Metallographic image segmentation using feature pyramid based recurrent residual U-Net

Understanding the fundamental microconstituents of metallic microstructure is crucial for comprehending the physical and mechanical properties of the metal. Artificial Intelligence based metallographic image processing techniques have been pivotal in analysing the materials. Particularly, deep learning-based methods for image segmentation are used to assign semantic labels to each pixel of the image, thereby partitioning the image into meaningful regions. In this paper, we employ a modified version of the popular U-Net architecture, which we name as MIMU-Net (Metallographic Image Segmentation using Modified U-Net) in order to enhance its performance for metallographic image segmentation. We apply recurrent residual blocks in the encoder part of the U-Net model to capture long-range dependencies and facilitate the smooth flow of information through the architecture. To obtain highly relevant contextual feature maps, we incorporate the concurrent spatial and channel squeeze and excitation module into the U-Net. In the decoder part of the U-Net, we employ feature pyramid pooling layers that upsample the high level feature maps and enable multi-scale feature extraction, thereby allowing the model to effectively handle objects of varying sizes and also capture fine-grained details. For the assessment of the proposed model, we conduct experiments on a recently developed metallographic image dataset, called MetalDAM. The obtained results demonstrate the promising performance of our modified U-Net architecture, outperforming several state-of-the-art methods. The source code of the present work is available in : Github