Navegando por Assunto "Aprendizado profundo"

Agora exibindo 1 - 6 de 6

Acesso aberto (Open Access)
Classificação de regiões de desmatamento via imagens do satélite landsat no nordeste do Pará
(Universidade Federal do Pará, 2023-12-18) CANAVIEIRA, Luena Ossana; COSTA, João Crisóstomo Weyl Albuquerque; http://lattes.cnpq.br/9622051867672434
Acesso aberto (Open Access)
Deep learning in education 5.0: proposing 3d geometric shapes classification model to improve learning on a metaverse application
(Universidade Federal do Pará, 2024-01-18) SANTOS, Adriano Madureira dos; SERUFFO, Marcos César da Rocha; http://lattes.cnpq.br/3794198610723464; https://orcid.org/0000-0002-8106-0560
The Brazilian educational system faces significant challenges, as evidenced by low educational development assessment scores. Due to the traditional educational model employed in the country, there are difficulties in the effective transmission of complex content, leading to high rates of academic failure and subsequent school dropout. The lack of innovation, especially in basic education settings, contributes to a scenario of low mathematical proficiency among Brazilian students. In this context, this work arises as a result of an innovation built to enhance the Geometa application, developed by the Inteceleri company, through the integration of Metaverse and Artificial Intelligence technologies to create an immersive and interactive educational environment. The intention is to train Artificial Intelligence for real-time three-dimensional geometric shape recognition from real-world object images. The proposal aims to mitigate challenges faced in Brazilian basic Mathematics education by adopting innovative technological approaches aligned with Education 5.0, which can be replicated for similar technologies involving the Metaverse. Furthermore, it is also intended to create a dynamic and sustainable educational environment that not only facilitates the mathematical concepts understanding but also promotes active student participation, encouraging their creativity and autonomy in the learning process. The method used relies on the ObjectNet dataset image reclassification from objects to three-dimensional geometric shapes. The reclassified images are used to train CNN, MobileNet, ResNet, ResNeXt, ViT and BEiT Deep Learning models, which are subsequently evalua ted through Machine Learning, inference time and dimension performance measures. Thus, the best-performance Artificial Intelligence model is selected for future integration into Geometa. As contributions of this work, the following were accomplished: (i) the defined models were trained for the three-dimensional geometric shapes recognition; (ii) the models were evaluated through Machine Learning, inference time and dimension performance measures; and (iii) the best-performance model was selected considering the highest assertiveness and smoothness based on models performances analysis. Concerning the obtained results, the ResNet surpassed BEiT, which was the second better performance model, in 5% Precision and 5 Inference Per Second. Finally, the ResNet model reached 84% Precision and 9 Inferences Per Second, being observed as the best-performance Artificial Intelligence for Geometa application integration flow.
Acesso aberto (Open Access)
Detecção de erosão em taludes baseada em deep learning
(Universidade Federal do Pará, 2023-03-31) LIMÃO, Caio Henrique Esquina; FRANCÊS, Carlos Renato Lisboa; http://lattes.cnpq.br/7458287841862567
The recent catastrophes triggered by the rupture of the Fundão and Córrego do Feijão dams caused around 300 deaths and countless irreparable socio-environmental damages. Since the use of more accurate monitoring systems and the proper execution of preventive and corrective maintenance would allow identifying, and even mitigating, the damage caused to society, it can be stated that there is a need for greater investment and incentive to create solutions of Structural Health Monitoring (SHM) capable of diagnosing occurrences that compromise the most crucial civil structures, such as bridges, buildings, dams and slopes. High-performance Artificial Intelligence (AI) techniques have been able to solve these structural analysis problems and presented superior results to previous solutions, their use has increased dramatically in the most diverse (SHM) scenarios. When it comes to image analysis and classification solutions, Convolutional Neural Network (CNN) is the type of neural network that delivers the best results. Therefore, this dissertation will describe the development process of a CNN with three convolutional layers that combines the use of the most consolidated technologies in the current scenario of computer vision, such as the Adam optimizer and batch normalization. The proposed CNN was trained with a database set up specifically for this dissertation, consisting of images of public work reports made by the Brazilian government, portfolios of companies that work with construction and maintenance of slopes and reports on landslides and/or catastrophes. These images were labeled, according to the context of each one of them, as stable or instable slopes. The results obtained were quite satisfactory, presenting an accuracy of 96.67% and proving that this solution is capable of identifying in a precise and improved way the instability indicators presented by the analyzed slopes, allowing a more adequate planning of the maintenance for each case, in the prevention of possible disasters, more efficient manpower management, cost reduction, greater safety and structural health to ensure its long-term integrity.
