Color fractal texture image complexity estimation using a convolutional neural network

In this paper we analyze the capacity of a convolutional neural network (CNN) to understand and model color texture. More specifically, we ask the question if a CNN is able to predict the complexity of a color texture image, in particular of a color fractal image. We used color entropy and color fractal dimension to compute the complexity of both natural color texture images from the VisTex image data base and synthetic color fractal images. We modified the last layer of a ResNet-18 CNN so that the network outputs a real number representing the input image complexity expressed either as color entropy or color fractal dimension. We trained the modified ResNet CNN in two scenarios: when fractal images were not part of the training set and when they were part of the training set. We report on our experimental results and draw the conclusions.