Sammendrag
Archaeological textile artifacts are often fragmented and human experts need to solve a puzzle to find matching pieces and reconstruct the original. This is a challenging task due to missing or degraded fragments, deformations and fading. Often it is not even known whether fragments come from the same object or not. A good example is the Oseberg Tapestry from the Viking Age [1]. Matching is usually done by experts who rely on technical analysis of the textile. One important criterion for identifying fragments from the same object is thread count (i.e. number of threads per centimeter) in the warp and weft directions. This is a tedious and time-consuming process. Computer science literature shows that current algorithms are not robust enough to solve the puzzle of archaeological textiles and proposes taking inspiration from the criteria that humans use [2-4]. To address this, we propose a framework for automatizing thread counting in textile images. We created a feature extractor, which captures thread count in vertical and horizontal directions. Thread count is usually measured by selecting several 1x1cm areas on the tapestries (we refer to this area as a patch), counting the threads and averaging the results. Archaeological textiles often have holes where threads are missing due to decomposition. The experts take this into account when they decide where to count. We modeled the thread count extraction as a pulse-counting process, inspired by how we intuitively count threads. For counting vertical threads, we trace the patch in the horizontal direction and count the thread by finding the number of occurrences of two consecutive vertical edges. Similarly for the horizontal threads we trace the patch in vertical directions. Our proposed thread flow begins with applying Otsu’s binary thresholding to the patch. Then we take all the pixel values for all columns in the same row to model the horizontal tracing of the patch. The pulse front is obtained by calculating the absolute approximate derivative of the trace. The number of pulses, which represent the number 𝑋(2: 𝑛) − 𝑋(1: 𝑛−1) of thread, is obtained by the number of non-zero elements of the absolute approximate derivative divided by two. This operation is repeated all the way from the top to bottom patch and the majority thread count is returned as the thread density of the patch. We compared our algorithm with the ground truth counted by experts. The preliminary results show that our method overestimates thread count. The final results will be presented at the conference. However, thread count features extracted with our method, provided promising results when used for clustering. Twelve out of 27 fragments clustered using thread count alone, were aligned with archaeologists’ hypotheses on which fragments belong together. The Oseberg Tapestry is a challenging case-study also because its thread systems consist of double threads, which introduces an additional complexity to the machine. The result for simpler scenarios will be also presented. Future work should explore more sophisticated methods for thresholding, such as Sobel to expose edges, and STRESS (Spatio-Temporal Retinex-inspired Envelope with Stochastic Sampling) enhancement preprocessing, to make them more separable.
Vis fullstendig beskrivelse
