per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 3 12 article Seyedehfatemeh Karimi ftmkm9776@gmail.com 1 Maryam Khodabakhsh m_khodabakhsh@shahroodut.ac.ir 2 M.Sc in Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran Assistant Professor of Faculty of Computer Engineering, Shahrood University of Technology, Shahrood, Iran With the rapid increase in the use of search engines, the need for developing more effective information retrieval and ranking methods has become critical. One of the key challenges in information retrieval is predicting query performance, which involves estimating how well a search engine can fulfill a user's information need. Accurate prediction of query performance allows search engines to take adaptive actions, such as query reformulation or ranking adjustment, to enhance retrieval effectiveness. Query Performance Prediction (QPP) methods fall into two main categories: pre-retrieval prediction and post-retrieval prediction. Pre-retrieval predictors estimate query difficulty before the retrieval process, relying on linguistic and statistical query features rather than retrieved documents. In contrast, post-retrieval prediction methods assess query performance based on the ranking list and document collection, providing deeper insights into retrieval effectiveness. In this study, we propose a novel unsupervised post-retrieval QPP method that evaluates query performance by analyzing the clustering behavior of retrieved documents. Our method defines five new metrics—CC, DCIC, DCNIC, DCNICR, and CCR— to measure the distribution and coherence of retrieved documents. These metrics help assess query difficulty by capturing how documents group into clusters, identifying outlier documents that do not fit well into clusters, and evaluating the overall structure of retrieved results. By leveraging these metrics, our approach provides a more fine-grained estimation of query performance without requiring human-labeled data. To evaluate the effectiveness of the proposed method, we conduct experiments on three datasets: TREC DL 2019, TREC DL 2020, and DL-Hard. The results demonstrate that our approach improves Spearman's correlation coefficient by 0.009 and 0.163 on the TREC DL 2019 and DL-Hard datasets, respectively. Additionally, it increases Pearson’s correlation coefficient by 0.037 on the TREC DL 2020 dataset compared to state-of-the-art unsupervised QPP methods. These improvements indicate that clustering-based QPP methods can effectively capture query difficulty and retrieval quality without the need for external supervision. http://jsdp.rcisp.ac.ir/article-1-1407-en.pdf Query Performance Prediction Information Retrieval Search Engines per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 13 24 article Ali Nazari ali4nazari4@gmail.com 1 Mostafa Fallah mostafa_fallah@comp.iust.ac.ir 2 Mohammad Javad Taheri taheri_mo96@comp.iust.ac.ir 3 Abolfazl Diyanat adiyanat@iust.ac.ir 4 M.Sc. in Computer Engineering, Iran University of Science and Technology, Tehran, Iran M.Sc. Student in Computer Engineering, Iran University of Science and Technology, Tehran, Iran M.Sc. in Computer Engineering, Iran University of Science and Technology, Tehran, Iran Assistant Professor, Faculty of Computer Engineering, Iran University of Science and Technology, Tehran, Iran The design and maintenance of the radio access network (RAN) in mobile telecommunications requires meticulous monitoring of network performance to ensure optimal quality and coverage. Drive testing is a prevalent methodology for collecting data on network status across delineated geographical areas. This process involves systematically traversing various routes, including streets and pathways, to assess network performance metrics such as Reference Signal Received Power (RSRP). Although drive testing provides essential insights for identifying regions with inadequate signal quality, it is inherently resource-intensive, involving considerable time and financial expenditures. This paper proposes an innovative optimization methodology aimed at enhancing the efficiency of data collection during drive tests. The proposed approach is organized into four fundamental steps: (1) partitioning the map into smaller sections, (2) selecting critical points within each section, (3) employing map-matching techniques to accurately align these points with actual streets and pathways, and (4) determining the optimal route for traversing the critical points. A rectangular area of interest is selected and divided into K smaller sub-regions, within which M critical points are identified according to a predefined criterion. These points, which may not initially correspond with the existing street network, are corrected through map-matching techniques to ensure feasible traversal paths. Lastly, an optimization algorithm is utilized to compute the shortest route that encompasses all identified critical points. The efficacy of the proposed method is assessed through experimental studies that manipulate key parameters K (denoting the number of sub-regions) and M (indicating the maximum critical points per zone). The evaluation emphasizes two critical metrics: the total distance traveled and the success rate in detecting areas with RSRP values below -100 dBm. Results indicate that the proposed approach significantly decreases the distance required for drive testing while achieving substantial coverage of areas with weak signals. For example, in an experiment where the optimized route covered only 18.54 kilometers—equivalent to 34% of the distance of an entire drive test—it successfully