Increase Recognition Accuracy Research Articles

A simple teacher behavior recognition method for massive teaching videos based on teacher set

The analysis of teacher behavior of massive teaching videos has become a surge of research interest recently. Traditional methods rely on accurate manual analysis, which is extremely complex and time-consuming for analyzing massive teaching videos. However, existing works on action recognition are difficultly transplanted to the teacher behavior recognition, because it is difficult to extract teacher’s behavior from complex teaching scenario, and teacher’s behaviors are given professional educational semantics. These methods are not adequate for the need of the teacher behavior recognition. Thus, a novel and simple recognition method of teacher behavior in the actual teaching scene for massive teaching videos is proposed, which can provide technical assistance for analyzing teacher behavior and fill the blank of automatic recognition of teacher behavior in actual teaching scene. Firstly, we discover the educational pattern which it be named “teacher set”, that is, the spatial region of the video of the whole class where teachers should exist. Based on this, the algorithm of teacher set identification and extraction (Teacher-set IE algorithm) is studied to identify the teacher in the teaching video, and reduce the interference factors of classroom background. Then, an improved behavior recognition network based on 3D bilinear pooling (3D BP-TBR) is presented to enhance fusion representation of three-dimensional features thus identifying the categories of teacher behavior, and experiments show that 3D BP-TBR can achieve better performance on public and self-built dataset (TAD-08). Hence, our whole approach can increase recognition accuracy of teacher behavior in the actual teaching scene to utilize the deep integration of educational characteristics and action recognition technology.

A multibiometric system based on the fusion of fingerprint, finger-vein, and finger-knuckle-print

Authentication systems play an essential role in our lives today. Human biological, behavioral, and morphological characteristics are usually used in an authentication system in a vast scope of applications ranging from unlocking consumer devices to surveillance and forensic analysis and offered an alternative to credit cards, ID cards, passports, driving licenses, etc. Unibiometric systems, which use a single biometric modality, suffer from some drawbacks such as low protection of user’s privacy against attacks. Multibiometric systems, which fuse features of biometric characteristics, can cope with unibiometric systems’ drawbacks and improve security and recognition accuracy. However, some main questions such as ‘what is the optimal number of biometric modalities?’, ‘how much accuracy do we need for our system?’, and ‘how much money do we want to invest for our system?’ have to be considered and answered before designing and implementing a multibiometric system. Unfortunately, the existing multibiometric systems have not considered and responded to all questions. Furthermore, identification mode for multibiometric systems is a challenging task and almost all of them have focused on verification mode. In this paper, we only consider a finger and employ the maximum possible modalities of the finger, i.e., fingerprint, finger-vein, and finger-knuckle-print, to increase user-friendliness and reducing the cost of implementation of the system. The proposed system only uses three cameras to capture all contactless fingerprint, finger-vein, and finger-knuckle images. Also, we propose an algorithm that makes enable the system to work in identification mode. The experiments on established databases exhibit that our proposed algorithm significantly increases recognition accuracy.

Understanding the limits of 2D skeletons for action recognition

With the development of motion capture technologies, 3D action recognition has become a popular task that finds great applicability in many areas, such as augmented reality, human–computer interaction, sports, or healthcare. On the other hand, the acquisition of 3D human skeleton data is an expensive and time-consuming process, mainly due to the high costs of capturing technologies and the absence of suitable actors. We overcome these issues by focusing on the 2D skeleton modality that can be easily extracted from ordinary videos. The objective of this work is to demonstrate a high descriptive power of such a 2D skeleton modality by achieving accuracy on the task of daily action recognition competitive to 3D skeleton data. More importantly, we thoroughly analyze the factors that significantly influence the 2D recognition accuracy, such as the sensitivity towards data normalization, scaling, quantization, and 3D-to-2D distortions in skeleton orientations and sizes, which are caused by the loss of depth dimension and fixed-angle camera view. We also provide valuable insights on how to mitigate these problems to increase recognition accuracy significantly. The experimental evaluation is conducted on three datasets different in nature. The ability to learn different types of actions better using either 2D or 3D skeletons is also reported. Throughout experiments, a generic light-weight LSTM network is used, whose architecture can be easily tuned to achieve the desired trade-off between its accuracy and efficiency. We show that the proposed approach achieves not only the state-of-the-art results in 2D skeleton action recognition but is also highly competitive to the best-performing methods classifying 3D skeleton sequences or the visual content extracted from ordinary videos.

Interactions of Emotion and Self-reference in Source Memory: An ERP Study.

