Unsupervised Segmentation of Micro-CT Scans of Polyurethane Structures By Combining Hidden-Markov-Random Fields and a U-Net

The paper presents the Selectively Suppressed Perfect Reconstruction (SUPER) Decoder Block, designed to enhance U-Net variants used in inverse problems. This method addresses the issue of information loss that limits the recovery of fine details in images. By leveraging the perfect reconstruction property of wavelets, SUPER selectively suppresses redundant features, thus improving representational richness and maintaining computational efficiency. Experiments demonstrate its effectiveness across various benchmarks, particularly in thin-crack segmentation and image denoising tasks.

A new segmentation pipeline utilizing a lightweight U-Net model has been developed for analyzing carbide microstructures in scanning electron microscopy (SEM) images of reactor-pressure-vessel steel. This model, trained on only 10 annotated SEM images, achieves a Dice-Sørensen coefficient of 0.98, significantly surpassing traditional methods that yield a coefficient of 0.85. The approach not only reduces annotation efforts by an order of magnitude but also generalizes to other steel types, enhancing automated carbide quantification for alloy design.

The article presents CLUE (Controllable Latent space of Unprompted Embeddings), a generative model framework designed for text-to-image synthesis. CLUE aims to generate diverse images while ensuring stability, utilizing fixed-format prompts without the need for additional data. Built on the Stable Diffusion architecture, it incorporates a Style Encoder to create style embeddings, which are processed through a new attention layer in the U-Net. This approach addresses challenges faced in specialized fields like medicine, where data is often limited.