Search

Coastal structures are prone to corrosion damage due to prolonged exposure to harsh corrosive environments, resulting in shortened service life and increased maintenance costs. Therefore, state-of-the-art technology is needed for rapid detection of corrosion in steel structures. This paper investigated the impact of training dataset quality on image recognition accuracy based on convolutional neural network technology. Multiple datasets were collected and used for neural network training, with the training results analysed using evaluation metrics. The research outcomes indicated that the quantity of data samples can improve at most by 1.2 times in terms of the average precision, while an excessive number of negative samples could interfere with the model’s ability to recognize positive samples, thereby reducing the precision. Instance segmentation datasets, compared to object detection datasets, were more sensitive to small objects due to their richer annotation information. This paper also proposed methods for improving datasets to enhance model prediction performance. The optimized model, combined with drone photography technology, was applied to practical engineering for bridge corrosion detection, providing technical support for intelligent bridge maintenance.

Deep learning has emerged as a prominent technique for extracting vegetation information from high-resolution satellite imagery. However, less attention has been paid to the quality of dataset labeling as compared to research into networks and models, despite data quality consistently having a high impact on final accuracies. In this work, we trained a U-Net model for tree cover segmentation in 30 cm WorldView-3 imagery and assessed the impact of training data quality on segmentation accuracy. We produced two reference tree cover masks of different qualities by labeling images accurately or roughly and trained the model on a combination of both, with varying proportions. Our results show that models trained with accurately delineated masks achieved higher accuracy (88.06%) than models trained on masks that were only roughly delineated (81.13%). When combining the accurately and roughly delineated masks at varying proportions, we found that the segmentation accuracy increased with the proportion of accurately delineated masks. Furthermore, we applied semisupervised active learning techniques to identify an efficient strategy for selecting images for labeling. This showed that semisupervised active learning saved nearly 50% of the labeling cost when applied to accurate masks, while maintaining high accuracy (88.07%). Our study suggests that accurate mask delineation and semisupervised active learning are essential for efficiently generating training datasets in the context of tree cover segmentation from high-resolution satellite imagery. [ABSTRACT FROM AUTHOR]

This research aimed to evaluate Mask R-CNN and U-Net convolutional neural network models for pixel-level classification in order to perform the automatic segmentation of bi-dimensional images of US dental arches, identifying anatomical elements required for periodontal diagnosis. A secondary aim was to evaluate the efficiency of a correction method of the ground truth masks segmented by an operator, for improving the quality of the datasets used for training the neural network models, by 3D ultrasound reconstructions of the examined periodontal tissue. Methods: Ultrasound periodontal investigations were performed for 52 teeth of 11 patients using a 3D ultrasound scanner prototype. The original ultrasound images were segmented by a low experienced operator using region growing-based segmentation algorithms. Three-dimensional ultrasound reconstructions were used for the quality check and correction of the segmentation. Mask R-CNN and U-NET were trained and used for prediction of periodontal tissue's elements identification. Results: The average Intersection over Union ranged between 10% for the periodontal pocket and 75.6% for gingiva. Even though the original dataset contained 3417 images from 11 patients, and the corrected dataset only 2135 images from 5 patients, the prediction's accuracy is significantly better for the models trained with the corrected dataset. Conclusions: The proposed quality check and correction method by evaluating in the 3D space the operator's ground truth segmentation had a positive impact on the quality of the datasets demonstrated through higher IoU after retraining the models using the corrected dataset. [ABSTRACT FROM AUTHOR]

Open and high-temporal- and spatial-resolution global land use/land cover (LULC) mapping data form the foundation of global change research and cross-scale land management planning. However, the consistency and reliability of the use of multisource LULC datasets in specific regions need to be quantitatively assessed. In this study, we selected the Indochina Peninsula as the research area; considered four datasets: LSV10, GLC_FCS30, ESRI10, and Globeland30; and analyzed them from four dimensions: the similarity of composition type, the degree of category confusion, spatial consistency, and data accuracy. The results show that: (1) the land composition descriptions of the different datasets are consistent. The study area is dominated by forest and cropland, supplemented by grassland, shrubland, and other land types. (2) The correlation coefficient between datasets is between 0.905 and 0.972; the spatial consistency of datasets is good; and the high-consistency area accounts for 77.87% of the total. (3) The overall accuracy of LSV10 is the highest (83.25%), and that of GLC_FCS30 is the lowest (72.27%). The accuracy of cropland, forest, water area, and built-up land is generally high (above 85%); the accuracy of grassland, shrubland, and bare land is low (below 60%). Therefore, researchers must conduct validation for specific regions and specific land types before using the above datasets. Our findings provide a basis for selecting LULC datasets in related research on the Indochina Peninsula and a reference method for assessing the reliability of multisource LULC datasets in other regions. [ABSTRACT FROM AUTHOR]