The way emotional information is encoded (e.g., deciding whether it is self-related or not) has been found to affect source memory. However, few studies have addressed how the emotional quality and self-referential properties of a stimulus interactively modulate brain responses during stimulus encoding and source memory recognition. In the current study, 22 participants completed five study-test cycles with negative, neutral, and positive words encoded in self-referential versus non-self-referential conditions, while event-related potentials of the electroencephalogram were recorded. An advantage of self-referential processing in source memory performance, reflected in increased recognition accuracy, was shown for neutral and positive words. At the electrophysiological level, self-referential words elicited increased amplitudes in later processing stages during encoding (700-1,200 ms) and were associated with the emergence of old/new effects in the 300-500 ms latency window linked to familiarity effects. In the 500-800 ms latency window, old/new effects emerged for all valence conditions except for negative words studied in the non-self-referential condition. Negative self-referential words also elicited a greater mobilization of post-retrieval monitoring processes, reflected in an enhanced mean amplitude in the 800-1,200 ms latency window. Together, the current findings suggest that valence and self-reference interactively modulate source memory. Specifically, negative self-related information is more likely to interfere with the recollection of source memory features.

Fault detection in mixture production process based on wavelet packet and support vector machine

Mixture production equipment is widely employed in road construction, and the quality of the produced mixture is the essential factor to ensure the quality of road construction. To detect the quality of the real-time produced mixture and solve the shortcomings of laboratory detection lag, a new fault detection method in the mixture production process is proposed, which is based on wavelet packet decomposition (WPD) and support vector machine (SVM). The proposed scheme includes feature extraction, feature selection, SVM classification, and optimization algorithm. During feature extraction, wavelet basis function is utilized to 4-layer decompose the aggregate and asphalt data mixed in real-time. The energy value calculated by wavelet packet coefficient is the extracted feature. During feature selection, a method combining the chi-square test and wrapper (CSW) is conducted to select the optimal feature subset from WPD features. Eventually, by adopting the optimal feature subset, SVM has been developed to classify various faults. Its parameters are optimized by differential evolution (DE) algorithm. In the test stage, multiple faults of different specifications of aggregates and asphalt are detected in the mixture production process. The results demonstrate that (1) accuracy produced by the CSW method with WPD features is 4.33% higher than the PCA method with statistical features; (2) SVM classification method optimized by DE algorithm brings an increase in recognition accuracy of identifying different types of mixture production faults produced by different equipment. Compared to other available methods, the proposed algorithm has a very outstanding detection performance.

Automated door with face recognition: using artificial neural network approach

The implementation of an accurate face recognition system in a hardware device is a very important aspect of various security applications. In a security system, any type of password is used to access private and confidential data, some are hard to remember and can be stolen or guessed, misplaced, forgotten, purloined or duplicated, or become corrupted and unreadable. While there have been significant improvements in the algorithms with increasing recognition accuracy, only a few types of research were conducted on implementing these face recognitions in hardware devices. Door access control systems based on face recognition is geared towards simplifying much-difficult face recognition problems in uncontrolled environments. This report comprised mainly of three sub-phases: face detection, face recognition, and automatic door access control. The face is detected by using the Viola-Jones method and face recognition is implemented by using the Principal Component Analysis (PCA). If a face is recognized, it is known, else it is unknown. The door will open automatically for the known person due to the command of the microcontroller (Atmega328P) received from the MATLAB program. On the other hand, the alarm will ring for an unknown person.

Combination of Pattern Classifiers Based on Naive Bayes and Fuzzy Integral Method for Biological Signal Applications

Background: Achieving the best possible classification accuracy is the main purpose of each pattern recognition scheme. An interesting area of classifier design is to design for biomedical signal and image processing. Materials and Methods: In the current work, in order to increase recognition accuracy, a theoretical frame for combination of classifiers is developed. This method uses different pattern representations to show that a wide range of existing algorithms could be incorporated as the particular cases of compound classification where all the pattern representations are used jointly to make an accurate decision. Results: The results show that the combination rules developed under the Naive Bayes and Fuzzy integral method outperforms other classifier combination schemes. Conclusion: The performance of different combination schemes has been studied through an experimental comparison of different classifier combination plans. The dataset used in the article has been obtained from biological signals.

