A novel advanced local binary pattern for image-based coral reef classification

High computational burden and low accuracy in non-uniform textures are the two main challenges of coral reef classification frameworks. To overcome these drawbacks, two novel forms of mapping approaches are proposed to enable Local Binary Patterns (LBP) scheme to extract discriminative features from textures. The mapping approach is a way to map the extracted features into a histogram (features vector) efficiently. In other words, the mapping method can merge some features into a feature and provides lower number of features efficiently. The proposed mapping techniques can be used for various types of LBPs; consequently, the extended LBPs can be applied to all types of textures. Benthic texture datasets are employed to assess the proposed method compared to the traditional ones. Regarding the multimodal distribution of the elicited features, K-Nearest Neighbor (KNN) is employed for classifying the extracted features. Here, the proposed mapping methods are tested on a special form of completed local binary patterns (CLBP). From the accuracy point of view, the extended CLBPs demonstrate higher accuracy compared to CLBP and also other state-of-the-art LBPs. Moreover, the proposed mapping approaches enhance the accuracy of rotation invariant LBPs, especially for large neighborhood. The proposed methods improve the classification accuracy for both noisy and noise-free images. From the computational complexity point of view, the extended CLBPs provide lower number of features compared to the others which leads to a faster recall time in KNN classifier.