<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Yunyeong Ju | DAIS Research Group</title><link>https://dais-research.github.io/en/authors/yunyeong-ju/</link><atom:link href="https://dais-research.github.io/en/authors/yunyeong-ju/index.xml" rel="self" type="application/rss+xml"/><description>Yunyeong Ju</description><generator>HugoBlox Kit (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Fri, 06 Nov 2026 00:00:00 +0000</lastBuildDate><item><title>Brightness-Band Projection for Residual Silicon Segmentation in C/SiC Composites</title><link>https://dais-research.github.io/en/publications/ack2026-csic-si-seg/</link><pubDate>Fri, 06 Nov 2026 00:00:00 +0000</pubDate><guid>https://dais-research.github.io/en/publications/ack2026-csic-si-seg/</guid><description>
&lt;blockquote class="border-l-4 border-neutral-300 dark:border-neutral-600 pl-4 italic text-neutral-600 dark:text-neutral-400 my-6"&gt;
&lt;p&gt;🏆 &lt;strong&gt;Industry-Academia-Research Collaboration Excellence Award&lt;/strong&gt; — ACK 2026 (Annual Conference of the Korea Information Processing Society), November 6, 2026, Sejong University, Seoul. The paper is written in Korean.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;div class="grid gap-6 md:grid-cols-2 mt-6 mb-8"&gt;
&lt;figure&gt;
&lt;div class="flex h-72 md:h-96 items-center justify-center"&gt;
&lt;img src="https://dais-research.github.io/media/gallery/ack2026-award.webp"
alt="Yunyeong Ju holding the award certificate in front of the ACK 2026 banner"
class="rounded-lg shadow-md max-h-full max-w-full object-contain" /&gt;
&lt;/div&gt;
&lt;figcaption class="text-sm text-center mt-2 opacity-75"&gt;
After the ACK 2026 award — Yunyeong Ju
&lt;/figcaption&gt;
&lt;/figure&gt;
&lt;figure&gt;
&lt;div class="flex h-72 md:h-96 items-center justify-center"&gt;
&lt;img src="https://dais-research.github.io/media/gallery/ack2026-certificate.webp"
alt="ACK 2026 Industry-Academia-Research Collaboration Excellence Award certificate"
class="rounded-lg shadow-md max-h-full max-w-full object-contain bg-white" /&gt;
&lt;/div&gt;
&lt;figcaption class="text-sm text-center mt-2 opacity-75"&gt;
Award certificate
&lt;/figcaption&gt;
&lt;/figure&gt;
&lt;/div&gt;
&lt;h2 id="background"&gt;Background&lt;/h2&gt;
&lt;p&gt;C/SiC ceramic composites produced by liquid silicon infiltration (LSI) retain
&lt;strong&gt;residual Si&lt;/strong&gt; in the matrix. Residual Si lowers both the maximum service
temperature and the mechanical strength, so its volume fraction has to be
measured and controlled for each set of process conditions. In practice it is
measured as an area fraction on cross-sectional SEM images.&lt;/p&gt;
&lt;p&gt;The measurement itself is the bottleneck. An operator picks a brightness
threshold per image in a labeling tool and edits the result with a brush;
because the brightness distribution differs from image to image, the threshold
has to be retuned every time. Throughput and reproducibility both suffer.&lt;/p&gt;
&lt;p&gt;Existing methods do not replace this. A threshold chosen automatically (Otsu and
the like) differs from the choice an operator would make and cannot express the
manual edits. A generic segmentation network, conversely, learns spatial features
but &lt;strong&gt;does not reflect in its output the fact that the labels were created by a
brightness rule in the first place.&lt;/strong&gt; Even when the network locates the Si
correctly, its probability-map boundary is off by pixels from the label
boundary, which is a brightness contour.&lt;/p&gt;
&lt;h2 id="method--soft-band-projection"&gt;Method — soft band projection&lt;/h2&gt;
&lt;p&gt;The idea is to put the label-creation rule back in at the last step, where the
mask is decided. For each image we pick the brightness interval [a, b] whose
mask overlaps most with the shape drawn by the network probability map P, and
emit that interval&amp;rsquo;s mask as the output.&lt;/p&gt;
&lt;figure class="mt-6 mb-8"&gt;
&lt;img src="https://dais-research.github.io/media/publications/ack2026-band-projection.webp"
alt="Soft band projection: SEM input, network probability map, candidate score map, projected output"
class="rounded-lg shadow-md w-full bg-white" /&gt;
&lt;figcaption class="text-sm text-center mt-2 opacity-75"&gt;
Soft band projection. The overlap J between the probability map P and the
mask of each candidate interval [a, b] is scored over all candidates, and the
maximizing interval ([202, 236] here) is emitted. Figure labels are in Korean.