identified 70% of regions with poor RSRP. Furthermore, this paper introduces a criterion for assessing the effectiveness of drive test routes, highlighting the balance between route length and coverage of low-signal areas. The findings substantiate that adequate data for network performance analysis can be secured with considerable savings in both cost and time compared to conventional exhaustive drive testing. While this study concentrates on RSRP measurements within 4G networks, the proposed methodology is adaptable for other metrics, including Reference Signal Received Quality (RSRQ). Future research may also examine the integration of alternative data sources, such as satellite imagery, to further refine map partitioning and critical point selection, thereby enhancing the overall efficacy of the proposed method. http://jsdp.rcisp.ac.ir/article-1-1428-en.pdf Mobile Networks Drive Test Route Optimization Reference Signal Received Power (RSRP) per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 25 38 article Soheil Rastari rastari.s@qut.ac.ir 1 Morteza Mohajjel kafshdooz mohajjel@qut.ac.ir 2 Mahboubeh Shamsi shamsi@qut.ac.ir 3 Master, Faculty of Electrical and Computer Engineering, Qom University of Technology, Qom, Iran Assistant Professor, Faculty of Electrical and Computer Engineering, Qom University of Technology, Qom, Iran Associate Professor, Faculty of Electrical and Computer Engineering, Qom University of Technology, Qom, Iran Creating neural networks in a non-automatic way is a slow process based on trial and error. When the number of network parameters or the number of layers increases, the non-automatic method becomes very expensive and the final result may be suboptimal. Automatic network architecture search algorithms are used to solve this problem. Recently, these algorithms have been able to achieve high accuracies on various datasets such as CIFAR-10, ImageNet, and Penn Tree Bank. These algorithms have the ability to search a wide space of architectures with different characteristics such as network depth, width, connection method, and operations in order to discover architectures with appropriate accuracy. However, one of the traditional challenges of these algorithms is their high search time (Approximately tens of thousands of GPU hours), which has been reduced to tens of hours with new research. Another challenge that usually exists in these methods is their focus on improving network accuracy, while other criteria such as network speed and consumed resources are not taken into account. As a result, these methods cannot be used directly to find the optimal architecture in embedded systems that have limited resources such as processing power, memory, and energy consumption. Therefore, search methods should be devised that are aware of these limitations. Research has been done in this field in recent years, but these methods do not focus specifically on coarse-grained multi-core architectures that do not have a GPU. In this article, we present a method for the automatic design of networks that are suitable for running on multi-core processors. In this method, based on gradient descent, a SuperNet with parallel paths and computational blocks is created. The number of parallel paths is equal to or less than the number of cores. We use a series of decision variables to select appropriate operations in each block of the path. In addition to deciding on the operations performed in each block, deciding is also made regarding synchronization points to utilize the intermediate results of parallel paths and improve the network's accuracy. Then, by training the decision variables (block type and synchronization points) simultaneously with the main network weights, an appropriate subnetwork is selected. Due to the use of the gradient descent method in this approach, the training process is performed only twice, resulting in the final structure of the network. As a result, it has a much lower execution time compared to other methods based on evolutionary search and reinforcement learning. Additionally, considering the constraints of the target system, such as the number of cores and memory consumption, can lead to a more suitable architecture compared to other methods. Experiments conducted on the CIFAR-10 dataset demonstrate that the proposed method can achieve satisfactory accuracy with very little search time. http://jsdp.rcisp.ac.ir/article-1-1418-en.pdf Neural network architecture search embedded systems parallelization multi-core processors gradient descent method. per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 39 52 article pardis moradbeiki p.moradbeiki@ec.iut.ac.ir 1 alireza basiri basiri@iut.ac.ir 2 Phd Student, Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan, Iran Assistant Professor of Electrical and Computer Engineering Department, Isfahan University of Technology, Isfahan, Iran The aim of the request recognition task in social networks is to understand the intent behind the posts, comments, or messages shared by users. Many businesses are actively present on various social networks, making it crucial to identify user needs for marketers in this space to foster the growth of online businesses and e-commerce. Detecting request messages automatically and filtering them is essential. However, social network messages often contain slang and numerous spelling errors, posing challenges for research in this domain. While extensive research has been conducted in English, studies on this task in Persian are limited. Telegram stands out as the most popular social network in Iran, with a large Persian-speaking