convolutional neural network (cnn)

One of the biometric detection methods is to identify people based on speech signals. The implementation of a speaker identification (SI) system can be done in many different ways, and recently, many researchers have been focusing on using deep neural networks. One of the types of deep neural networks is recurrent neural networks, where memory and recurrent parts are handled by layers such as LSTM or Gated Recurrent Unit (GRU). In this paper, we propose a new structure as a classifier in the speaker identification system, which significantly improves the recognition rate by combining a convolutional neural network with two layers of GRU (CNN+ GRU). MFCC coefficients that have been extracted as cell arrays from each period of Pt speech will be used as sequence vectors for the input of proposed classifier.  The performance of the SI system has improved in comparison to basic methods according to experiments conducted on two databases, LibriSpeech and VoxCeleb1. When Pt is longer, the system performs better, so that on the LibriSpeech database with 251 speakers, recognition accuracy is equal to 92.94% for Pt=1s, and it rises to 99.92% for Pt=9s. The proposed CNN+GRU classifier has a low sensitivity to specific genders, which can be said to be almost zero.

A brain-computer interface (BCI) is a form of assistive technology that facilitates communication between users and machines by interpreting brain signals. The P300 wave, an event-related potential (ERP) in oddball paradigms, is generated approximately 300 milliseconds after the presentation of the target stimulus selected by the user in the brain. Accurate recognition of these waves in a P300 spelling system enables the user to write letters. Classification of P300 waves in an EEG-based spelling system faces several challenges, including accurate detection of P300 waves and handling the high dimensionality of these signals. This study presents a Convolutional Deep Learning Framework (CDLF) for character recognition using EEG signals. The proposed model uses CNN with a one-dimensional kernel to extract features over time. The proposed model was applied to two public datasets: BCI Competition III dataset II and BCI Competition II dataset IIb. The proposed model showed an average character recognition rate of 95% at epoch 15 for BCI Competition III dataset II and 100% at epoch 15 for BCI Competition II dataset IIb, without using any feature and channel selection methods before classification. The proposed model is promising for brain-computer interface classification applications in the spelling domain.

Signature identification plays an important role in many areas such as banking, administrative and judicial systems. For this purpose, in this paper, an automatic intelligent framework is developed by combining a deep pre-trained network with a recurrent neural network. The results of the proposed model were evaluated on several valid datasets and collected datasets. Since there was no suitable Persian signature dataset, we collected a Persian signature dataset based on US ASTM guidelines and standards, which can be very effective and profound for deep approaches. Due to the very promising results of the proposed model in comparison with recent studies and conventional methods, to evaluate the resistance of the proposed model to different noises, we added Gaussian Noise, Salt and Pepper Noise, Speckle Noise, and Local var Noise in different SNRs to the raw data. The results show that the proposed model can still be resistant to a wide range of SNRs; So at 15 dB, the accuracy of the proposed method is still above 90%.

Diagnosing benign and malignant glands in thyroid ultrasound images is considered as a challenging issue. Recently, deep learning techniques have significantly resulted in extracting features from medical images and classifying them. Convolutional networks ignore the hierarchical structure of entities within images and do not pay attention to spatial information as well as the need for a large number of training samples. Capsule networks consist of different hierarchical capsules equivalent to the same layers in the CNN neural network. This study tried to extract textural features using a deep learning model based on a capsule network. Thyroid ultrasound images were given to the capsule network as input data, and finally the features learned in the capsule network were used to teach the Support Vector Machine classifier, in order to diagnose thyroid cancer. Experimental results showed that the proposed method with 98% accuracy has achieved better results compared to convolutional networks.