Generating photo-realistic training data to improve face recognition accuracy

Face recognition has become a widely adopted biometric in forensics, security and law enforcement thanks to the high accuracy achieved by systems based on convolutional neural networks (CNNs). However, to achieve good performance, CNNs need to be trained with very large datasets which are not always available. In this paper we investigate the feasibility of using synthetic data to augment face datasets. In particular, we propose a novel generative adversarial network (GAN) that can disentangle identity-related attributes from non-identity-related attributes. This is done by training an embedding network that maps discrete identity labels to an identity latent space that follows a simple prior distribution, and training a GAN conditioned on samples from that distribution. A main novelty of our approach is the ability to generate both synthetic images of subjects in the training set and synthetic images of new subjects not in the training set, both of which we use to augment face datasets. By using recent advances in GAN training, we show that the synthetic images generated by our model are photo-realistic, and that training with datasets augmented with those images can lead to increased recognition accuracy. Experimental results show that our method is more effective when augmenting small datasets. In particular, an absolute accuracy improvement of 8.42% was achieved when augmenting a dataset of less than 60k facial images.

A unified generative model using generative adversarial network for activity recognition

The recent advancement of deep learning methods has seen a significant increase in recognition accuracy in many important applications such as human activity recognition. However, deep learning methods require a vast amount of sensor data to automatically extract the most salient features for activity classification. Therefore, in this paper, a unified generative model is proposed to generate verisimilar data of different activities for activity recognition. The proposed generative model not only able to generate data that have a similar pattern, but also data with diverse characteristics. This allows for data augmentation in activity classification to improve the overall recognition accuracy. Three similarity measures are proposed to assess the quality of the synthetic data in addition to two visual evaluation methods. The proposed generative model was evaluated on a public dataset. The training data was prepared by systematically varying the combination of original and synthetic data. Results have shown that classification using the hybrid training data achieved a comparable recognition accuracy with the classification using the original training data. The performance of the classifiers maintained at the recognition accuracy of 85%.

Pattern Recognition of DC Partial Discharge on XLPE Cable Based on ADAM-DBN

Pattern recognition of DC partial discharge (PD) receives plenty of attention and recent researches mainly focus on the static characteristics of PD signals. In order to improve the recognition accuracy of DC cable and extract information from PD waveforms, a modified deep belief network (DBN) supervised fine-tuned by the adaptive moment estimation (ADAM) algorithm is proposed to recognize the four typical insulation defects of DC cable according to the PD pulse waveforms. Moreover, the effect of the training sample set size on recognition accuracy is analyzed. Compared with naive Bayes (NB), K-nearest neighbor (KNN), support vector machine (SVM), and back propagation neural networks (BPNN), the ADAM-DBN method has higher accuracy on four different defect types due to the excellent ability in terms of the feature extraction of PD pulse waveforms. Moreover, the increase of training sample set size would lead to the increase of recognition accuracy within a certain range.

Accuracy increase of software and hardware appliance for muscle activity measuring and monitoring by filtration of carrier component and frequencies higher than measured signal range

Subject of Research. The paper proposes a method of muscle activity filtering measurements for a mobile hardware- software appliance used in surface electromyography. The method extends the dynamic range of measurements by capacity growth of analog-to-digital converter aimed at the increase of recognition accuracy and range of muscle activity. Method. A filter model for signals from sensors for a muscle activity controller was developed in the Proteus Design Suite software package. The filter of the signal carrier component based on the RC high-pass filter provides isolation of the measuring unit from the reference voltage of the sensor. An active low-pass filter amplifies the signal from the sensor and filters out the noise higher than the frequency range of muscle activity signals. Main Results. Filtering of the signal carrier component and increasing the order of low-pass filter show positive results in simulation. The paper presents amplitude-frequency characteristics plots and model structures with and without RC filter, with an active low-pass filter of the first order and an active low-pass filter of the second order. An amplifier unit electrical circuit for a muscle activity controller is developed based on the methodology for muscle activity measurement filtering. The results obtained are applicable for improvement of the prototype for the mobile hardware-software appliance used in surface electromyography. Practical Relevance. The developed complex can be applied in a system for muscle activity measuring and monitoring as the rehabilitation process maintenance during the movement of patients with injuries and disorders of the musculoskeletal system. This complex can be used in various neurophysiological studies where the monitoring of muscle activity dynamics in the process of the examined subject movement is required.

Deep Learning Based Human Emotional State Recognition in a Video

Human emotions play significant role in everyday life. There are a lot of applications of automatic emotion recognition in medicine, e-learning, monitoring, marketing etc. In this paper the method and neural network architecture for real-time human emotion recognition by audio-visual data are proposed. To classify one of seven emotions, deep neural networks, namely, convolutional and recurrent neural networks are used. Visual information is represented by a sequence of 16 frames of 96 × 96 pixels, and audio information - by 140 features for each of a sequence of 37 temporal windows. To reduce the number of audio features autoencoder was used. Audio information in conjunction with visual one is shown to increase recognition accuracy up to 12%. The developed system being not demanding to be computing resources is dynamic in terms of selection of parameters, reducing or increasing the number of emotion classes, as well as the ability to easily add, accumulate and use information from other external devices for further improvement of classification accuracy.