user base. This study utilized a standard labeled Persian dataset from Telegram for training and testing purposes, comprising 85741 messages from the platform, evenly split between request and non-request categories. To tackle the significant challenges posed by sarcastic messages and spelling mistakes on social media platforms, we devised a multi-step hybrid strategy. The initial step involves preprocessing. Social media data typically consists of unstructured and slang-ridden user messages, necessitating preprocessing to enhance Persian text processing and reduce slang usage. The pre-processing phase is crucial when dealing with social media platforms. Because Telegram is unique compared to other platforms the data cleaning process varies. This study's accomplishment includes developing a unique dataset and filtering out noise from Telegram enhancing improvement in the pre-processing phase. Also, this involves normalizing different word forms, such as "beautiful" and "beauty," to maintain the integrity of word meanings. The subsequent step focuses on feature extraction. Various approaches to feature extraction come with their own set of advantages and drawbacks. Hence, we employed hybrid feature extraction methods to address this complexity. While Tf-Idf methods assess word importance without considering meaning, FastText retains semantic similarity. By combining the bag of words and FastText methods, our research aims to enhance accuracy. The final step involves classification, where deep learning networks are utilized to evaluate these features. Experimental findings indicate that our final model achieves precision, recall, and f-score rates of nearly 90%, representing a 5% improvement on average compared to previous methodologies. http://jsdp.rcisp.ac.ir/article-1-1425-en.pdf e-commerce request detection social networks messaging deep-learning based method. per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 53 70 article samira asadzadeh samira.assadzadeh@gmail.com 1 Elham parvinnia Elham.parvinnia@iau.ac.ir 2 PhD Student, Department of Computer Engineering, Shiraz Branch, Islamic Azad University, Shiraz, Iran Associate Professor, Department of Computer Engineering, Shiraz Branch, Islamic Azad University, Shiraz, Iran With the rapid growth and development of urban areas, the demand for secure and efficient transportation systems in urban logistics has become increasingly critical. One of the most pressing challenges in this domain is the routing of bank cash-in-transit (CIT) vehicles, which, due to their sensitive and high-risk nature, require precise and intelligent planning. The primary challenges for CIT vehicles include ensuring security, optimizing timing, managing traffic congestion, and selecting the most efficient routes. This paper proposes an innovative method for the intelligent routing of CIT vehicles by mapping the urban routing problem to the Traveling Salesman Problem (TSP). The proposed approach leverages the Ant Colony Optimization (ACO) algorithm, enhanced with real-world constraints such as heavy traffic, low-security areas, and road hazards, which are incorporated as additional weights in the optimization process. The data used in this study includes the urban traffic map of Shiraz, Iran, and the locations of various banks. The results demonstrate that the proposed method effectively selects routes that avoid high-traffic zones, crime-prone areas, and hazardous roads while optimizing travel time. This approach not only enhances security and operational efficiency but also contributes to reducing operational costs. The core innovation of this research lies in its ability to map the urban routing problem to the TSP, a well-known combinatorial optimization problem, and to utilize the ACO algorithm, which is inspired by the foraging behavior of ants. In nature, ants leave pheromone trails to communicate and find the shortest path between their nest and food sources. Similarly, the ACO algorithm employs artificial ants to explore possible routes, leaving virtual pheromones to guide subsequent ants toward optimal paths. In this study, the ACO algorithm is further enhanced by incorporating heuristic information such as traffic volume, security rates, and unsafe driving conditions, which are treated as critical factors in the routing process. The implementation of the proposed method utilizes real-world data from the urban traffic map of Shiraz, including the locations of eight major banks and the routes connecting them. The distances between these locations are calculated using the Haversine formula, which accounts for the Earth's curvature to provide accurate geographical distances. The algorithm is tested with various parameters, including different numbers of artificial ants (ranging from 10 to 200), evaporation rates (0.1 to 0.5), and exploration-exploitation trade-offs (alpha and beta values). The results show that the proposed method can effectively identify routes that minimize travel time while avoiding high-traffic areas, crime-prone zones, and hazardous roads. One of the key contributions of this research is the integration of multiple heuristic factors into the ACO algorithm. Traditional routing algorithms often focus solely on minimizing distance or travel time, neglecting critical real-world constraints. In contrast, the proposed method assigns weights to factors such as traffic volume, security levels, and unsafe driving conditions, allowing the algorithm to prioritize safer and more efficient routes. For example, routes passing through areas with high crime rates or heavy traffic are