Pulabaigari Viswanath
Pulabaigari Viswanath has a total of 63 co-authors. Here are the top co-authors by number of shared publications:
- Boddi Eswara Reddy
JNTUA College of Engineering
7 shared publications - Thogarachetti Hitendra Sarma
Vasavi College of Engineering
7 shared publications - Chigarapalle Shobha Bindu
JNTUA College of Engineering
4 shared publications - Snehasis Mukherjee
Institution of Eminence Delhi NCR India
4 shared publications - Avinash Gautam
Department of Computer Science and Information Systems, Birla Institute of Technology and Science, Pilani, India
3 shared publications - Chandra Sekhar Vorugunti
Indian Institute of Information Technology, Sri City, India
3 shared publications
Latest Publications
- 2 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 742-751
TSOSVNet: Teacher-student collaborative knowledge distillation for Online Signature Verification
Online signature verification (OSV) is a standardized personal authentication scheme with wide social acceptance in critical real-time applications include access control, m-commerce, etc. Even though the current advances in Deep learning (DL) technologies catalysed state-of-the-art frameworks for challenging domains like computer vision, speech recognition, etc., the DL-based frameworks are voluminous with huge trainable parameters and are hard to deploy in real-time systems demanding faster inference. To adopt DL into OSV for improved performance, we propose an OSV framework made up of teacher-student collaborative knowledge distillation (TSKD) technique. A heavy Transformer based teacher is trained first and the teacher knowledge is distilled into a very lightweight Convolutional Neural Network (CNN) based student. A well trained teacher network results in an efficient deep representative feature learning by the student and results in a performance improvement. In a thorough set of experiments with three popular and standard datasets, i.e., the MCYT-100, SUSIG, and SVC, TSOSVNet framework, with a CNN based student model requiring only 3266 trainable parameters results in an EER of 12.42% compared to the recent SOTA 13.38% by a model with 206277 parameters in skilled_01 category of MCYT-100 dataset. In comparison to cutting-edge CNN-based OSV models, the proposed TSOSVNet produced a state-of-the-art EER in the most of the test categories with an average of 90% lesser trainable parameters.
- 2 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 742-751
TSOSVNet: Teacher-student collaborative knowledge distillation for Online Signature Verification
Online signature verification (OSV) is a standardized personal authentication scheme with wide social acceptance in critical real-time applications include access control, m-commerce, etc. Even though the current advances in Deep learning (DL) technologies catalysed state-of-the-art frameworks for challenging domains like computer vision, speech recognition, etc., the DL-based frameworks are voluminous with huge trainable parameters and are hard to deploy in real-time systems demanding faster inference. To adopt DL into OSV for improved performance, we propose an OSV framework made up of teacher-student collaborative knowledge distillation (TSKD) technique. A heavy Transformer based teacher is trained first and the teacher knowledge is distilled into a very lightweight Convolutional Neural Network (CNN) based student. A well trained teacher network results in an efficient deep representative feature learning by the student and results in a performance improvement. In a thorough set of experiments with three popular and standard datasets, i.e., the MCYT-100, SUSIG, and SVC, TSOSVNet framework, with a CNN based student model requiring only 3266 trainable parameters results in an EER of 12.42% compared to the recent SOTA 13.38% by a model with 206277 parameters in skilled_01 category of MCYT-100 dataset. In comparison to cutting-edge CNN-based OSV models, the proposed TSOSVNet produced a state-of-the-art EER in the most of the test categories with an average of 90% lesser trainable parameters.
- 25 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-10
OSVConTramer: A Hybrid CNN and Transformer based Online Signature Verification
The advances in Deep Learning (DL) resulted in the development Convolutional Neural Network (CNN) and Recurrent Neural Networks (RNN) based Online Signature Verification (OSV) frameworks. The main drawback with LSTM based networks is the limited parallelization of model training. The CNN based frameworks are efficient in learning local feature dependencies, but fail to apprehend long term feature dependencies. The current works confirmed the success of Transformer based models in long term time series classification (LTTSC) problems due to efficient capturing of context-dependent global feature interactions. Hence, to achieve higher classification accuracy, in this work, we propose a first of its kind of an attempt, in which, we combine CNN and Transformer for Online Signature Verification, named OSVConTramer. The proposed OSVConTramer efficiently learns optimal local and global dependencies of an input signature feature vector and outperforms previous CNN and LSTM based OSV frameworks achieving state-of-the-art classification accuracy. On the widely used MCYT-100, SVC, and SUSIG datasets, specific to one shot learning, our model achieves a SOTA EER of 10.85%, 5.45%, and 6.32%, respectively. The results of the experimental analysis confirms that the accuracy outcomes of OSV frameworks is improved significantly by the optimal learning of the relationships between local and global feature dependency.