APPLICATION OF KOHONEN SELF-ORGANIZING MAP TO SEARCH FOR REGION OF INTEREST IN THE DETECTION OF OBJECTS

Today, there is a serious need to improve the performance of algorithms for detecting objects in images. This process can be accelerated with the help of preliminary processing, having found areas of interest on the images where the probability of object detection is high. To this end, it is proposed to use the algorithm for distinguishing the boundaries of objects using the Sobel operator and Kohonen self-organizing maps, described in this paper and shown by the example of determining zones of interest when searching and recognizing objects in satellite images. The presented algorithm allows 15–100 times reduction in the amount of data arriving at the convolutional neural network, which provides the final recognition. Also, the algorithm can significantly reduce the number of training images, since the size of the parts of the input image supplied to the convolution network is tied to the image scale and equal to the size of the largest recognizable object, and the object is centered in the frame. This allows to accelerate network learning by more than 5 times and increase recognition accuracy by at least 10 %, as well as halve the required minimum number of layers and neurons of the convolutional network, thereby increasing its speed.

PROCESSING IMAGES OF SALES RECEIPTS FOR ISOLATING AND RECOGNISING TEXT INFORMATION

Objectives. This article presents an application for the processing of scanned images of sales receipts for subsequent extraction of text information using the Tesseract OCR Engine. Such an application is useful for maintaining a family budget or for accounting in small companies. The main problem of receipt recognition is the low quality of ink and printing paper, which results in creasing and tears, as well as the rapid fading of printed characters.Methods. The study is based on a number of algorithms based on mathematical morphology methods for opening, closing and morphological gradient operations, as well as image conversion, which can significantly improve the final recognition of characters by Tesseract.Results. In order to solve this problem, a special image normalisation algorithm is proposed, which includes locating a receipt on an image, processing the received image section, removing image capture and carrier defects, as well as point processing for restoring missing characters. The developed application supports increased recognition accuracy of text information when using Tesseract OCR.Conclusion. The developed system recognises characters with fairly high accuracy, demonstrates a result that is better than that obtained when using the unmodified Tesseract method, but which is still inferior to the recognition accuracy of ABBY FineReader. Methods are also been proposed aimed at improving the developed algorithm.

EchoFace: Acoustic Sensor-Based Media Attack Detection for Face Authentication

Face authentication systems have gained widespread popularity because of their user-friendly usage and increasing recognition accuracy. Unfortunately, the boom in mobile social networks has bought with it media-based facial forgery; a critical threat where an adversary forges or replays the victim’s photograph/video to fool the system. In this article, we propose EchoFace, an effective and robust liveness detection system to enhance face authentication in defending against media-based attacks, which works with today’s smartphones/smartwatches without any hardware modification. EchoFace uses active acoustic sensing to differentiate the uneven stereostructure of the face and the flat forged media. Our proposed scheme effectively extracts the desired reflection profiles from the target. Moreover, we propose effective similarity measurements of reflection profiles to distinguish live users from forged media, which works robustly under various environmental conditions. EchoFace only requires low cost and universally equipped acoustic sensors without human intervention for liveness detection, which can be easily deployed in a variety of application scenarios. We implement EchoFace on commercial smartphones, and experiment results show that EchoFace achieves an average detection accuracy higher than 96% and false alarm rate lower than 4% across various media attacks and different levels of background noise. This shows its great potential to enhance the security of widely deployed face authentication systems in real scenarios.

The use of convolutional neural networks to identify artifacts of cells micrographs in biomedical research

The work is devoted to the use of artificial neural networks to solve the problem of recognition and isolation of objects (cells) in digital micrographs used in the practice of microbiological research. This task is relevant due to a combination of two factors: a large amount of data (hundreds and thousands of photographs, hundreds of objects on each) arising from such studies, as well as the high complexity of manual image processing and the risk of operator errors. To solve the problem of recognition and separation of objects (cells) in digital biomedical micrographs, artificial convolutional neural networks are used. However, in most cases, the use of artificial neural networks requires selection of the network architecture and its training for each task individually. Each time when solving the problem of training the network, it is necessary to collect huge training samples, the more complex the more complex the architecture. The task of cell isolation was solved using a convolutional neural network. The neural network architecture of Mask R-CNN Facebook Research was used. We used a pretrained neural network, retrained in digital micrographs obtained and marked out by the authors. The solution is implemented in Python using TensorFlow, an open source machine learning software library developed by Google. As a result of testing the system, on the available data, the correctness of cell recognition in microphotographs was more than 95%, despite the small size of the digital images used. The applied approach definitely showed its efficiency on the available experimental data and has development prospects in the direction of increasing the size of processed images, increasing recognition accuracy, expanding the composition of secreted objects, working not only with cells, but also with structures in tissues.

