Deep transfer learning for star cluster classification: I. application to the PHANGS–HST survey

ABSTRACT We present the results of a proof-of-concept experiment that demonstrates that deep learning can successfully be used for production-scale classification of compact star clusters detected in Hubble Space Telescope(HST) ultraviolet-optical imaging of nearby spiral galaxies ($D\lesssim 20\, \textrm{Mpc}$) in the Physics at High Angular Resolution in Nearby GalaxieS (PHANGS)–HST survey. Given the relatively small nature of existing, human-labelled star cluster samples, we transfer the knowledge of state-of-the-art neural network models for real-object recognition to classify star clusters candidates into four morphological classes. We perform a series of experiments to determine the dependence of classification performance on neural network architecture (ResNet18 and VGG19-BN), training data sets curated by either a single expert or three astronomers, and the size of the images used for training. We find that the overall classification accuracies are not significantly affected by these choices. The networks are used to classify star cluster candidates in the PHANGS–HST galaxy NGC 1559, which was not included in the training samples. The resulting prediction accuracies are 70 per cent, 40 per cent, 40–50 per cent, and 50–70 per cent for class 1, 2, 3 star clusters, and class 4 non-clusters, respectively. This performance is competitive with consistency achieved in previously published human and automated quantitative classification of star cluster candidate samples (70–80 per cent, 40–50 per cent, 40–50 per cent, and 60–70 per cent). The methods introduced herein lay the foundations to automate classification for star clusters at scale, and exhibit the need to prepare a standardized data set of human-labelled star cluster classifications, agreed upon by a full range of experts in the field, to further improve the performance of the networks introduced in this study.

Download Full-text

A Novel Convolutional Neural Network Architecture for SAR Target Recognition

Journal of Sensors ◽

10.1155/2019/1246548 ◽

2019 ◽

Vol 2019 ◽

pp. 1-9 ◽

Cited By ~ 5

Author(s):

Yinjie Xie ◽

Wenxin Dai ◽

Zhenxin Hu ◽

Yijing Liu ◽

Chuan Li ◽

...

Keyword(s):

Neural Network ◽

Convolutional Neural Network ◽

Network Architecture ◽

Target Recognition ◽

Automatic Target Recognition ◽

Neural Network Architecture ◽

Data Set ◽

Stationary Target ◽

Optical Images

Among many improved convolutional neural network (CNN) architectures in the optical image classification, only a few were applied in synthetic aperture radar (SAR) automatic target recognition (ATR). One main reason is that direct transfer of these advanced architectures for the optical images to the SAR images easily yields overfitting due to its limited data set and less features relative to the optical images. Thus, based on the characteristics of the SAR image, we proposed a novel deep convolutional neural network architecture named umbrella. Its framework consists of two alternate CNN-layer blocks. One block is a fusion of six 3-layer paths, which is used to extract diverse level features from different convolution layers. The other block is composed of convolution layers and pooling layers are mainly utilized to reduce dimensions and extract hierarchical feature information. The combination of the two blocks could extract rich features from different spatial scale and simultaneously alleviate overfitting. The performance of the umbrella model was validated by the Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark data set. This architecture could achieve higher than 99% accuracy for the classification of 10-class targets and higher than 96% accuracy for the classification of 8 variants of the T72 tank, even in the case of diverse positions located by targets. The accuracy of our umbrella is superior to the current networks applied in the classification of MSTAR. The result shows that the umbrella architecture possesses a very robust generalization capability and will be potential for SAR-ART.

Download Full-text

Neural Network Analysis of Dynamic Fracture in a Layered Material

MRS Advances ◽

10.1557/adv.2018.673 ◽

2019 ◽

Vol 4 (19) ◽

pp. 1109-1117 ◽

Cited By ~ 2

Author(s):

Pankaj Rajak ◽

Rajiv K. Kalia ◽

Aiichiro Nakano ◽

Priya Vashishta

Keyword(s):

Neural Network ◽

Dynamic Fracture ◽

Network Architecture ◽

Md Simulation ◽

Structural Phase ◽

Feature Learning ◽

Training Data ◽

Model Complexity ◽

Feed Forward Neural Network ◽

Data Set

AbstractDynamic fracture of a two-dimensional MoWSe2 membrane is studied with molecular dynamics (MD) simulation. The system consists of a random distribution of WSe2 patches in a pre-cracked matrix of MoSe2. Under strain, the system shows toughening due to crack branching, crack closure and strain-induced structural phase transformation from 2H to 1T crystal structures. Different structures generated during MD simulation are analyzed using a three-layer, feed-forward neural network (NN) model. A training data set of 36,000 atoms is created where each atom is represented by a 50-dimension feature vector consisting of radial and angular symmetry functions. Hyper parameters of the symmetry functions and network architecture are tuned to minimize model complexity with high predictive power using feature learning, which shows an increase in model accuracy from 67% to 95%. The NN model classifies each atom in one of the six phases which are either as transition metal or chalcogen atoms in 2H phase, 1T phase and defects. Further t-SNE analyses of learned representation of these phases in the hidden layers of the NN model show that separation of all phases become clearer in the third layer than in layers 1 and 2.

