In this paper we propose a surrogate approach to extract fibers and voids from polymer matrix composites by combining results obtained from model-based methods to train convolutional neural networks. This approach focuses on microscopy images where labeled data is not readily available, but purely model based approaches can be too slow due to their computational complexity. In addition, we propose an encoder-decoder alternative to a fiber instance segmentation paradigm, showing a speedup in training and inference times without a significant decrease in accuracy with respect to alternative methods. The neural networks approach represent a significant speedup over model based approaches and can correctly capture most fibers and voids in large volumes for further statistical analysis of the data.