Classification of Sambas Traditional Fabric “Kain Lunggi” Using Texture Feature

Traditional fabric is a cultural heritage that has to be preserved. Kain Lunggi is Sambas traditional fabric that saw a decline in its crafter. To introduce Kain Lunggi in a broader national and global society in order to preserve it, a digital image processing based system to perform Kain Lunggi pattern recognition need to be built. Feature extraction is an important part of digital image processing. The visual feature that does not represent the character of an object will affect the accuracy of a recognition system. The purposes of this research are to perform feature selection on sets of feature to determine the best feature that can increase recognition accuracy. This research conducted in several steps which are image acquisition of Kain Lunggi pattern, preprocessing to reduce image noise, feature extraction to obtain image features, and feature selection. GLCM is implemented as a feature extraction method. Feature extraction result will be used in a feature selection process using CFS (Correlation-based Feature Selection) methods. Selected features from CFS process are Angular Second Moment, Contrast, and Correlation. Selected features evaluation is conducted by calculating classification accuracy with the KNN method. Classification accuracy prior to feature extraction is 85.18% with K values K=1 ; meanwhile, the accuracy increases to 88.89% after feature selection. The highest accuracy improvement of 20.74% in KNN occurred when using K value K= 4.

Decoding emotions from nonverbal vocalizations: How much voice signal is enough?

How much acoustic signal is enough for an accurate recognition of nonverbal emotional vocalizations? Using a gating paradigm (7 gates from 100 to 700 ms), the current study probed the effect of stimulus duration on recognition accuracy of emotional vocalizations expressing anger, disgust, fear, amusement, sadness and neutral states. Participants (n = 52) judged the emotional meaning of vocalizations presented at each gate. Increased recognition accuracy was observed from gates 2 to 3 for all types of vocalizations. Neutral vocalizations were identified with the shortest amount of acoustic information relative to all other types of vocalizations. A shorter acoustic signal was required to decode amusement compared to fear, anger and sadness, whereas anger and fear required equivalent amounts of acoustic information to be accurately recognized. These findings confirm that the time course of successful recognition of discrete vocal emotions varies by emotion type. Compared to prior studies, they additionally indicate that the type of auditory signal (speech prosody vs. nonverbal vocalizations) determines how quickly listeners recognize emotions from a speaker’s voice.

Fault diagnosis of planetary gearbox using a novel semi-supervised method of multiple association layers networks

Abstract The vibration signal can effectively represent the equipment fault information and be utilized for fault diagnosis by further extracting sensitive features. However, traditional supervised diagnosis requires a huge amount of labeled samples, and the features extraction and selection are mainly done manually, which are costly and time consuming. Deep Learning (DL) can automatically extract fault-sensitive features and effectively improve the recognition rate of fault diagnosis, while Semi-Supervised Learning (SSL) trains a model together with labeled and unlabeled data to increase recognition accuracy. In this paper, a novel deep Semi-Supervised method of Multiple Association Layers Networks (SS-MALN) framework of planetary gearbox vibration-based fault diagnosis is proposed. The SS-MALN model has the advantages of SSL and DL simultaneously, which can reduce the amount of labeled samples and improve the accuracy of recognition. The wavelet packet transform was employed to highlight various impulse components and present time-frequency features of the original vibration signal. Then, the transformed samples were divided into two parts, a major part of which deletes the category information, i.e. the labels. After that, the labeled and unlabeled samples were employed to train the improved SS-MALN model. Validation by Drivetrain Diagnostic Simulator testbed data on several planetary gearbox fault datasets manifests that our SS-MALN semi-supervised method can deliver enhanced performance over other traditional deep networks of planetary gearbox fault classification with less labeled data.

An Enhanced Temporal Feature Integration Method for Environmental Sound Recognition

Temporal feature integration refers to a set of strategies attempting to capture the information conveyed in the temporal evolution of the signal. It has been extensively applied in the context of semantic audio showing performance improvements against the standard frame-based audio classification methods. This paper investigates the potential of an enhanced temporal feature integration method to classify environmental sounds. The proposed method utilizes newly introduced integration functions that capture the texture window shape in combination with standard functions like mean and standard deviation in a classification scheme of 10 environmental sound classes. The results obtained from three classification algorithms exhibit an increase in recognition accuracy against a standard temporal integration with simple statistics, which reveals the discriminative ability of the new metrics.