- 25 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-10
OSVConTramer: A Hybrid CNN and Transformer based Online Signature Verification
The advances in Deep Learning (DL) resulted in the development Convolutional Neural Network (CNN) and Recurrent Neural Networks (RNN) based Online Signature Verification (OSV) frameworks. The main drawback with LSTM based networks is the limited parallelization of model training. The CNN based frameworks are efficient in learning local feature dependencies, but fail to apprehend long term feature dependencies. The current works confirmed the success of Transformer based models in long term time series classification (LTTSC) problems due to efficient capturing of context-dependent global feature interactions. Hence, to achieve higher classification accuracy, in this work, we propose a first of its kind of an attempt, in which, we combine CNN and Transformer for Online Signature Verification, named OSVConTramer. The proposed OSVConTramer efficiently learns optimal local and global dependencies of an input signature feature vector and outperforms previous CNN and LSTM based OSV frameworks achieving state-of-the-art classification accuracy. On the widely used MCYT-100, SVC, and SUSIG datasets, specific to one shot learning, our model achieves a SOTA EER of 10.85%, 5.45%, and 6.32%, respectively. The results of the experimental analysis confirms that the accuracy outcomes of OSV frameworks is improved significantly by the optimal learning of the relationships between local and global feature dependency.
- 15 April 2023
- Published by Springer Nature
- p. 659-670
Unsupervised Domain Adaptation Supplemented with Generated Images
- 15 April 2023
- Published by Springer Nature
- p. 659-670
Unsupervised Domain Adaptation Supplemented with Generated Images
- 25 November 2022
- Published by Springer Nature
- p. 83-97
Impact of Type of Convolution Operation on Performance of Convolutional Neural Networks for Online Signature Verification
- 25 November 2022
- Published by Springer Nature
- p. 83-97
Impact of Type of Convolution Operation on Performance of Convolutional Neural Networks for Online Signature Verification
- Open Access
Deep Model Compression based on the Training History
Deep Convolutional Neural Networks (DCNNs) have shown promising performances in several visual recognition problems which motivated the researchers to propose popular architectures such as LeNet, AlexNet, VGGNet, ResNet, and many more. These architectures come at a cost of high computational complexity and parameter storage. To get rid of storage and computational complexity, deep model compression methods have been evolved. We propose a "History Based Filter Pruning (HBFP)" method that utilizes network training history for filter pruning. Specifically, we prune the redundant filters by observing similar patterns in the filter's L1-norms (absolute sum of weights) over the training epochs. We iteratively prune the redundant filters of a CNN in three steps. First, we train the model and select the filter pairs with redundant filters in each pair. Next, we optimize the network to ensure an increased measure of similarity between the filters in a pair. This optimization of the network facilitates us to prune one filter from each pair based on its importance without much information loss. Finally, we retrain the network to regain the performance, which is dropped due to filter pruning. We test our approach on popular architectures such as LeNet-5 on MNIST dataset; VGG-16, ResNet-56, and ResNet-110 on CIFAR-10 dataset, and ResNet-50 on ImageNet. The proposed pruning method outperforms the state-of-the-art in terms of FLOPs reduction (floating-point operations) by 97.98%, 83.42%, 78.43%, 74.95%, and 75.45% for LeNet-5, VGG-16, ResNet-56, ResNet-110, and ResNet-50, respectively, while maintaining the less error rate.
Open AccessDeep Model Compression based on the Training History
Deep Convolutional Neural Networks (DCNNs) have shown promising performances in several visual recognition problems which motivated the researchers to propose popular architectures such as LeNet, AlexNet, VGGNet, ResNet, and many more. These architectures come at a cost of high computational complexity and parameter storage. To get rid of storage and computational complexity, deep model compression methods have been evolved. We propose a "History Based Filter Pruning (HBFP)" method that utilizes network training history for filter pruning. Specifically, we prune the redundant filters by observing similar patterns in the filter's L1-norms (absolute sum of weights) over the training epochs. We iteratively prune the redundant filters of a CNN in three steps. First, we train the model and select the filter pairs with redundant filters in each pair. Next, we optimize the network to ensure an increased measure of similarity between the filters in a pair. This optimization of the network facilitates us to prune one filter from each pair based on its importance without much information loss. Finally, we retrain the network to regain the performance, which is dropped due to filter pruning. We test our approach on popular architectures such as LeNet-5 on MNIST dataset; VGG-16, ResNet-56, and ResNet-110 on CIFAR-10 dataset, and ResNet-50 on ImageNet. The proposed pruning method outperforms the state-of-the-art in terms of FLOPs reduction (floating-point operations) by 97.98%, 83.42%, 78.43%, 74.95%, and 75.45% for LeNet-5, VGG-16, ResNet-56, ResNet-110, and ResNet-50, respectively, while maintaining the less error rate.
