0% classification rate (perfect classification performance) for 2 datasets (Outex TC00000, Kylberg). These results confirm that our proposed chess-pat method is highly accurate.