&lt;/figcaption&gt;
&lt;/figure&gt;
&lt;p&gt;Three components:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Base segmentation&lt;/strong&gt; — U-Net++ with an ImageNet-pretrained EfficientNet-B5
encoder. The input is linearly normalized so that the 2nd and 99th percentiles
of brightness inside the analysis region map to 0 and 255 (applying one shared
threshold to all images gives IoU 0.55 on raw brightness and 0.80 after
normalization). The loss is BCE + Dice plus soft-clDice weighted 0.5, which
accounts for the connectivity of thin Si structures. Probability maps from five
seeds are averaged.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Band candidates&lt;/strong&gt; — a lower bound alone would admit every bright pixel above
the threshold, so an upper bound is needed: pixels excluded by hand have a
median brightness of 251, brighter than the 217 of the Si that was kept.
Candidates are all integer pairs 150 ≤ a ≤ b ≤ 255, covering the 178–255 range
of thresholds actually used in practice.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Soft score&lt;/strong&gt; — binarizing the probability map at 0.5 before measuring overlap
wipes out the signal on images where Si is sparse, so candidates are scored by a
soft IoU against the un-binarized P.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="results"&gt;Results&lt;/h2&gt;
&lt;p&gt;33 images from six specimens were evaluated with &lt;strong&gt;specimen-wise 6-fold
cross-validation&lt;/strong&gt;, holding out one specimen at a time so that no specimen
appears in both training and test.&lt;/p&gt;
&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Method&lt;/th&gt;
&lt;th&gt;IoU&lt;/th&gt;
&lt;th&gt;Si area-fraction error (%p)&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Otsu (binary)&lt;/td&gt;
&lt;td&gt;0.220&lt;/td&gt;
&lt;td&gt;33.37&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GMM, 8 components&lt;/td&gt;
&lt;td&gt;0.786&lt;/td&gt;
&lt;td&gt;2.96&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fixed-threshold baseline&lt;/td&gt;
&lt;td&gt;0.804&lt;/td&gt;
&lt;td&gt;2.87&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;U-Net++ (base)&lt;/td&gt;
&lt;td&gt;0.793&lt;/td&gt;
&lt;td&gt;0.94&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;+ single-threshold projection&lt;/td&gt;
&lt;td&gt;0.842&lt;/td&gt;
&lt;td&gt;0.71&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;+ soft-clDice&lt;/td&gt;
&lt;td&gt;0.861&lt;/td&gt;
&lt;td&gt;0.65&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;+ band projection &amp;amp; ensemble&lt;/td&gt;
&lt;td&gt;0.870&lt;/td&gt;
&lt;td&gt;0.55&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;+ soft objective (proposed)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;0.892&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;0.51&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;oracle-t (single-threshold upper bound, reference)&lt;/td&gt;
&lt;td&gt;0.977&lt;/td&gt;
&lt;td&gt;0.22&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;
&lt;ul&gt;
&lt;li&gt;The mean absolute area-fraction error is 2.87 %p for the fixed-threshold
baseline and 0.51 %p for the proposed method. The worst-case error is 24.0 %p
for the baseline and 1.8 %p for the proposed method, whose median per-image IoU
is 0.958.&lt;/li&gt;
&lt;li&gt;Across &lt;strong&gt;40 configurations of classical methods&lt;/strong&gt; applied without training, the
best IoU was 0.786 (8-component GMM), below even the fixed-threshold baseline at
0.804. On a cross-section of three phases — carbon, SiC and Si — five binary
thresholding methods also picked up SiC as Si, staying under IoU 0.24.&lt;/li&gt;
&lt;li&gt;Bootstrapping over the six specimens gives 95 % confidence intervals of
0.809–0.951 for mean IoU and 0.33–0.64 %p for mean area-fraction error.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;How far manual edits are reproduced&lt;/strong&gt; — with an upper bound b, the band ruled
out 71.6 % of the 41,395 pixels excluded by hand (0.7 % for the control with b
fixed at 255). Conversely, the 76,293 pixels added by hand — dark because of
polishing cracks — sit below the lower bound a, so although the probability map
calls 50.4 % of them Si, only 19.2 % survive the projection.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id="limitations-and-future-work"&gt;Limitations and future work&lt;/h2&gt;
&lt;p&gt;31 of the 33 images had an area-fraction error of 1.34 %p or less, so the
predicted area fraction of a new specimen image falls within ±1.34 %p with about
90 % probability. The three images with low IoU have Si covering only about 1 % of
the analysis region, where the small pixel count makes boundary error dominate IoU.&lt;/p&gt;
&lt;p&gt;Four follow-ups: (1) a model that restores Si darkened by polishing cracks and
therefore dropped by the projection; (2) a detector for foreign particles whose
brightness overlaps Si; (3) an explicit expert criterion for the fine segregation
between carbon bundles; (4) extending automatic analysis-region estimation with
text recognition on the information bar and boundary detection for non-rectangular
specimens. Beyond that, we aim to build a system that rapidly produces reference
labels reflecting expert review and editing criteria.&lt;/p&gt;
&lt;h2 id="acknowledgments"&gt;Acknowledgments&lt;/h2&gt;
&lt;p&gt;This work was supported by the National Research Foundation of Korea, funded by
the Ministry of Science and ICT, under the Nano &amp;amp; Material Technology Development
Program (Establishment of database hub of ultra-high temperature heat-resistant
ceramic materials for space launch vehicles based on AI exploration design and
full-cycle data management, No. RS-2025-25442774).&lt;/p&gt;</description></item></channel></rss>