- 9 May 2022
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 81 (28) , 40431-40449
An information-rich sampling technique over spatio-temporal CNN for classification of human actions in videos
We propose a novel video sampling scheme for human action recognition in videos, using Gaussian Weighing Function. Traditionally in deep learning-based human activity recognition approaches, either a few random frames or every kth frame of the video is considered for training the 3D CNN, where k is a small positive integer, like 4, 5, or 6. This kind of sampling reduces the volume of the input data, which speeds-up the training network and also avoids overfitting to some extent, thus enhancing the performance of the 3D CNN model. In the proposed video sampling technique, consecutive k frames of a video are aggregated into a single frame by computing a Gaussian-weighted summation of the k frames. The resulting frame preserves the information in a better way than the conventional approaches and experimentally shown to perform better. In this paper, a 3-Dimensional deep CNN is proposed to extract the spatio-temporal features and follows Long Short-Term Memory (LSTM) to recognize human actions. The proposed 3D CNN architecture is capable of handling the videos where the camera is placed at a distance from the performer. Experiments are performed with KTH, WEIZMANN, and CASIA-B Human Activity and Gait datasets, whereby it is shown to outperform state-of-the-art deep learning based techniques. We achieve 95.78%, 95.27%, and 95.27% over the KTH, WEIZMANN, and CASIA-B human action and gait recognition datasets, respectively.
- 9 May 2022
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 81 (28) , 40431-40449
An information-rich sampling technique over spatio-temporal CNN for classification of human actions in videos
We propose a novel video sampling scheme for human action recognition in videos, using Gaussian Weighing Function. Traditionally in deep learning-based human activity recognition approaches, either a few random frames or every kth frame of the video is considered for training the 3D CNN, where k is a small positive integer, like 4, 5, or 6. This kind of sampling reduces the volume of the input data, which speeds-up the training network and also avoids overfitting to some extent, thus enhancing the performance of the 3D CNN model. In the proposed video sampling technique, consecutive k frames of a video are aggregated into a single frame by computing a Gaussian-weighted summation of the k frames. The resulting frame preserves the information in a better way than the conventional approaches and experimentally shown to perform better. In this paper, a 3-Dimensional deep CNN is proposed to extract the spatio-temporal features and follows Long Short-Term Memory (LSTM) to recognize human actions. The proposed 3D CNN architecture is capable of handling the videos where the camera is placed at a distance from the performer. Experiments are performed with KTH, WEIZMANN, and CASIA-B Human Activity and Gait datasets, whereby it is shown to outperform state-of-the-art deep learning based techniques. We achieve 95.78%, 95.27%, and 95.27% over the KTH, WEIZMANN, and CASIA-B human action and gait recognition datasets, respectively.
- 13 September 2021
- Published by Springer Nature in Sādhanā
- Vol. 46 (4) , 1-15
A multi-pronged accurate approach to optical character recognition, using nearest neighborhood and neural-network-based principles
- 13 September 2021
- Published by Springer Nature in Sādhanā
- Vol. 46 (4) , 1-15
A multi-pronged accurate approach to optical character recognition, using nearest neighborhood and neural-network-based principles
- 4 January 2021
- Published by Springer Nature in Neural Computing and Applications
- Vol. 33 (13) , 8055-8065
AutoFCL: automatically tuning fully connected layers for handling small dataset
- 4 January 2021
- Published by Springer Nature in Neural Computing and Applications
- Vol. 33 (13) , 8055-8065
AutoFCL: automatically tuning fully connected layers for handling small dataset
- 1 January 2021
- Published by Elsevier in Neural Networks
- Vol. 133, 112-122
AutoTune: Automatically Tuning Convolutional Neural Networks for Improved Transfer Learning
- 1 January 2021
- Published by Elsevier in Neural Networks
- Vol. 133, 112-122
AutoTune: Automatically Tuning Convolutional Neural Networks for Improved Transfer Learning