Combining efficient textural features with CNN — Based classifiers to segment regions of interest in aerial images

In this paper, we develop an methodology and corresponding algorithms that segments regions of interest like vegetation and flood from aerial images. To this end different textural features are used, particularly second order type (extracted from co-occurrence matrix — Haralick features) and histogram of oriented gradients (HOG). These features are calculated from image patches and the obtained values are considered as input of a convolutional neural network for classification and segmentation. A logical OR operator is used between the decisions of the classifier based on Haralick' features and the classifier based on HOG descriptors. The algorithms are tested and validated on 100 images taken from a UAV mission. The percentage of the occupancy of ROIs in the images is also computed.