Abstract: Texture recognition is one of the most important tasks in computer vision, with numerous applications in several areas. Despite the recent success of end-to-end deep learning models in image recognition in general, when it comes to texture images, the classical approach based on the combination of a local predefined descriptor with a pooling encoder is still competitive and may be advantageous, especially when we do not have access to a large amount of data for training or when understanding what each step in the algorithm pipeline performs is an important matter. In this context, here we propose a method for texture recognition based on a hybrid approach that combines a randomized neural network with multifractal transform for a local complete and robust representation. The first stage of the method employs a randomized neural network similar to that used in extreme learning machines. We use pixel intensities within each particular neighborhood as inputs and pixels strategically placed around the neighborhood center as our (multiple) targets. The learnable weights (connecting hidden and output layers) are computed over the entire image. In the second stage, we compute a multifractal measure over each pixel and repeat the procedure with the randomized model. Finally, we combine the learnable weights of both representations to compose our final descriptors. The effectiveness of our approach is verified on texture classification, over benchmark databases, and on a practical task of plant species identification. In both cases, our method achieves promising results, being competitive with several approaches in the state-of-the-art of the area Ver menos