Acesso aberto (Open Access)
Machine learning algorithms for damage detection in structures under changing normal conditions
(Universidade Federal do Pará, 2017-01-31) SILVA, Moisés Felipe Mello da; SALES JÚNIOR, Claudomiro de Souza de; http://lattes.cnpq.br/4742268936279649; COSTA, João Crisóstomo Weyl Albuquerque; http://lattes.cnpq.br/9622051867672434
Engineering structures have played an important role into societies across the years. A suitable management of such structures requires automated structural health monitoring (SHM) approaches to derive the actual condition of the system. Unfortunately, normal variations in structure dynamics, caused by operational and environmental conditions, can mask the existence of damage. In SHM, data normalization is referred as the process of filtering normal effects to provide a proper evaluation of structural health condition. In this context, the approaches based on principal component analysis and clustering have been successfully employed to model the normal condition, even when severe effects of varying factors impose difficulties to the damage detection. However, these traditional approaches imposes serious limitations to deployment in real-world monitoring campaigns, mainly due to the constraints related to data distribution and model parameters, as well as data normalization problems. This work aims to apply deep neural networks and propose a novel agglomerative cluster-based approach for data normalization and damage detection in an effort to overcome the limitations imposed by traditional methods. Regarding deep networks, the employment of new training algorithms provide models with high generalization capabilities, able to learn, at same time, linear and nonlinear influences. On the other hand, the novel cluster-based approach does not require any input parameter, as well as none data distribution assumptions are made, allowing its enforcement on a wide range of applications. The superiority of the proposed approaches over state-of-the-art ones is attested on standard data sets from monitoring systems installed on two bridges: the Z-24 Bridge and the Tamar Bridge. Both techniques revealed to have better data normalization and classification performance than the alternative ones in terms of false-positive and false-negative indications of damage, suggesting their applicability for real-world structural health monitoring scenarios.
Acesso aberto (Open Access)
Metodologia para a classificação automática de doenças em plantas utilizando redes neurais convolucionais.
(Universidade Federal do Pará, 2019-11-07) REZENDE, Vanessa Castro; OLIVEIRA, Roberto Célio Limão de; http://lattes.cnpq.br/4497607460894318; SANTOS, Adam Dreyton Ferreira dos; http://lattes.cnpq.br/2616572481839756
Convolutional neural networks (CNNs) are one of the deep learning techniques that, due to the computational advancement of recent years, have enhanced the field of computer vision by enabling substantial gains in various classification problems, especially those involving digital images. Given the advantages of using these networks, a variety of applications for automatic plant diseases identification have been developed for specialized assistance or automated screening tools, contributing to more sustainable farming practices and improved food production security. In this context, this work aims to propose a methodology for the classification of multiple pathologies from distinct plant species, having as input a database composed of digital images of plant diseases. Initially, this methodology involved image preprocessing activities on the plant disease database to provide the appropriate input for selected CNN models (VGG16, RestNet101v1, ResNet101v2, ResNetXt50 and DenseNet169), as well as to generate ten new bases, ranging from 50 to 66 classes with greater representativeness, to submit the models to different situations. After model training, a comparative study was conducted based on widely used classification metrics such as test accuracy, f1-score, and area under the curve. To attest the significance of obtained results, the Friedman nonparametric statistical test and two post-hoc procedures were performed, which showed that ResNetXt50 and DenseNet169 obtained superior results when compared with VGG16 and ResNets.
Acesso aberto (Open Access)
Redes neurais convolucionais aplicadas à inspeção de componentes do vagão ferroviário
(Universidade Federal do Pará, 2020-02-03) ROCHA, Rafael de Lima; GOMES, Ana Claudia da Silva; http://lattes.cnpq.br/9898138854277399; SILVA, Cleison Daniel; http://lattes.cnpq.br/1445401605385329; https://orcid.org/0000-0001-8280-2928
The railcar is one of the most important assets in a mining company, where tons of ore are transported daily by it, besides, the railcar can be used to transport people. Therefore, the inspection of defects in structural components of the railcar is a very important activity, making it possible to avoid problems in railway logistics, as well as to prevent accidents. The inspection task is performed visually by an operating technician who is exposed to accidents where the inspection is performed, in addition to the possibility of human error due to stress, fatigue, and others. The pad is a rail component analyzed in this work, where it is responsible for the primary suspension, a role that is important in the railcar dynamics. Thus, the purpose of this work is to use deep learning techniques, specifically convolutional neural networks (CNN) for the component inspection. CNN classifies the image of the structural component analyzed concerning the possible state it is in the railway, absent pad, undamaged pad, and damaged pad. Also, it intends to investigate the contribution of the component image in the frequency domain obtained through the magnitude and phase of the discrete Fourier transform (DFT) of the original image (spatial domain) in the CNN classification process. Histogram equalization and increasing the number of images through data augmentation techniques are also examined to evaluate their collaborations in improving classification performance. The results of CNN inspection of the pad prove to be quite inspiring, especially when the spatial component image is used together with the DFT magnitude image of the original image as CNN inputs, which are superior when only the original (spatial) image of the component is used, achieving a classification accuracy of 95.65%. In particular, the method that uses the increase in the number of training images by the data augmentation and the spatial domain and frequency (magnitude) images achieves the highest accuracy, with 97.47%, which represents approximately 385.5 correctly classified images from a total of 395.2 images.