نشریه سیستم های پردازشی و ارتباطی چند رسانه ای هوشمند
سال چهارم شماره 2 (پیاپی 12، تابستان 1402)

Face recognition is one of the most important biometric technologies for people identification, also used in access control. Face recognition is one of the important steps before identity recognition. Usually, one method is used to detect the presence of faces in images. Still, in this research, to increase the accuracy of detection, the combination of two methods is used to improve the detection performance: Viola-Jones algorithm and the matching of image components and texture with face and skin components. In the first step, we use the Viola-Jones algorithm to detect the facial features. In the next step, the features of the eye and nose tissues are analyzed with regression neural networks, and facial features are recognized better according to the facial features. In this research, the texture features of the right and left eyes and the nose of the face are used to increase the matching accuracy. We have used the faces of the FDD-Fold dataset to evaluate the proposed method. Comparing the performance of this method with the RCNN deep network method with a much smaller number of training data, we reached an accuracy of 96.36%, more than the deep learning network. This method gives good results in systems with limited computing ability and average amount of data.The face recognition system is one of the biometric identification systems and one of the most important technologies for people identification, which is also used in access control. Face identification is one of the few biometric methods that, with the advantages of high accuracy and low level of human intervention, is used in cases such as information security, law enforcement and monitoring, traffic control, and registration in attendance systems. This method creates more convenience and development with fewer requirements. then, this method has received more attention during the last twenty years.Face detection is a local binary classification problem that shows the presence of faces in the given image using boxes surrounding them. Although the Viola-Jones method is less accurate than modern methods such as convolutional neural networks; Its much lower efficiency and training parameters compared to the millions of parameters of a typical CNN result in faster training, better accuracy with limited data, and its use in devices with limited computing power such as cameras and mobile phones. The innovation of this method is matching the geometric pattern of the edges to identify the presence of the face in the image, along with matching the skin texture. This method seems to be faster and more accurate than the previous ones.

In this research, in the first step, we use Viola-Jones, one of the optimal face recognition algorithms in the image, to detect facial components. In the next step, we use the adaptation of the general shape of facial parts such as eyes, and match the textures in the image with the predicted texture for human skin, to improve the recognition performance and increase the recognition accuracy, in such a way that the regression neural networks examine the eye and nose tissue characteristics and according to the characteristics of the facial tissue, the facial components are recognized by the regression neural network. The investigated features in the texture include minimum and maximum color intensity, mean and median, and variance of the image. The data is given to the regression neural network for training. Here Remarkable thing is matching the overall shape of the human head and face, and in the next step matching the overall shape of the facial parts such as the eyes to improve the accuracy of the presented method. We also use the matching of textures in the image with the texture predicted for human skin to further improve the accuracy of the program's performance.

In this paper, we introduce the ParsAirCall toolkit, which is a tool for automatic recognition of Persian numbers in airport systems. It leverages deep transfer learning to improve performance in real and operational scenarios of voice-controlled smart telephone systems at airports across the country. In today's world, with the advancements in artificial intelligence, traditional systems for interacting with callers in telephone calls are not efficient, and this efficiency will be enhanced through automation and the automation of repetitive tasks.

ParsAirCall distinguishes itself by surpassing competing models in the Persian language, achieving heightened accuracy with fewer parameters and optimized computing resources. Addressing the challenge posed by limited data for Persian speech recognition, we meticulously curated a 30-hour telephony dataset, serving as the cornerstone for training the final ParsAirCall model. Embracing the innovative QuartzNet architecture, our deep transfer learning strategy empowers ParsAirCall to capture nuanced features in Persian speech, ensuring superior performance in number recognition tasks associated with airport telephone calls.

Experiments were conducted on both our collected telephony dataset and the Common Voice project, demonstrating ParsAirCall’s efficiency in achieving a 2.7% WER (Word Error Rate) in number recognition in airport telephone calls.