Download Full-text

Neural Network Interpretation of Ultrasonic Response for Concrete Condition Assessment

Transportation Research Record Journal of the Transportation Research Board ◽

10.1177/0361198196153200112 ◽

1996 ◽

Vol 1532 (1) ◽

pp. 80-86

Author(s):

Samir N. Shoukry ◽

D.R. Martinelli

Keyword(s):

Neural Network ◽

Concrete Structures ◽

Neural Nets ◽

Training Data ◽

Neural Network Modeling ◽

Correct Classification ◽

Data Set ◽

Ultrasonic Signals ◽

The Individual

Ultrasonic testing of concrete structures using the pitch-catch method is an effective technique for testing concrete structures that cannot be accessed on two opposing surfaces. However, the ultrasonic signals so measured are extremely noisy and contain a complicated pattern of multiple frequency-coupled reflections that makes interpretation a difficult task. In this investigation, a neural network modeling approach is used to classify ultrasonically tested concrete specimens into one of two classes: defective or nondefective. Different types of neural nets are used, and their performance is evaluated. It was found that correct classification of the individual ultrasonic signals could be achieved with an accuracy of 75 percent for the test set and 95 percent for the training set. These recognition rates lead to the correct classification of all the individual test specimens. The study shows that although some neural net architectures may show high performance using a particular training data set, their results might not be consistent. In this paper, the consistency of the network performance was tested by shuffling the training and testing data sets.

Download Full-text

Ultrasonic Acoustic Health Monitoring of Ball Bearings Using Neural Network Pattern Classification of Power Spectral Density

2010 Joint Rail Conference, Volume 2 ◽

10.1115/jrc2010-36240 ◽

2010 ◽

Cited By ~ 1

Author(s):

William Kirchner ◽

Steve Southward ◽

Mehdi Ahmadian

Keyword(s):

Neural Network ◽

Artificial Neural Network ◽

Health Monitoring ◽

Training Data ◽

Ultrasonic Frequency ◽

Ball Bearings ◽

Frequency Range ◽

Data Set ◽

Artificial Neural

This work presents a generic passive non-contact based acoustic health monitoring approach using ultrasonic acoustic emissions (UAE) to facilitate classification of bearing health via neural networks. This generic approach is applied to classifying the operating condition of conventional ball bearings. The acoustic emission signals used in this study are in the ultrasonic range (20–120 kHz), which is significantly higher than the majority of the research in this area thus far. A direct benefit of working in this frequency range is the inherent directionality of microphones capable of measurement in this range, which becomes particularly useful when operating in environments with low signal-to-noise ratios that are common in the rail industry. Using the UAE power spectrum signature, it is possible to pose the health monitoring problem as a multi-class classification problem, and make use of a multi-layer artificial neural network (ANN) to classify the UAE signature. One major problem limiting the usefulness of ANN’s for failure classification is the need for large quantities of training data. This becomes a particularly important issue when considering applications involving higher value components such as the turbo mechanisms and traction motors on diesel locomotives. Artificial training data, based on the statistical properties of a significantly smaller experimental data set is created to train the artificial neural network. The combination of the artificial training methods and ultrasonic frequency range being used results in an approach generic enough to suggest that this particular method is applicable to a variety of systems and components where persistent UAE exist.

Download Full-text

Multi-attribute self-attention guided vehicle local region detection based on convolutional neural network architecture

International Journal of Advanced Robotic Systems ◽

10.1177/1729881420944343 ◽

2020 ◽

Vol 17 (4) ◽

pp. 172988142094434

Author(s):

Jingbo Chen ◽

Shengyong Chen ◽

Linjie Bian

Keyword(s):

Neural Network ◽

Convolutional Neural Network ◽

Quantitative Evaluation ◽

Network Architecture ◽

Local Region ◽

Training Data ◽

Detection Methods ◽

Data Set ◽

Region Detection ◽

Bounding Boxes

Many pieces of information are included in the front region of a vehicle, especially in windshield and bumper regions. Thus, windshield or bumper region detection is making sense to extract useful information. But the existing windshield and bumper detection methods based on traditional artificial features are not robust enough. Those features may become invalid in many real situations (e.g. occlude, illumination change, viewpoint change.). In this article, we propose a multi-attribute-guided vehicle discriminately region detection method based on convolutional neural network and not rely on bounding box regression. We separate the net into two branches, respectively, for identification (ID) and Model attributes training. Therefore, the feature spaces of different attributes become more independent. Additionally, we embed a self-attention block into our framework to improve the performance of local region detection. We train our model on PKU_VD data set which has a huge number of images inside. Furthermore, we labeled the handcrafted bounding boxes on 5000 randomly picked testing images, and 1020 of them are used for evaluation and 3980 as the training data for YOLOv3. We use Intersection over Union for quantitative evaluation. Experiments were conducted in three different latest convolutional neural network trunks to illustrate the detection performance of the proposed method. Simultaneously, in terms of quantitative evaluation, the performance of our method is close to YOLOv3 even without handcrafted bounding boxes.

