Block-based selection random forest for texture classification using multi-fractal spectrum feature

This paper proposes a block-based selection random forest (BBSRF) for texture classification task using multi-fractal spectrum (MFS) feature descriptor. The random feature selection method for node splitting in random forest may omit some features which would be informative and critical to represent the instances. The BBSRF ensures that each feature would be considered via the block-based selection strategy. In BBSRF, all features are divided into \(k\) blocks; next, we generate synthesis feature subset which is made up of all features in one block and \(m\) random features from the remaining \((k-1)\) blocks; finally, each node splitting of the random tree is operated on one synthesis feature subset. After all blocks have been searched, all features are re-divided into new \(k\) blocks. The above process works iteratively until the satisfactory result is obtained. Once the random trees have been built, a testing instance is classified by voting from them. We conducted the experiments on five texture benchmark datasets with the help of MFS feature. Experimental results demonstrate the excellent performance of the proposed method in comparison with state-of-the-art results on these datasets.