A Modified Automatic Image Stitching Model for Textile and Garment Research Area

Qi Zhang, Yihong Zhang, Jing XuAdvances 2013 · DOI

Abstract

This paper presents techniques for constructing panoramic image stitching from sequences of images captured by cameras from different angles in the garment and textile research area. The image stitching representation associates a transformation matrix with each input image. An improved SIFT (Scale Invariant Feature Transform) algorithm is used to locate feature points in scanned images, specifically targeting the problem of automatically stitching textile images scanned in divided pieces. The RANSAC (Random Sample Consensus) method is proposed to estimate image transformation parameters and find the best consensus with the data. To reduce accumulated registration errors, a global alignment (block adjustment) method is proposed for the entire image sequence, resulting in an optimally registered stitch. Additionally, a local alignment technique warps each image based on pairwise local registrations to compensate for motion parallax and other unmodeled distortions.

Introduction: The Challenge of Textile Imaging

In the textile and garment industry, high-resolution digital representations of large fabric samples or full garments are essential for detailed analysis and defect detection. However, these items are often too large for a single camera shot or standard scanner bed, requiring them to be captured in overlapping "tiles."

Stitching these tiles back together presents several technical hurdles. Traditional methods frequently struggle with scale and rotation differences between shots, as well as "accumulated error." In a long sequence of images, small misalignments in the initial pairs can "drift," resulting in massive gaps or overlaps by the end of the sequence. Furthermore, simple stitching often leaves "ghosting" (double images) or blurred edges where patterns fail to align perfectly—artifacts that are unacceptable for precise textile research.

Figure 2: Illustration of input tiles and the resulting alignment gaps (gap1, gap2, gap3) that occur without advanced adjustment techniques.

Methodology: The Stitching Pipeline

The proposed model builds upon a robust pipeline of computer vision techniques tailored for the consistency required in fabric research. The process begins with feature extraction using SIFT, which identifies unique keypoints robust to changes in zoom, rotation, and lighting.

Figure 1: The proposed automatic image stitching pipeline, from input images to final panoramic results.

As shown in the flowchart above, the system follows these critical steps:

  1. Feature Matching and RANSAC: To filter out "bad" matches (outliers), the system employs Random Sample Consensus (RANSAC). This algorithm iteratively tests small groups of matches to find the mathematical transformation (homography) that the majority of points agree on.
  2. Probabilistic Verification: A Bayesian framework calculates the posterior probability that an image match is correct, allowing the system to automatically recognize which images belong together in a sequence without human intervention.
  3. Global Alignment (Block Adjustment): To solve the drift problem, the authors perform a global optimization. Instead of matching images sequentially (A to B, B to C), they optimize the positions of all images simultaneously to minimize total error across the entire sequence.
  4. Local Alignment: Finally, a local warping technique is applied to compensate for small distortions caused by camera motion parallax or the physical unevenness of the fabric itself.

Experimental Results

The model was evaluated using a custom dataset of 57 high-resolution images (2272 × 1704 pixels) of a garment. The sequence was captured using a camera's automatic mode, meaning aperture, exposure, and flash varied between shots, testing the algorithm's robustness to illumination changes.

Despite the complexity of solving for 228 parameters across 57 images, the system registered the entire sequence in just 9 seconds on a 2.6GHz Pentium M processor. The final output was a 360° × 100° spherical panorama. Qualitative analysis demonstrated that while basic stitching leaves visible gaps and "ghosts" in fabric patterns, the proposed global and local alignment produces a visually seamless texture suitable for industrial application.

Conclusion

By combining SIFT-based feature matching with multi-stage global and local alignment, this research provides a robust solution for the textile industry's need for high-fidelity digital fabric representations. The system effectively handles the inconsistencies of consumer-grade hardware and the geometric challenges of large-scale garment scanning, paving the way for future automated defect detection systems.

Summary updated

Abstract

This paper presents some techniques for constructing panoramic image stitching from sequences of images captured by cameras from different angle in garment and textile research area. The image stitching representation associates a transformation matrix with each input image. In this paper, an improved SIFT(Scale Invariant Feature Transform) algorithm was used to locate the feature points in the scanned images targeting at the problem of automatically stitching of textile images which were scanned in divided pieces. RANSAC (random sample consensus) method is proposed to to estimate image transformation parameters and to find a solution that has the best consensus with the data. Techniques for estimating and rening camera focal lengths are also presented. In order to reduce accumulated registration errors, global alignment (block adjustment) method is proposed for the whole sequence of images, which results in an optimally registered image stitching. A local alignment technique is also developed which warps each image based on the results of pairwise local image registrations to compensate for small amounts of motion parallax introduced by translations of the camera and other unmodeled distortions.
BibTeX
@inproceedings{zhang2013modified,
  author = {Qi Zhang and Yihong Zhang and Jing Xu},
  title = {{A Modified Automatic Image Stitching Model for Textile and Garment Research Area}},
  booktitle = {Advances},
  year = {2013},
  doi = {10.2991/icmt-13.2013.85},
}
Figure