Download Full-text

Performance Evaluation of Neural Networks in Concrete Condition Assessment

Transportation Research Record Journal of the Transportation Research Board ◽

10.3141/1739-10 ◽

2000 ◽

Vol 1739 (1) ◽

pp. 76-82 ◽

Cited By ~ 1

Author(s):

D. R. Martinelli ◽

Samir N. Shoukry

Keyword(s):

Neural Network ◽

High Performance ◽

Neural Nets ◽

Training Data ◽

Neural Network Modeling ◽

Data Sets ◽

Neural Net ◽

Correct Classification ◽

Data Set

A neural network modeling approach is used to identify concrete specimens that contain internal cracks. Different types of neural nets are used and their performance is evaluated. Correct classification of the signals received from a cracked specimen could be achieved with an accuracy of 75 percent for the test set and 95 percent for the training set. These recognition rates lead to the correct classification of all the individual test specimens. Although some neural net architectures may show high performance with a particular training data set, their results might be inconsistent. In situations in which the number of data sets is small, consistent performance of a neural network may be achieved by shuffling the training and testing data sets.

Download Full-text

Robust Approach to Supervised Deep Neural Network Training for Real-Time Object Classification in Cluttered Indoor Environment

Applied Sciences ◽

10.3390/app11157148 ◽

2021 ◽

Vol 11 (15) ◽

pp. 7148

Author(s):

Bedada Endale ◽

Abera Tullu ◽

Hayoung Shi ◽

Beom-Soo Kang

Keyword(s):

Neural Network ◽

Deep Learning ◽

Real Time ◽

Network Architecture ◽

Input Data ◽

Deep Neural Network ◽

Data Augmentation ◽

Object Classification ◽

Training Data ◽

Gradient Descent Algorithm

Unmanned aerial vehicles (UAVs) are being widely utilized for various missions: in both civilian and military sectors. Many of these missions demand UAVs to acquire artificial intelligence about the environments they are navigating in. This perception can be realized by training a computing machine to classify objects in the environment. One of the well known machine training approaches is supervised deep learning, which enables a machine to classify objects. However, supervised deep learning comes with huge sacrifice in terms of time and computational resources. Collecting big input data, pre-training processes, such as labeling training data, and the need for a high performance computer for training are some of the challenges that supervised deep learning poses. To address these setbacks, this study proposes mission specific input data augmentation techniques and the design of light-weight deep neural network architecture that is capable of real-time object classification. Semi-direct visual odometry (SVO) data of augmented images are used to train the network for object classification. Ten classes of 10,000 different images in each class were used as input data where 80% were for training the network and the remaining 20% were used for network validation. For the optimization of the designed deep neural network, a sequential gradient descent algorithm was implemented. This algorithm has the advantage of handling redundancy in the data more efficiently than other algorithms.

Download Full-text

Study on Consistency Analysis in Text Categorization

Applied Mechanics and Materials ◽

10.4028/www.scientific.net/amm.539.181 ◽

2014 ◽

Vol 539 ◽

pp. 181-184

Author(s):

Wan Li Zuo ◽

Zhi Yan Wang ◽

Ning Ma ◽

Hong Liang

Keyword(s):

Text Categorization ◽

Training Data ◽

Experimental Result ◽

Final Decision ◽

Consistency Analysis ◽

Training Set ◽

Weak Classifier ◽

Data Set ◽

Basic Premise

Accurate classification of text is a basic premise of extracting various types of information on the Web efficiently and utilizing the network resources properly. In this paper, a brand new text classification method was proposed. Consistency analysis method is a type of iterative algorithm, which mainly trains different classifiers (weak classifier) by aiming at the same training set, and then these classifiers will be gathered for testing the consistency degrees of various classification methods for the same text, thus to manifest the knowledge of each type of classifier. It main determines the weight of each sample according to the fact is the classification of each sample is accurate in each training set, as well as the accuracy of the last overall classification, and then sends the new data set whose weight has been modified to the subordinate classifier for training. In the end, the classifier gained in the training will be integrated as the final decision classifier. The classifier with consistency analysis can eliminate some unnecessary training data characteristics and place the key words on key training data. According to the experimental result, the average accuracy of this method is 91.0%, while the average recall rate is 88.1%.

Download Full-text