Machine learning-based classification of fiber specklegram for pressure point detection
Date Issued
May 2025
Author(s)
Abstract
In this paper, we demonstrate a method for determining the pressure applied to an optical
fiber by analyzing its specklegram images using a neural network pipeline. In two steps, we
fine-tune EfficientNetB0: first, we train a small classifier on top of frozen layers, and then we
gently update the entire network with a decaying learning rate to produce stable 256-element
feature vectors. During training, we employ RandAugment, CutMix, and MixUp to expand
and diversify our data. These feature vectors are then fed into an XGBoost ensemble, and we
combine its output with the CNN’s softmax scores. On a hold-out test set of specklegrams,
this hybrid model achieves an accuracy of over 94% and demonstrates balanced performance
across all pressure classes. Our method is efficient and is running quickly on an A100 GPU.
Future work could extend this approach to predict exact force values using regression models
trained on more detailed datasets.
fiber by analyzing its specklegram images using a neural network pipeline. In two steps, we
fine-tune EfficientNetB0: first, we train a small classifier on top of frozen layers, and then we
gently update the entire network with a decaying learning rate to produce stable 256-element
feature vectors. During training, we employ RandAugment, CutMix, and MixUp to expand
and diversify our data. These feature vectors are then fed into an XGBoost ensemble, and we
combine its output with the CNN’s softmax scores. On a hold-out test set of specklegrams,
this hybrid model achieves an accuracy of over 94% and demonstrates balanced performance
across all pressure classes. Our method is efficient and is running quickly on an A100 GPU.
Future work could extend this approach to predict exact force values using regression models
trained on more detailed datasets.
File(s)![Thumbnail Image]()
Name
MichaelStavrou_2025_BSC-abstract.pdf
Size
303.76 KB
Format
Adobe PDF
Checksum (MD5)
4860833e35c1c71ac2e1f5f3dbf505f0