Deep learning became an ubiquitous paradigm due to its extraordinary effectiveness and applicability in numerous domains. However, the approach suffers from the high demand for data required to achieve the potential of this type of model. An ever-increasing subfield of Artificial Intelligence, Image Synthesis, aims to address this limitation through the design of intelligent models capable of creating original and realistic images, endeavor which could drastically reduce the need for real data. The Stable Diffusion generation paradigm recently propelled state-of-the-art approaches to exceed all previous benchmarks. In this work, we propose the ContRail framework based on the novel Stable Diffusion model ControlNet, which we empower through a multi-modal conditioning method. We experiment with the task of synthetic railway image generation, where we improve the performance in rail-specific tasks, such as rail semantic segmentation by enriching the dataset with realistic synthetic images.

In autonomous driving, a tremendous amount of imagery data is collected at all times. Manual annotation of such high-resolution images represents a costly and inefficient process. Active Learning comes to aid annotators in their process to focus on labelling meaningful samples which leads to competitive Machine Learning models, useful for various prediction tasks. In this paper, we introduce a novel Active Learning sampling technique, inspired by the Ant Colony Optimization algorithm, that considers both uncertainty and diversity features. We also introduce two hybrid sampling techniques that use weighted sums. We validate the proposed method on the Semantic Segmentation task, on a popular dataset from the railway domain. We also showcase the effectiveness of Active Learning in the scenario of Rail Semantic Segmentation by using only a quarter of the data to obtain competitive results of up to 78% mean Intersection over Union.

Breast cancer remains a critical global health concern, given the crucial role of early detection in achieving successful treatment outcomes. By harnessing the power of deep learning, our proposed methodology aims to discern intricate patterns and nuances in breast tissue textures, enabling robust discrimination between benign and malignant tumors. We start to search for solutions using two directions: an intelligent system that uses Convolutional Neural Networks (CNNs) over the images and another model that uses CNNs over textural features extracted from mammograms. This research comes as an extension to our previous work on the Classification of mammograms into benign and malignant types using textural features and shallow classifiers. While these methods provided valuable insights, we sought to explore the future of CNNs in increasing the accuracy of breast cancer detection.

SENTIROM introduces an innovative contribution to sentiment analysis research in the Romanian language, a field previously limited by a lack of resources and computational tools. Focusing on enhancing the LaRoSeDa dataset with advanced NLP techniques like Word2Vec, K-Means clustering, and a specialized BERT configuration with a Romanian tokenizer, this study seeks to bridge the gap in Romanian linguistic analysis. Through experimenting with various computational models and comparing the results to an English dataset, SENTIROM aims to demonstrate the viability and superiority of these approaches for understanding and processing Romanian sentiment. Key contributions include the development of a system that can accurately classify sentiment in Romanian text, proposing a methodology to streamline data collection for future high-quality corpora, introducing a novel BERT and K-Means based approach outperforming existing models, and evaluating the dataset’s translatability and comparability to English sentiment analysis frameworks.

The USA presidential election in 2020 has aroused the interest of society as a whole. Social media plays an essential role in campaigns due to how people express their ideas about a candidate. Therefore, analyzing messages posted on environments like Twitter is a challenging task that can be accomplished using sentiment analysis techniques. Our approach focuses on clustering algorithms that can determine positive and negative tweets related to the presidential candidates Joe Biden and Donald Trump.In addition, a new model is defined for representing a tweet, called hash index, by using the hashtag feature offered by Twitter. On the other hand, external validation is applied to the collected tweet by setting the sentiment derived from the Vader lexicon. Consequently, the experiments are evaluated using internal metrics like the Silhouette index or external ones like accuracy metrics.The experimental results show interesting achievements for the unsupervised approach that can determine a candidate’s popularity during the 2020 election.