penalized, reducing their likelihood of being selected. This approach ensures that the final route is not only the shortest but also the safest and most reliable. Comparative evaluations indicate that the proposed algorithm offers a more realistic and comprehensive solution compared to other models. By balancing multifaceted aspects of routing such as safety, timeliness, and cost, the method proves to be highly effective. The algorithm's ability to avoid routes with heavy traffic, low security, and poor road conditions significantly enhances the safety and time efficiency of CIT vehicles. Furthermore, the proposed method represents a significant step forward in improving the efficiency and security of banking operations by providing an innovative approach to intelligent routing for CIT vehicles. By mapping the urban routing problem to the TSP and utilizing the ACO algorithm with real-world constraints, the method delivers optimal and secure routes. In comparison to other studies, the proposed method demonstrates superior performance in terms of accuracy, execution time, energy consumption, and security. The use of real-time traffic data and the algorithm's ability to adapt to dynamic changes further enhance the practicality and reliability of the proposed solution. This adaptability makes the method particularly suitable for urban environments with fluctuating traffic patterns and evolving security challenges. In conclusion, this paper presents a novel and effective approach to intelligent routing for CIT vehicles in urban environments. By combining the strengths of the ACO algorithm with real-world constraints, the proposed method offers a comprehensive solution that balances efficiency, security, and reliability. Future work could explore the integration of machine learning techniques to further enhance the algorithm's predictive capabilities, enabling it to anticipate and respond to emerging traffic and security challenges proactively. This research marks a significant advancement in the field of urban logistics, providing a robust framework for the safe and efficient routing of high-risk transportation systems. http://jsdp.rcisp.ac.ir/article-1-1424-en.pdf Intelligent routing Ant Colony Optimization Traveling Salesman Problem route optimization traffic security per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 71 82 article Zahra Firuz Mahjanabadi zahra.firooz2020@gmail.com 1 Pouria Jafari pjafari@ece.usb.ac.ir 2 Mehdi Rezaei mehdi.rezaei@ece.usb.ac.ir 3 M-Eng Department of Telecommunications, Faculty of Electrical and Computer Engineering, University of Sistan and Baluchestan, Zahedan, Iran Assistant Professor of Department of Electrical and Electronics Engineering, Faculty of Electrical and Computer Engineering, University of Sistan and Baluchestan, Zahedan, Iran Associate Professor of Department of Electrical and Electronics Engineering, Faculty of Electrical and Computer Engineering, University of Sistan and Baluchestan, Zahedan, Iran The properties of steels are intrinsically dependent on their microstructural components, known as phases, which form during the manufacturing process. Different steel phases can be observed in microscopic images of steel surfaces. Automatic detection and classification of these phases from images can significantly enhance the understanding of steel properties with improved speed and accuracy. This paper introduces, for the first time, an intelligent and automated method for classifying steel phases from microscopic images. This process requires defining and extracting suitable texture features unique to these images and segmenting the images into highly irregular regions based on the extracted features. To achieve this, the input image is initially divided into blocks, and texture features are extracted independently for each block. The dimensionality of these features is then reduced using Principal Component Analysis, and the refined features are subsequently fed into a Softmax neural network for classification. The implementation results indicate that the proposed method achieves an accuracy of over 99% in distinguishing between two phases: acicular ferrite and granular ferrite. Furthermore, it attains an accuracy exceeding 86% when classifying three phases: granular ferrite, acicular ferrite, and Widmanstätten ferrite. This suggests that the widely used and conventional k-means clustering method, as a traditional machine learning approach, is incapable of effectively distinguishing microscopic steel phase blocks using extracted texture features. Notably, as of the writing of this paper, no prior research has been conducted on the automatic classification of different ferrite phases, making this study a novel contribution to the field. In this research, an automated classification algorithm for ferrite phase structures in SEM images of steel is proposed using texture feature extraction methods and machine learning models. The dataset comprises images of 1024×768 resolution, which were divided into 128×128 blocks, with classification performed independently for each block. Due to the limited number of blocks available for training machine learning models, data augmentation techniques such as rotation and scaling were applied to increase the dataset size. Various image processing methods were used to extract 128 texture features. These extracted features were then used to classify different ferrite phases using two machine learning models: k-means clustering and the Softmax neural network. Additionally, PCA was employed to reduce feature dimensionality, which positively impacted the