ParsAirCall emerges as a versatile tool, poised for seamless integration as a service into any Persian-language airport telephone system. Its practical application extends to number recognition in airport call centers, exemplifying the transformative impact of advanced technologies in streamlining communication processes within critical operational environments. ParsAirCall can be easily integrated as a service into any Persian-language airport telephone system, making it a practical tool for number recognition in airport call centers and telephone systems.

Due to the increasing volume and services available on the web, tools such as recommender systems in websites and applications that can help users find information and services of interest can be provided. For this reason, suitable guidance and suggestions for users in different choices, according to the user's priorities in different areas of a specific position, have been provided.

Recommender systems are information systems that assist in the decision-making process by modeling the behavior of users in operational environments in ranking, comparing, selecting and choose items by users, narrowing the information search through high-quality and accurate recommendations. In this research, a multi-agent recommender system is proposed as an intelligent shopping assistant in the process of buying suitable offers. The proposed model is used to analyze the sales data set of a UK-based store containing 1,067,371 records of online sales data.

By simulating the proposed model, the results of applying the model to the relevant data were analyzed. The proposed model in this research was simulated in MATLAB software version 2022 and the results of applying the proposed model on the data related to the sale of an online shopping were analyzed. According to the results, in this evaluation, the accuracy of the proposed model was 91.5% on average, compared to the neural network model, it was 86.41%, compared to the KNN model, 78.32%, compared to the SOM Ensembles model, 74.38%, compared to the Global Top-N model, 69.78%, compared to the Weighted item-based model, 72.31%, and compared to The Naïve Bayesian model has an accuracy of 59.68%, a higher accuracy in the right suggestion to users.

In this research, while studying recommender systems, the challenges in this field were examined and multi-agent systems were used to provide suggestions and recommendations with high accuracy and quality in ranking, comparison, selection and preferences of users' items in the decision-making process in operational environments. By combining multi-agent systems, multi-agent recommender systems were proposed that can provide suitable recommendations as a purchasing assistant in the purchasing process. The results of applying the proposed model on the data related to the purchase history of the customers of an online shopping showed that the proposed model has a good efficiency in evaluating the parameters used in comparison with the common methods in this property field.

Optimization of nonlinear time-varying functions, as a subset of nonlinear programming, has been widely observed in various economic and engineering models. In energy management, one example of optimizing nonlinear functions with time-variable components is the efficient allocation of energy resources and managing changes in demand and supply, leading to increased efficiency and reduced energy waste. In this article, we intend to use Zhang neural networks for optimizing nonlinear functions with time-varying components. By harnessing the parallel processing power of neural networks, Zhang networks search the solution space faster than traditional methods, significantly reducing the required computational time.

In this research, the proposed neural network receives data using MATLAB software. The data is first standardized using standard normalization methods. The data is then divided into four stages: training, testing, experimenting and validation which are further evaluated in five phases. The training data is based on the Luenberger-Madala algorithm for the first layer and a linear function for the second layer. Subsequently, the best network structure is considered with the transformation function and the proposed neural network model is tested in five stages. In this research, the Taylor series is used for data normalization and the zero-stability model of the n discrete time method is used to calculate error which reduces the error. The data in this research examined and evaluated in four stages of training, test, test and validation. Data training is based on Lunberg-Maud algorithm model for the first layer and linear function for the second layer. The reason for using Lunberg-Maud algorithm for research analysis is its convergence speed and higher efficiency due to not being in local minima and small error level.

The best network structure with transformation function was considered and tested in 5 steps based on the proposed neural network model. The mean square of error in the third and fourth experiments has gradually increased compared to the first two stages. This amount of difference in the performance error as well as the coefficient of determination is different in each iteration and is caused by getting stuck in local minima.

Due to the results obtained in the five test stages, it can be said that the algorithm based on the proposed neural network improves the performance of the network by increasing the learning rate. However, this algorithm is highly sensitive to local minima. This problem exists even when the learning rate is small and therefore the step of the algorithm is small. To avoid this sensitivity to local minima, the algorithm used in the proposed network was tested with momentum with different learning rates in five stages and the best result was evaluated. Also, at each stage, the process of training, testing and validation was evaluated separately.