Breast cancer is one of the most common types of cancer amongst women, but it is also one of the most frequently cured cancers. Because of this, early detection is crucial, and this can be done through mammography screening. With the increasing need of an automated interpretation system, a lot of methods have been proposed so far and, regardless of the algorithms, they all share a step: segmentation. That is, identifying the region of interest in order to further analyze and classify it either as benign or malignant. However, due to the different types of mammary tissues, mammography segmentation can prove to be a difficult task. Various techniques of mammography segmentation have been proposed so far. Yet, since obtaining the ground-truth for mam-mographic images might be problematic, recent literature leans towards unsupervised techniques. In this paper we present a segmentation approach based on the GrowCut algorithm. The original method starts with a number of seed points inside and outside the region of interest, selected by a human expert, and iterates over the pixels multiple times, until it reaches a stable state where all the pixels have been assigned to a class. Our proposal aims to reduce the human intervention by eliminating the need of selecting initial background seeds and, also, to reduce the computational time by limiting the process of parsing the entire image to a small, fixed number of iterations. The proposed approach was compared to the original method, using three variants: (1) automatically constructing the initial background seeds surrounding the foreground ones; (2) using mammograms’s background as initial background seeds; (3) not using initial background seeds. Experimental results obtained for the Mini-MIAS dataset show that, for the variations that use background seeds outside the breast and that do not use any background seeds, our approach yields much better results than the original method.

As the world changes, so does the future of our students. In this respect, the evolution of the technology comes up with specific environments for educational purpose. Building smart learning environments supported by e-learning platforms is an important area of research in education domain within our days. The evolution of these smart learning environments is justified by some events (Covid19) that force students to learn remotely. The paper proposes a formalisation using ontology for providing an inclusive approach of universities’ websites, having as instance a software application component using Alexa smart speaker, that currently remains at a design level, which integrates different services (Amazon Web Services, Microsoft Services) for a proper virtual environment platform, for both students and teachers. It addresses the main concerns of the current educational system and provides a smart solution through the use of Artificial Intelligence based tools. The proposed approach not only achieves unifying data and knowledge-share mechanisms in a remotely mode, but it brings also a good learning experience, increasing the effectiveness and the efficiency of the learning process.

Edge detection is an important component in many computer vision tasks since edges convey information about the objects in an image. This paper presents a comparative analysis of the proposed edge detector with respect to one of the state-of-the-art methods, the Canny edge detector. Our edge detection model involves the supervised optimization of a cellular automaton rule with particle swarm optimization. Using this scheme we obtain transferable rules that can be applied on multiple images with similar properties. We test the two methods on clean and noisy images and the proposed method outperforms Canny on average on our data set containing a variety of edges.

Extensive research has been performed in image processing to find the best edge detector, from the gradient-based operators to evolved Cellular Automata (CA). Some of these detectors have weak points, such as disconnected edges, the incapacity of detecting the branching edges or the need of a ground truth that is not always available. To overcome these issues, we propose a CA-based edge detector adapted to the particularities of the image. The adaption means to identify the best CA rule, which is an optimization problem solved by a Genetic Algorithm (GA). The GA requires a fitness function and we propose to use an unsupervised fitness based on edge dissimilarity. The performed numerical experiments are meant to evaluate the proposed approach and to emphasize that some of the weak points of a well-known detector (Canny) can be overcome by our method.

Children of nowadays grow in a digital landscape, so education has embraced the advantages brought by the multimedia technology progress. Appropriate interactive learning experiences positively influence learners’ performance. However, new challenges occur when the learners are preschoolers, as they are not able to articulate and communicate their experience towards interaction with edutainment applications. In this paper we explore the appropriateness of using Machine Learning-techniques in identifying preschoolers’ emotions while interacting with edutainment applications. The investigated scenarios reveal promising directions for assessing children satisfaction with edutainment applications.

Mobile applications are software systems that are highly used by all modern people; a vast majority of those are intricate systems. Due to their increase in complexity, the architectural pattern used plays a significant role in their lifecycle. Architectural patterns can not be enforced on a codebase without the aid of an external tool; with this idea in mind, the current paper describes a novel technique for an automatically analysis of Model View Controller mobile application codebases from an architectural point of view. The analysis takes into account the constraints imposed by this layered architecture and offers insightful metrics regarding the architectural health of the codebase, while also highlighting the architectural issues. Both open source and private codebases have been analysed by the proposed approach and the results indicate an average accuracy of 89.6% of the evaluation process.