classification of granular and acicular ferrite. While k-means clustering, as a conventional and widely used machine learning method, failed to achieve satisfactory classification accuracy, the proposed approach using a smooth maximum neural network demonstrated exceptional performance. Despite the complex and irregular nature of ferrite shapes, the selected features and the proposed algorithm successfully achieved over 99% accuracy for two-phase classification and over 86% accuracy for three-phase classification. http://jsdp.rcisp.ac.ir/article-1-1400-en.pdf Scanning Electron Microscope Automated Classification Steel Phases Neural Networks K-Means Algorithm. per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 83 112 article Mohammad Hosein Bayat mhbayat@ut.ac.ir 1 Bahram Tarvirdizadeh bahram@ut.ac.ir 2 Mohammad Shahbazi shahbazi@iust.ac.ir 3 PhD Student, Department of Mechatronics Engineering, College of Interdisciplinary Science and Technology, Tehran University, Tehran, Iran Associated Professor, Department of Mechatronics Engineering, College of Interdisciplinary Science and Technology, Tehran University, Tehran, Iran Assistant Professor, School of Mechanical Engineering, Iran University of Science and Technology, Tehran, Iran Vision-based object tracking, as one of the most challenging fields in machine vision, means following the target(s) in the sequence of image frames in the presence of various challenges. In general, tracking algorithms can be classified to the single-target and multi-target based on the number of objects that should be tracked in frames. Trackers use two basic features in tracking: the appearance and motion. The appearance features are extracted from independent images but the motion features are produced through sequence of frames. According to the evaluations, motion models improve the tracking performance and take less process compared to the appearance features. Our investigations show that in contrast of single-target algorithms, the multi-target algorithms consider more contribution for the motion models, and due to the multiplicity of objectives in the scene they focus less on the appearance features. Despite the wide range of methods and significant progress in machine vision, reliable and flawless performance cannot be expected in the use of tracking algorithms with real-time criteria. This will be aggravated if one of the challenges occurs. Challenges such as sudden and fast movements by the target, occlusion by obstacles or other targets in the scene, extreme changes in the appearance and dimensions of the target, as well as entering and exiting the scene, which cause tracking algorithms to fail. Having a good trade-off between the accuracy and the execution speed is one of the main problems for applied tracking algorithms. Detection algorithms, which are known to detect different targets in an independent image, have shown acceptable accuracy, but it is not possible to use them in every frame for a real-time tracking, because either due to the high processing volume of these algorithms, the execution speed of the detector is limited or they are only able to identify certain classes. But the purpose of a general tracking is to follow an object in a sequence of images regardless of its type and class as well as considering temporal and spatial dependencies among successive frames. With the development of recurrent neural networks and their great ability to process sequential data such as text, audio and video, their use in tracking algorithms is increasing. The use of these networks has helped to improve the performance of tracking algorithms due to their short-term and long-term memory in maintaining important features during tracking. Different methods of integrating convolutional and recurrent neural networks are presented and showed grate performance in tracking, but the main drawback of most of them is the low execution speed of the algorithms. Our studies show that direct feeding the high-dimensional inputs, such as features extracted from images, to the recurrent networks greatly reduces their processing speed. Therefore, in some methods with the approach of real-time execution, the dimensions of the recurrent networks input are downsampled and reduced to the smaller size, although the accuracy is also slightly reduced. Our investigations show that the use of motion models in single-target tracking algorithms is less explored compared to the multi-target methods. Meanwhile, the studies show the success of these models in improving tracking performance. Before the introduction of convolutional networks and their remarkable success in extracting deep features from the image, motion models were mostly used, but in recent methods, especially in single-target trackers, appearance features are used more. In single-target algorithms, the presence of only one object in the image and less computational volume compared to multi-target algorithms allows for more free use of appearance features, but this is not possible in multi-target tracking due to the multiplicity of targets so the motion models are more useful in these algorithms. Therefore, in this paper, a more detailed investigation of motion models and their effect on tracking performance is done. The results show that motion models have a profound effect on improving tracking performance while being simple and impose low processing volume. In this paper, a comprehensive review and implementation of different tracking algorithms is discussed and appropriate methods are introduced for practical implementations. On the other hand, different