Детальная информация

The subject of the graduate qualification work is “Enhancing the Reliability of Mosaic Assembly for Micro-Images of Electronic Structures (PCB/MEMS) Using a Pose Graph”. The work addresses reliable mosaic assembly for microscopic images of electronic structures. It is based on the semester reports and task statement that define the project as a combination of literature review, confidence-aware mosaic assembly, pose graph optimization with confidence weights, and results discussion. The object of study is a regular-grid microscopic image stitching workflow for PCB/MEMS-like images and NIST MIST microscopy data. The aim is to increase the reliability and interpretability of the mosaic assembly process by combining local image registration, edge-level confidence estimation, diagnostic visualization, and confidence-weighted pose graph optimization. The implemented workflow uses Python and OpenCV for tile parsing, serpentine-order correction, overlap-band search, local registration, intra-row and inter-row stitching, reliability heatmap generation, and pose graph visualization. The final extension adds a robust least-squares optimization backend with a regular-grid prior, soft-L1 loss, confidence-based edge weighting, and diagnostic residual outputs. The Phase dataset demonstrates stable optimization behavior: the mean observation residual decreases from 5.5000 pixels to 0.9008 pixels, and the weighted observation residual decreases from 4.1233 pixels to 0.7627 pixels. The Cy5 dataset behaves as a stress test: the row-step median is pulled closer to the expected grid prior and the row-step MAD decreases from 58.5 pixels to 9.07 pixels, corresponding to an 84.5% improvement, but the final mean observation residual remains high. These results show that pose graph optimization improves global consistency when local observations are weak but compatible, while extremely sparse fluorescence data still require a stronger modality-aware local registration module.