Buildings could help solve the supply and demand mismatch of electricity by providing energy flexibility to the grid. Thermal models of a dwelling are required to properly exploit their flexibility. The first step in developing such models is gathering on-site measurements. The quality of these measurements impacts the performance of the resulting models and, in turn, the harnessed flexibility. This research aims to evaluate the effect of the measurement sampling period on the achieved flexibility in a dwelling. In this paper, flexibility is realized through passive heat storage in the building by using model predictive control. The effect of the dataset sampling period on modeling accuracy and flexibility is evaluated. To assess the impact of the sampling period, four types of modeling techniques are trained by datasets with different sampling periods. The performance of these models for flexibility activation is assessed in two case studies. These case studies have similar structures with the main difference being their heating system. One building is equipped with electric heaters while the other uses underfloor heating. The findings show that decreasing the sampling period below 1 h does not lead to noticeable improvements in energy flexibility. On the other hand, deploying a sampling period larger than 1-h leads to a controller that underperforms. Using a sampling period of 2 h could sabotage the harnessed flexibility up to 51 %. • Applying feature selection to predictive models used for flexibility activation. • Assessing the impact of the data sampling period on the MPC-activated flexibility. • Inappropriate choice of the sampling period could reduce energy flexibility by 51 %. • High accuracy of a predictive model does not guarantee a high level of flexibility. [ABSTRACT FROM AUTHOR]

In deep learning, achieving high performance on image classification tasks requires diverse training sets. However, the current best practice-maximizing dataset size and class balance-does not guarantee dataset diversity. We hypothesized that, for a given model architecture, model performance can be improved by maximizing diversity more directly. To test this hypothesis, we introduce a comprehensive framework of diversity measures from ecology that generalizes familiar quantities like Shannon entropy by accounting for similarities among images. (Size and class balance emerge as special cases.) Analyzing thousands of subsets from seven medical datasets showed that the best correlates of performance were not size or class balance but A -"big alpha"-a set of generalized entropy measures interpreted as the effective number of image-class pairs in the dataset, after accounting for image similarities. One of these, A 0 , explained 67% of the variance in balanced accuracy, vs. 54% for class balance and just 39% for size. The best pair of measures was size-plus- A 1 (79%), which outperformed size-plus-class-balance (74%). Subsets with the largest A 0 performed up to 16% better than those with the largest size (median improvement, 8%). We propose maximizing A as a way to improve deep learning performance in medical imaging.

Open and high-temporal- and spatial-resolution global land use/land cover (LULC) mapping data form the foundation of global change research and cross-scale land management planning. However, the consistency and reliability of the use of multisource LULC datasets in specific regions need to be quantitatively assessed. In this study, we selected the Indochina Peninsula as the research area; considered four datasets: LSV10, GLC_FCS30, ESRI10, and Globeland30; and analyzed them from four dimensions: the similarity of composition type, the degree of category confusion, spatial consistency, and data accuracy. The results show that: (1) the land composition descriptions of the different datasets are consistent. The study area is dominated by forest and cropland, supplemented by grassland, shrubland, and other land types. (2) The correlation coefficient between datasets is between 0.905 and 0.972; the spatial consistency of datasets is good; and the high-consistency area accounts for 77.87% of the total. (3) The overall accuracy of LSV10 is the highest (83.25%), and that of GLC_FCS30 is the lowest (72.27%). The accuracy of cropland, forest, water area, and built-up land is generally high (above 85%); the accuracy of grassland, shrubland, and bare land is low (below 60%). Therefore, researchers must conduct validation for specific regions and specific land types before using the above datasets. Our findings provide a basis for selecting LULC datasets in related research on the Indochina Peninsula and a reference method for assessing the reliability of multisource LULC datasets in other regions.