Complex networks are data structures with great importance in representing real world interactions which surrounds us. While their structures might look chaotic at a first glance, the focus of most on-going studies in this field is in understanding how their topological properties influence the dynamics of a complex network’s structure in order to prove a possible order in the apparent chaos that they display. Based on the evidence found in our previous studies, which revealed a significant correlation between the existence of articulation points and meso-level components such as network motifs, this paper tries to extend this study by presenting analytical research between a consistent set of micro-level topological properties from Graph and Complex Networks Theory and the appearance of the motifs. The purpose of this study is to use network properties to provide a better understanding of how and why network motifs appear, a further step toward the goal of proposing a generator model for networks with specific concentrations of motifs.

Due to the increased number of mobile applications and their popularity, many software developers have begun to focus on mobile platforms.While this focus has positive effects (e.g. a larger developer community, new open source projects, new tools), it also has a down side.With the migration of developers from different software development areas, where they have used other programming paradigms or architectural approaches, the topic of software architecture on mobile platforms become more trending and hype in the mobile development communities.Even though several new architectural solution were proposed for solving some of the issues which arise from using the classical architectural patterns popularised by the creators of the mobile platforms, we want to emphasise the principles of software architecture in mobile computing, why they have to be respected and how their adoption impacts the development process.Therefore, this paper focuses on showing that the Model View Controller (MVC) -one of the most common classical architectural patterns -can be used successfully for building mobile applications and the problems which might arise are by products of the wrong usage of the pattern rather than pattern issues.We show that by analysing the most common architectural misuses of the MVC pattern in both open-source and private projects and offers solutions to those problems.

Classical kernel-based classifiers only use a single kernel, butthe real world applications have emphasized the need to con-sider a combination of kernels also known as a multiple kernel in order to boost the performance. Our purpose isto automatically find the mathematical expression of a multiple kernel by evolutionary means. In order to achieve this purpose we propose a hybrid model that combines a Genetic Programming (GP) algorithm and a kernel-based Support Vector Machine (SVM) classifier. Each GP chromosome isa tree encoding the mathematical expression of a multiple kernel. Numerical experiments show that the SVM embedding the evolved multiple kernel performs better than the standard kernels for the considered classification problems.

The result of the program encoded into a Genetic Programming(GP) tree is usually returned by the root of that tree. However, this is not a general strategy. In this paper we present and investigate a new variant where the best subtree is chosen to provide the solution of the problem. The other nodes (not belonging to the best subtree) are deleted. This will reduce the size of the chromosome in those cases where its best subtree is different from the entire tree. We have tested this strategy on a wide range of regression and classification problems. Numerical experiments have shown that the proposed approach can improve both the search speed and the quality of results.

hybrid model for evolving support vector machine (SVM) kernel functions is developed in this paper. The kernel expression is considered as a parameter of the SVM algorithm and the current approach tries to find the best expression for this SVM parameter. The model is a hybrid technique that combines a genetic programming (GP) algorithm and a support vector machine (SVM) algorithm. Each GP chromosome is a tree encoding the mathematical expression for the kernel function. The evolved kernel is compared to several human-designed kernels and to a previous genetic kernel on several datasets. Numerical experiments show that the SVM embedding our evolved kernel performs statistically better than standard kernels, but also than previous genetic kernel for all considered classification problems.

According to the No Free Lunch (NFL) theorems all black-box algorithms perform equally well when compared over the entire set of optimization problems. An important problem related to NFL is finding a test problem for which a given algorithm is better than another given algorithm. In this paper we propose an evolutionary approach for solv- ing this problem: we will evolve multi-objective test func- tions for which a given algorithm A is better than another given algorithm B. The evolved functions are represented as binary strings. Several numerical experiments involv- ing PESA and NSGA II are performed. The results show the effectiveness of the proposed approach. Several multi- objective problems for which PESA performs better than NSGA II and several multi-objective test problems for which NSGA II performs better than PESA have been evolved.