This work aims to extend the previously proposed method of segmentation of irregular microstructural elements of materials by applying transfer learning and analyzing the approach’s effectiveness on other deep neural network architectures. The research enriched input information with additional structural-textural channels obtained from the k-means algorithm, the Sobel operator, and superpixel segmentation (Felzenszwalb). Previous research was limited to U-Net-based architectures; the DeepLabv3 model with ResNet50 encoder was also analyzed in this work. To improve segmentation quality, transfer learning was applied using pretrained weights from two datasets: the universal ImageNet dataset and the specialized MicroNet dataset containing microscopic images. Comparative experiments were carried out for models trained from scratch and using transfer learning, assessing the quality of segmentation based on standard measures such as IoU, Dice score, Precision, Recall, and F1-score. The results show that integrating extended input data with a suitably adapted architecture can significantly improve the quality of segmentation of irregular structures. The best results were achieved for Attention U-Net with pre-training on ImageNet, which confirms the potential of combining nonsupervised methods and transfer learning in tasks of precise segmentation of material microstructures.