The impact of data characteristics on dataset quality is of significant guidance in the field of thermal comfort field surveys. In this paper, a Chinese thermal comfort database has been established, comprising quality-controlled on-site survey data. The database encompasses 41,977 sets of on-site survey data collected from 50 cities across 24 provinces, spanning a period of nearly 20 years from 2001 to 2021. It includes the five thermal zones in China and encompasses three seasons—summer, transitional, and winter—as well as four types of buildings: office buildings, residential buildings, student dormitories, and classrooms. The database exhibits an extensive data distribution, demonstrating a certain level of representativeness in terms of environmental temperature and survey locations. Based on this established database, the thermal sensation model was employed as the criterion for evaluating dataset quality. This study analyzed the impact of data characteristics on dataset qualityfrom three perspectives: sample size, data distribution, and data range. A sample size calculation method using interval estimation was adopted, determiningthat a minimum sample size of 350 is required for on-site thermal comfort surveys in office and residential buildings. A comparison revealed that compared to data with the normal distribution and uniform distribution, data with the positively skewed distribution exhibit lower neutral temperatures, while data with the negatively skewed distribution exhibit higher neutral temperatures. The thermal sensation model constructed based on data with uniform distribution demonstrates greater robustness. As the temperature range decreases, the neutral temperature increases in winter and decreases in summer, resulting in reduced accuracy of the constructed thermal sensation model. [ABSTRACT FROM AUTHOR]

Human activity recognition (HAR) is a popular field of study. The outcomes of the projects in this area have the potential to impact on the quality of life of people with conditions such as dementia. HAR is focused primarily on applying machine learning classifiers on data from low level sensors such as accelerometers. The performance of these classifiers can be improved through an adequate training process. In order to improve the training process, multivariate outlier detection was used in order to improve the quality of data in the training set and, subsequently, performance of the classifier. The impact of the technique was evaluated with KNN and random forest (RF) classifiers. In the case of KNN, the performance of the classifier was improved from 55.9% to 63.59%.

Recommender systems based on collaborative filtering (CF) rely on datasets containing users’ taste preferences for various items. Accuracy of various prediction approaches depends on the amount of similarity between users and items in a dataset. As a heuristic estimate of this data quality aspect, which could serve as an indicator of the prediction ability, we define the Global User Correlation Measure (GUCM) and the Global Item Correlation Measure (GICM) of a dataset containing known user–item ratings. The proposed measures range from 0 to 1 and describe the quality of the dataset regarding the user–user and item–item similarities: a higher measure indicates more similar pairs and better prediction ability. The experiments show a correlation between the proposed measures and the accuracy of standard prediction models. The measures can be used to quickly estimate whether a dataset is suitable for collaborative filtering and whether we can expect high prediction accuracy of user-based or item-based CF approaches. [ABSTRACT FROM AUTHOR]

Osteoporosis causes harmful influences on both men and women of all races. Bone mass, also referred to as "bone density," is frequently used to assess the health of bone. Humans frequently experience bone fractures as a result of trauma, accidents, metabolic bone diseases, and disorders of bone strength, which are typically led by changes in mineral composition and result in conditions like osteoporosis, osteoarthritis, osteopenia, etc. Artificial intelligence holds a lot of promise for the healthcare system. Data collection and preprocessing seem to be more essential for analysis, so bone images from different modalities, such as X-ray, Computed Tomography (CT), and Magnetic Resonance Imaging (MRI), are taken into consideration that help to recognize, classify, and evaluate the patterns in clinical images. This research presents a comprehensive overview of the performance of various image processing techniques and deep learning approaches used to predict osteoporosis through image segmentation, classification, and fault detection. This survey outlined the proposed domain-based deep learning model for image classification in addition to the initial findings. The outcome identifies the flaws in the existing literature's methodology and lays the way for future work in the deep learning-based image analysis model., (Copyright© Bentham Science Publishers; For any queries, please email at epub@benthamscience.net.)

Human activity recognition (HAR) is a popular field of study. The outcomes of the projects in this area have the potential to impact on the quality of life of people with conditions such as dementia. HAR is focused primarily on applying machine learning classifiers on data from low level sensors such as accelerometers. The performance of these classifiers can be improved through an adequate training process. In order to improve the training process, multivariate outlier detection was used in order to improve the quality of data in the training set and, subsequently, performance of the classifier. The impact of the technique was evaluated with KNN and random forest (RF) classifiers. In the case of KNN, the performance of the classifier was improved from 55.9% to 63.59%. [ABSTRACT FROM AUTHOR]