Melanoma is a very dangerous type of skin cancer and its detection in the early stages is necessary for proper treatment. The article proposes an intelligent system, based on the fusion of the decisions of several neural networks to increase the performance in melanoma detection from dermatoscopic images. The system implementation method is based on the optimal choice of the number and type of neural networks involved by testing the possible combinations. According to the selection procedure, a system was implemented with four neural networks DenseNet 201, VGG 19, MobileNet, and EfficientNet. The results obtained on two different databases (ISIC 2019 - for learning, validation, and testing - and PH2 for testing) were better than those obtained on individual networks or in other works from literature.

The automatic recognition of real people is particularly important in many fields of activity. The article proposes a complex system for recognizing people taking into account possible fraud attacks. A Long Short-Term Memory neural network was used to detect possible frauds in the video sequences, and to recognize the person in the case after validation as a real person, two Siamese neural networks implemented with VGG 16 and ResNet 50 were tested. The obtained results were compared for Euclidean distance, Cosine distance, and for two training methods: Contrastive Loss and Triple Loss. The best results were obtained for VGG 16, Cosine distance, and Triple Loss.

Nowadays, multidisciplinary research is found everywhere, especially due to the accelerated development of computer science and information and communication technology in all economic and social fields. In many universities, including the POLITEHNICA University in Bucharest, research as a subject is considered multidisciplinary through case studies and concrete applications. The paper presents a multidisciplinary project implemented under ERA-NET COFOUND, ICT-AGRI-FOOD 2019 Joint Call. This project implies multiple collaborations inside and between different groups: disciplines (computer science-image processing-artificial intelligence-robotics-entomology-agriculture), domains (universities-research-farms), countries, and teams (teachers-researchers-students-farmers). The resulting impact of this collaboration is also multiple: educational, economic, and social. From an educational point of view, the results consisted of course chapters, themes for undergraduate, master, and doctoral degrees, and papers in important international conferences or journals made by teams of students, researchers, and professors.

Artificial intelligence applications are recently showing promising results in advancing prenatal ultrasound examination. Regarding deep learning contribution for early pregnancy evaluation in ultrasound images, research is still in its infancy. There are some ultrasound elements whose measurements offer essential information on early pregnancy. Proper determination of the gestational age relates mainly to the dimension of the gestational sac and the length of the embryo. In this regard, semantic segmentation can lead to improved imagistic diagnosis by automatized measurement of the mentioned structures. The paper presents a precise pregnancy detection system in ultrasound images based on artificial intelligence by proposing various combined convolutional neural networks focused on semantic segmentation tasks of some regions of interest like an embryo and gestational sac and selecting the most accurate ones in terms of multiple performance metrics. The proposed convolutional neural networks are DeepLabV3+networks based on multiple state-of-the-art convolutional neural networks, used as encoders, such as ResNet-50, ResNet-18, InceptionResNet-V2, Xception, and MobileNet-V2.

The paper's goal was to create some neural networks-based models for the detection and classification of insects such as Halyomorpha Halys in ecological orchards, from acquired images in the trees. The detecting operations were performed using models from two of the most efficient deep learning families in this area: R-CNN and YOLO. Using the proposed models, (Faster R-CNN, YOLOv5-s, YOLOv5-m, and YOLOv5-1) to early detection of harmful insects, a real contribution to anticipating damage in orchards is possible. The dataset is composed of images taken from the Maryland Biodiversity dataset. All training and testing operations were performed with the help of GPU processors provided by Google, the resulting models being saved on Google Drive Cloud. The images were evaluated from the detection and the classification perspective based on specific metrics such as precision, recall, and mAP. The best results were obtained for YOLOv5-m.

The connected projects MUROS-MAARS are funded by the Romanian Space Agency program STAR and it is implemented through collaboration between the University POLITEHNICA of Bucharest and three industrial partners. The paper highlights the educational aspect and impact of this type of collaboration. The main goal of the project is to implement a hybrid system of wireless sensor networks and unmanned aerial vehicles integrated in IoT to monitor regions of interest at ground level. The projects had a positive impact on the educational process expressed by the interest of the students and academic staff for the project implementation. Courses and applications for the industrial partners and for the students, were developed. The results of the educational activities were themes for bachelor, master, and doctoral degrees. Result dissemination in research papers at prestigious international conferences or journals was made by teams of students, industry specialists, and professors.

To reduce the damage caused by Halyomorpha Halys (HH) in ecological orchards, the first step is to detect and identify this insect species. The paper's goal was to evaluate the effectiveness of convolutional neural networks (CNNs) in detecting harmful insects such as HH in orchards. The images were acquired by a Mavic 2 Pro DJI drone. To detect HH from images, four CNNs were tested and compared: GoogLeNet, ResNet101, DenseNet201, and VGG19. Transfer learning and data augmentation were used to minimize computational effort in the learning phase. For each proposed CNN, the detection and classification performances are evaluated in the testing phase based on statistical indicators derived from confusion matrices. Also, the learning time was considered.

The paper presents an unmanned aerial system (UAS), based on multi-UAV (unmanned aerial vehicle) architecture, integrated in Internet, able to collect data from ground wireless sensor networks (WSNs), deployed on large areas for prevention agriculture. The contributions were focused on the way of mission planning and data collection: navigation algorithms, transfer of flight trajectory in UAS, storing trajectories to UAS ground part, and data transfer between WSN and UAV. The experimental results, focusing on communication performances, prove the real time data collection and high accuracy, both in simulation environment and real experiments on crop monitoring.

