YOLO Image Segmentation
YOLO Image Segmentation
Embedded Files
Fine tuned the You Only Look Once (YOLO) Image segmentation model to distinguish between Compact binary coalescence signals and noise in the time-frequency data
Fine tuned the You Only Look Once (YOLO) Image segmentation model to distinguish between Compact binary coalescence signals and noise in the time-frequency data
Transient noise when present very close to the gravitational wave signals makes it nearly impossible for the gravitational waves to be recognized by the default matched filtering techniques. Here I fine tuned and trained a segmentation model to not only identifiy and classify signals vs noise but also localize the instances in the data by returning us accurate pixel masks of signals and noise in time-frequency spectrogtrams. Check out my work at https://arxiv.org/abs/2508.17399
Transient noise when present very close to the gravitational wave signals makes it nearly impossible for the gravitational waves to be recognized by the default matched filtering techniques. Here I fine tuned and trained a segmentation model to not only identifiy and classify signals vs noise but also localize the instances in the data by returning us accurate pixel masks of signals and noise in time-frequency spectrogtrams. Check out my work at https://arxiv.org/abs/2508.17399
Here a faint signal is identified ("Chirp") and localized in the data.
Slow Scatter Noise reduction
Slow Scatter Noise reduction
Applied data analysis to find the location of scatter noise coupling, which led to a noise reduction by a factor of 100. More details in the journal publication here.
Applied data analysis to find the location of scatter noise coupling, which led to a noise reduction by a factor of 100. More details in the journal publication here.
Stray light creates excess power in the gravitational wave detectors such as the one shown on the left. This noise was due to the reflection of stray light from the gold rings between the LIGO test mass mirrors shown on right.
Stray light creates excess power in the gravitational wave detectors such as the one shown on the left. This noise was due to the reflection of stray light from the gold rings between the LIGO test mass mirrors shown on right.
In this Project, I worked on understanding the cause of this noise which led to its source. Once the noise source was fixed, there was a dramatic reduction in the rate of noise.
In this Project, I worked on understanding the cause of this noise which led to its source. Once the noise source was fixed, there was a dramatic reduction in the rate of noise.
The figure on the right shows noise rate before the fix in blue and after the fix in red
The figure on the right shows noise rate before the fix in blue and after the fix in red
Using CNNs to identify noise in LIGO
Using CNNs to identify noise in LIGO
GravitySpy is a ML based noise recognition tool. I trained GravitySpy with image dataset of a new source of noise. This led to recognition of Fast Scatter noise in the data. The figure on the right shows the old classification vs the new classification for the two detectors. More details in the paper.
GravitySpy is a ML based noise recognition tool. I trained GravitySpy with image dataset of a new source of noise. This led to recognition of Fast Scatter noise in the data. The figure on the right shows the old classification vs the new classification for the two detectors. More details in the paper.
I analyzed the Fast Scatter noise and identified two different sub population of the noise. The consequent modeling led to its source identification and noise reduction. The figure on the right compares the noise rate during logging (tree cutting) before and after the fix. More details in the paper.
I analyzed the Fast Scatter noise and identified two different sub population of the noise. The consequent modeling led to its source identification and noise reduction. The figure on the right compares the noise rate during logging (tree cutting) before and after the fix. More details in the paper.
Noise vs Signal Classifier
Noise vs Signal Classifier
Developed an algorithm to distinguish noise from signal based on the energy in the Q-transform of the time-series.
Developed an algorithm to distinguish noise from signal based on the energy in the Q-transform of the time-series.
The image on the right shows a 4 sec Q-transform of the time-series data. This. Q transform is divided into 15 blocks across time and frequency. Depending on how much energy is in a given block, a signal can be classified as noise or real event or real event + noise. More details on this multiclassification algorithm in this paper.
The image on the right shows a 4 sec Q-transform of the time-series data. This. Q transform is divided into 15 blocks across time and frequency. Depending on how much energy is in a given block, a signal can be classified as noise or real event or real event + noise. More details on this multiclassification algorithm in this paper.
Page updated
Google Sites
Report abuse