Random Forest for Semantic Segmentation Using Pre Trained CNN (VGG16) Features

For years the semantic segmentation of images has become a high-level task, and a great challenge in the field of computer vision. From a technical point of view, semantic segmentation consists in segmenting an image by grouping the parts that belong to the same class of object (for example: beach, road, trees…). Today, several research works are proposed to treat this research axis, and still the challenge remains to have a powerful model to improve or augment these existing techniques. In this paper, we have proposed a hybrid method for semantic image segmentation, combining two popular and powerful techniques. These are the convolutional neural network CNN with the VGG16 model and the Random Forest algorithm, which are two outstanding classifiers. In fact, first of all we apply for the input image the pre-trained model based on the CNN architecture which is VGG16 to automatically extract the features. Then, these extracted features are used for another classification process with Random Forest to make the prediction of the labels of the input models. The proposed approach combines the strengths of two techniques which explains the improvement and accuracy of the obtained results. To make a comparison between the different methods, we adopted a set of evaluation measures: IOU, SSIM, SAM, and MS-SSIM. From the results, we can say that our approach shows its effectiveness in solving the problem of image segmentation in general.