The most convenient method to assess flood damage in rural areas is to analyze the images taken over by a UAV (unmanned aerial vehicle) team. The paper presents such an aerial unmanned system, implemented by the authors in a research project. The images are directly transmitted via internet to the image processing sub-system. After creating an orthophotoplan, the images are partitioned in patches and then a convolutional neural network is used to classify the patch pixels in flooded type or non-flooded. A set of 100 images with flooded and non flooded zones was used and corresponding 5000 patches (3000 for the learning phase and 2000 for the testing phase). The experimental results show good performances regarding the accuracy and the calculation time.

The paper presents a unmanned aerial system (UAS), based on a multi-UAV (unmanned aerial vehicle) architecture, integrated in Internet, able to collect data from ground wireless sensor networks (WSNs), deployed on large areas. The contributions were focused on the way of mission planning and data collecting: navigation algorithm, transfer of flight trajectory in UAS, storing trajectories to UAS ground part, and data transfer between WSN and UAV. The system can be used for aerial monitoring in many civilian fields of activity. The experimental results, focusing on communication performances, prove the real time data collection in different scenarios with the ground control station at distance (internet connection).

Wireless Sensor Network (WSN) has been a focus for research in the last years due to the promising technology it embeds. This appears to be the most sustainable technology for environmental sensing whether it's about limited or large-scale monitoring, thanks to the ad-hoc wireless links, scalability and ease of implementation. However, main drawbacks are stemming from the limited capacity of network nodes for data storage, computing and accessing. To overcome these limitations, virtualized resources were appended allowing access to increased storage, processing and user-friendly accessibility. This came as a natural development of the common WSN architectures in the trend of modern concepts emerged with the IoT (Internet of Things) technologies proliferation. Despite the increasing usage of cloud-based WSN monitoring systems, there are still issues due to the drawbacks of cloud computing such as latency and storage costs. This paper discusses the improvements made to a cloud-based WSN architecture by adding a layer of computing at the edge of the network, a method that follows the novel model of analysing and acting on IoT data, entitled Fog Computing. Comparative analytics were performed to prove the improvements achieved through edge of the network computing.

Designing and deploying a safe and reliable wireless sensor network involves deep knowledge of the propagation environment, as the performance can be compromised by unknown or unpredictable factors like interference, vegetation, terrain or excessive humidity. An accurate propagation model is essential for developing a cost-effective wireless sensor network, however, a guide that will allow the frequency selection based on some input data that can be considered as default constraints for any RF planner (e.g. distance between sensors, lack of electricity, minimum throughput, etc.) can always serve as an alternative to rather expensive RF propagation studies.

This paper presents a novel evaluation method of areas affected by natural disasters with the purpose of managing these crisis situations. Since it is necessary to have a real overview of a specific area in the shortest time, our methodology proposes a neural network with backpropagation approach for flood detection from UAV images. For this, the Local Binary Pattern (LBP) texture operator is used for areas classification. The LBP operator labels each pixel of the analyzed image by comparing it with its neighbors, which ends with the computation of a binary number that it is converted to decimal format named LBP code. Thus, based on the generated LBP codes, a histogram type feature is computed and used in both training and testing phases of the proposed neural network. Over 50 images obtained with the aid of UAV technology were tested with the proposed neural network and good results in terms of accuracy for flood areas detection were obtained.

The detection, localization and evaluation of small flooded areas can contribute to decrease the economical damages of such disasters. The cheapest and most accurate method is to segment the aerial images taken from UAV. In this paper, we propose a new method for detection of regions of interest, like flooding in rural areas, using Generative Adversarial Networks (GAN) and Graphics Processing Units (GPU). The classical GPU is used to create, by parallel calculation of textural features, extracted from the co-occurrence matrix, the supervised mask of flood segmentation in the images from the learning set. Based on these images and their associated real masks, the weights of the generator and discriminator are established. A set of 40 images were used for the learning phase and another set of 60 images were used for method validation. The results demonstrate that the proposed method provide a high accuracy and robustness, comparing with other papers for flooding evaluation. Even if it is a relative long time to learn the GAN, in the operational phase the time for image segmentation process is very short.

In this paper we consider a multi-UAV formation whose goal is to efficiently gather data from sensors deployed in a cluttered environment while in the same time keeping communication with ground terminals. We formulate these requirements and constraints as a nonlinear constrained optimization problem and recast them in a mixed-integer form with the help of a hyperplane arrangement construction. Particular attention is given to line-of-sight constraints which ensure permanent communication between UAVs and ground terminals. Reference trajectories are generated in simulation over illustrative examples.

Wireless Sensor Networks (WSNs) comprising a large number of sensing nodes deployed within the area of interest, are able to measure, process and share specific parameters. Besides enabling effective area coverage, recent research has proven that unmanned aerial vehicles (UAVs) represent a viable addition to large area monitoring through remote sensing and data collecting functions. The proposed interlinking between autonomous UAV and on-ground WSNs overcomes the limitations that prevent sensing nodes from adequately managing large scale applications. This paper presents an overview regarding state of art WSNs and UAVs technologies used for large area monitoring, introduces a novel approach for smart data collecting and further explores the in-network data paradigm. Various simulations have been implemented and analyzed from a comparative standpoint.

This paper proposes an improved method for automated segmentation of images containing small flooded areas, in order to evaluate the material damage in rural zones. The solution consists on a hybrid wireless sensor network composed of two parts: the aerial mobile nodes (for surveillance and monitoring of flood affected areas) and the fixed nodes at the ground (for control, image processing and flood area evaluation). By appropriate design of trajectories, the mobile nodes cover the entire surface to be evaluated. The methodology for remote image segmentation is based on combining the features extracted from co-occurrence matrices, on different color channels, with the fractal features and then by searching of the similarity with the representatives of the classes. The method gives good results on small areas of flood.