tracking structures are investigated and categorized based on spatial, temporal, appearance and motion features. Also, due to the development of deep learning methods and their impact on tracking, deep architectures, training datasets and standard evaluation methods are studied and the future horizon of this field is discussed. Our studies show that temporal and motion features have received less attention despite their favorable impact on tracking performance. With the development of deep memory networks, the use of these features is increasing and they have taken a greater portion in tracking. http://jsdp.rcisp.ac.ir/article-1-1409-en.pdf Vision-Based Object Tracking Appearance Features Motion Features Deep Learning Machine Vision. per Research Center on Developing Advanced Technologies Signal and Data Processing 2538-4201 2538-421X 2025-05 22 1 113 141 article Mohammad Mahdi Haji-Esmaeili neltherion@gmail.com 1 Gholamali Montazer montazer@modares.ac.ir 2 PHD Student Industrial and Systems Engineering, Tarbiat Modares University, Tehran, Iran Professor Faculty of Industrial and Systems Engineering, Tarbiat Modares University, Tehran, Iran The rapid increase in the volume, diversity, and complexity of visual content in the digital world has made the need for designing and implementing visual content search and retrieval systems highly evident. Currently, we are facing a massive scale of visual data on the web, for which the conventional approaches based on manual and human-generated metadata are not sufficient to handle the diversity and sheer volume. The enormous volume of data generated on the web, without a high-accuracy and high-speed solution for understanding and retrieving it, will join the digital archives forever and never be found again. Recently, there have been significant efforts for retrieving these images, particularly in the fields of Content-Based Image Retrieval (CBIR) and Semantic Image Retrieval (SIR). Content-based and semantic image retrieval systems have the capability to search and retrieve images based on their internal content and high-level human-understandable semantics, rather than just the metadata that may be associated with them. This paper provides a comprehensive review of the latest advancements in the field of content-based image retrieval in recent years. It aims to critically discuss the strengths and weaknesses of each research area in content-based retrieval, and provide an overall framework of this process and the progress made in areas such as image preprocessing, feature extraction and embedding, machine learning, benchmark datasets, similarity matching, and performance evaluation. Finally, the paper presents novel research approaches, challenges, and suggestions for better advancing research in this field. The sections of the paper are organized as follows: After the introduction, Section 2 describes the components of a CBIR system framework, and with a cursory look at classical and traditional methods, it will delve into the workings of modern approaches and their associated challenges. In Section 3, we will provide an overview of the concept of "relevance feedback" and explain the need for this method to enhance the retrieval performance in CBIR systems, followed by an introduction to the prominent solutions in this domain. Finally, in Section 4, we will present a review of the image datasets commonly used in the field of content-based image retrieval, along with a discussion of their characteristics. IGiven the recent advancements in the field of computer vision and image processing, especially in the area of "image-text relationship" and how to integrate the two to improve retrieval performance, the focus of a large part of this study has been on the solutions in this area and the performance of the prominent methods. The current main research in this field is monopolized by large companies and organizations with access to vast financial resources, which has slowed down the progress of research and academic work in this field. These companies, with access to unimaginable data and financial resources, have trained well-known and sometimes unknown models on a very large scale (billions of images and texts), and after the training is complete, they have placed the final model in various web services without publishing the details of the research conducted. The important point is that the scale law applies in this field, and any entity that has more access to computational and storage resources will be able to train better and more accurate models, which has made it less possible for small research units and universities to enter this field and wait for the publication of research by the aforementioned organizations and companies. There is a dire need to introduce effective solutions in this field that require limited resources and are capable of achieving high accuracy and competitiveness with the massive models, with a fraction of the budget required to train them. This has happened in the field of large language models, and after two years, multiple research groups have been able to achieve the accuracy of the Chat-GPT4 language model from OpenAI and with the ability to run on home devices, and it is necessary for research in this field to shift from focusing on achieving accuracy with greater scale to focusing on achieving accuracy with lower cost, otherwise this field will remain in the monopoly of companies focused on greater profits http://jsdp.rcisp.ac.ir/article-1-1432-en.pdf Content-Based Image Retrieval Image Processing Computer Vision Machine Learning Deep Learning Semantic Gap.