Demachi Laboratory

Research of

Nuclear Security

Nuclear Security & Development of Insiders' Sabotage Detection Method

Research of

​Medical Application

2017 Internation Conference on Physical Protection of Nuclear Material and Nuclear Facilities (11/13-11/17)-Shi Chen(D1)

Graduation Thesis: Yutaro Nakajima (B4)

Graduation Thesis: Ryo Kubota (B4)

Resilience Assessment for Nuclear Facility based on Information-Theoretic Method

The Slide of Laboratory Meeting(4/15)-Jonathan Poli(M2)

Development of Early Abnormal Detection in Dynamic Equipment

The Slide of Laboratory Meeting(4/15)-Naoko Inadome(M2)

Graduation Thesis: Reiji Terayama (B4)

Development of Movie Prediction Method

What is the movie prediction method ?

👉Currently, there are numerous systems using video, but studies that make a prediction for the video itself are rare. In Demachi laboratory, we captured the video as time-series data, then analyzed multivariate time series data by applying a capable MSSA method and principal component analysis. Later, video prediction method was developed which enabled an output of a future movie.

 

Overview of movie prediction method

What is PCA?

👉Statistical methods are widely used in the field of image processing, and one of the typical methods amongst them is PCA (Principal Component Analysis).
Principal Component Analysis uses an orthogonal transformation to convert a set of observations from correlated variables into a set of values of linearly uncorrelated variables called principal components. This principal components eliminate the correlation amongst the variables from multivariate data analysis technique, which attempts to characterize the original measurements from the main component.

​Image of PCA

PCA and face recognition

 Most famous application example in the field of image processing of principal component analysis is the recognition of face image.
 In making the recognition of the general face image, rather than using the knowledge of specific components in the face, such as eyes and nose, the general method is to treat the entire face image as a pattern and using a statistical pattern recognition technique.
 However, there is a problem that the dimension of the pattern becomes enormous if face image is handled as a pattern.
 One of the solutions is the Principal Component Analysis (PCA). By using PCA, high-precision face recognition can be implemented by the compressed information of face image pattern.

The example of face recognition (Left: Face images used as training set 

Right:Principal components) [1]

What is MSSA ?

The SSA method, time series data to retrieve the principal component by decomposing the singular values, is a technique that predicts time-series data using the data reconstructed by the principal components.
One of the major features of the SSA method is that this is not limited to the movement of the periodic ones, but it could also be used to predict and show related content of the chaotic behavior, such as air temperature and seawater temperature.
MSSA method is an extension of the SSA method to multidimensional.

 

Algorithm of MSSA method

Algorithm For Image Predicion

 In this method, the movie data is expressed as a linear combination of principal components and coefficients using principal component analysis, then predict coefficients in MSSA method and output future movie data.
 That is, by reducing dimension using principal component analysis, it allows the prediction of the complex data of movie from a prediction of only by the coefficient, which is the emphasis of the proposed method.

Algorithm for Image Prediction

Prediction Result

 We applied this technique to a variety of subjects and attempted to predict.
Here we introduce the prediction result of a hand and with a pedestrian in the video data.

​Video Description (click on "i" button on the video screen for more info)

The prediction result of hands (prediction of 50 pieces of movie data with a lead time of 5 seconds)
※Left: the original image, middle: the predicted image, right: the difference

The Deformation prediction result of human (prediction of 60 pieces of movie data with a lead time of 5 seconds)
※Left: the original image, right: the predicted image

The walking prediction result of human (prediction of 60 pieces of movie data with a lead time of 5 seconds)
※Left: the original image, right: the predicted image

Reference

[1] M.Turk and A.Pentland: Eigenfaces for recognition, Journal of Cognitive Neuroscience, Vol.3, No.1, pp.71-86, 1991.

Movie Prediction for Lung Tumor Tracking in Radiation Therapy

The Slide of Laboratory Meeting(4/15)-Ritu Bhusal Chhatkuli(Researcher)

Pre-defense: Kenichiro Kuwabara (M2)

Pre-defense: Maoxin Tang (M2)

Flaw Depth Identification from Eddy Current Testing Signal by Deep Learning

Graduation Thesis: Tomoyuki Hori (M2)

 
 
 
 
 

© 2018 by Demachi Laboratory. Created with Wix.com