sciwork 2023

How Can We “Perfectly and Rapidly” Stitch Images? Exploring Improved End-to-end Techniques
2023-12-10, 14:20–14:50 (Asia/Taipei), NYCU

We present fundamental theories of image stitching, using MATLAB based on feature detection and projection. Our technique, enhanced through mismatch detection, smooth blending, and reshaping, creates flawless panoramas. Impressively, it processes 0.23 times faster than Photoshop.

This presentation delves deep into how we improved the fundamental theories and built up an end-to-end pipeline of image stitching. Attendees can anticipate insights into:

  1. Basic Theories:
    An exploration into the mathematics and principles behind image stitching, delving into Scale Invariant Feature Transform, projection transformation, and continuous blending.

  2. Demonstration of Current Techniques:
    Utilizing MATLAB programming, we will showcase image stitching techniques grounded in the basic theories outlined above. This provides clarity into its operation and offers a firsthand look at the tangible outcomes.

  3. Optimizations and Advancements:

3-1. Feature Mismatch Detection:
Discussing strategies to detect and correct feature mismatches during the stitching process.

3-2. Smooth Blending:
We introduce an advanced technique wherein the Poisson equation is solved to generate a smooth mask, ensuring seamless transitions between stitched images. To enhance efficiency, a specialized fast algorithm is employed to solve the large-scale linear system.

3-3. Reshaping:
We delve into the use of computational geometry methods to deform the irregularly stitched image into a rectangle, refining its shape after stitching to achieve perfection.

3-4. Results Display:
A side-by-side comparison of stitching outcomes, both before and after implementing the improved techniques, affording attendees a vivid understanding of the optimization's impact.

Prior Knowledge: Image Processing, Calculus, Linear Algebra

Ask questions at slido

Prior Knowledge Expected? –

Yes, previous knowledge expected

 Language –

Mandarin talk w. English slides
Jiawei

Jiawei is a Ph.D. student in Computer Science and Information Engineering at NTU. His research focuses on Deep Learning for Computer Vision. He also serves as a Research Assistant at Academia Sinica, where he leads a team investigating Generative AI for image generation. In addition, he has a keen interest in Data Science and completed a year-long tenure as a Data Analyst Intern at Appier. During the internship, he engaged in client data analysis, feature engineering, business model development, and established key business metrics. His commitment to clean code architecture, design patterns, and high-quality programming was nurtured during his time at Appier. He possesses strong communication and presentation skills, honed through extensive cross-department collaboration. Passionate about contributing to innovation, he is eager to apply his skills to solve challenging problems in the technology industry and open to discussions about potential collaborations.

Linkedln: http://linkedin.com/in/jwliao1209
GitHub: https://github.com/jwliao1209
Jing-En Huang

As a mathematics student with a strong passion for artificial intelligence, I enjoy leveraging my mathematical knowledge to solve real‐world computer vision tasks through AI integration. For example, I have recently developed a real‐time automatic system for tracking and analyzing basketball matches using computer vision, deep learning models, and various mathematical techniques. As hobbies, I have served as the team leader in three AI Cup competitions, each with commendable results.
( https://angusbb.github.io/AngusHuang.github.io/ )