DETransMVSnet: Research on Terahertz 3D Reconstruction of Multi-View Stereo Network With Deep Equilibrium Transformers

Terahertz waves, positioned between microwaves and infrared in the electromagnetic spectrum, are distinguished by their exceptional penetration capabilities, minimal energy requirements, and consistent absorption profiles for specific substances. Their versatility in applications such as non-destructive evaluation, human security scans, and biological diagnostics has propelled them to the forefront of scientific inquiry. However, existing terahertz equipment poses limitations in terms of compromised resolution, diffraction-induced blurring, and degradation of clarity due to texture overlaps. Consequently, numerous multi-view 3D reconstruction algorithms struggle to produce high-quality results with terahertz imagery. To address these challenges—particularly the scarcity of terahertz datasets and texture conflation—we integrated X-ray images from the DTU dataset with our collected terahertz projections. By observing inconsistent texture projections in multi-view terahertz images resulting from angle shifts, we developed DETransMVSnet—a state-of-the-art multi-view 3D reconstruction approach based on the Multi-Scale Deep Equilibrium Layer (MDEQ) paradigm. Leveraging equilibrium layers within homography-projected feature maps enables us to extract masks that differentiate different layers within a scene. The Intra-Attention and Mask-Attention Blocks further refine feature selection by preserving relevant terahertz details while suppressing disruptive background elements. As evidence of its effectiveness, DETransMVSnet achieves comparable performance to conventional algorithms on the DTU dataset but notably outperforms them when applied to terahertz datasets by successfully reconstructing images where previous methods have failed.