Article
Engineering, Electrical & Electronic
Yuan Chen, Chang Liu, Yunjie Yang, Mathieu Lucquiaud, Jiabin Jia
Summary: This study investigates the feasibility of using electrical capacitance tomography (ECT) and convolutional neural networks (CNNs) as an intensified alternative to conventional flooding prediction methods. ECT combined with CNNs allows for more accurate calculation of liquid hold-up, especially at high gas flow rates.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2022)
Article
Computer Science, Information Systems
Jose P. Amorim, Pedro H. Abreu, Joao Santos, Marc Cortes, Victor Vila
Summary: This study introduces a new approach for evaluating saliency map methods, which have various applications in medical image classification. By introducing natural perturbations and studying their impact on evaluation metrics adapted from saliency models, the effectiveness of this method is validated. The results show that this approach accurately evaluates the reliability of saliency map methods and has potential for application in other medical imaging tasks.
INFORMATION PROCESSING & MANAGEMENT
(2023)
Article
Environmental Sciences
Muhammad Afaq Hussain, Zhanlong Chen, Ying Zheng, Yulong Zhou, Hamza Daud
Summary: The study provides an updated inventory of landslides on the Karakoram Highway using SBAS-InSAR and PS-InSAR technology. The landslide occurrences and causal variables are investigated to create a landslide susceptibility model. The CNN 2D technique shows superior performance in creating the landslide susceptibility map.
Article
Construction & Building Technology
Man Wang, Zhe Wang, Yang Geng, Borong Lin
Summary: This study proposes a method to interpret neural network models using gradients, quantifying the influence of inputs on outputs. The method reduces calculation time and provides reasonable and informative feature importance compared to other models.
BUILDING AND ENVIRONMENT
(2022)
Article
Computer Science, Artificial Intelligence
Jing Yuan, Shuwei Cao, Gangxing Ren, Fengxian Su, Huiming Jiang, Qian Zhao
Summary: In this paper, a new mechanical feature extraction and fault diagnosis method LW-Net is introduced, which achieves accurate extraction of impact fault features and improves fault diagnosis effectiveness through smart lifting wavelet kernels and the design of lifting layer.
NEURAL COMPUTING & APPLICATIONS
(2022)
Article
Computer Science, Artificial Intelligence
Rui Shi, Tianxing Li, Yasushi Yamaguchi
Summary: This paper presents attribution methods for explaining decisions made by convolutional neural networks (CNNs), and proposes an objective function for generating baselines iteratively using gradient descent. The research shows that this method outperforms other approaches in terms of attribution results.
IMAGE AND VISION COMPUTING
(2022)
Article
Engineering, Marine
Wenhua Li, Wenrui Song, Guang Yin, Muk Chen Ong, Fenghui Han
Summary: In this paper, a two-phase flow experimental system is constructed to identify the flow pattern of subsea jumpers using the EfficientNet-B5 convolutional neural network (CNN). The research achieves high accuracy in flow pattern recognition and provides a new method for monitoring flow patterns in deep-sea oil exploration and transportation.
Article
Biodiversity Conservation
Guoli Zhang, Ming Wang, Kai Liu
Summary: This paper compares and analyzes the application of two feedforward neural network models (CNNs and MLPs) in global wildfire susceptibility prediction, and explores the interpretability of the CNNs model. By constructing response variables and monthly wildfire predictors, four MLPs and CNNs architectures were built, and five statistical measures were used to evaluate the prediction performance of the models. The contextual-based CNN-2D model was found to have the highest accuracy, while the MLPs model was more suitable for pixel-based classification, and the performance ranking of the four models was CNN-2D > MLP-1D > MLP-2D > CNN-1D.
ECOLOGICAL INDICATORS
(2021)
Article
Computer Science, Artificial Intelligence
Hui Dou, Furao Shen, Jian Zhao, Xinyu Mu
Summary: Neurons are fundamental units of neural networks, and this paper proposes a method for explaining neural networks by visualizing the learning process of neurons. The method can analyze the working mechanism of different neural network models without requiring any changes to the architectures. The effectiveness of the method is demonstrated through experiments on various neural network architectures for image classification tasks, providing insights into the interpretability of neural networks from diverse perspectives.
Article
Environmental Sciences
Jack D. Hollister, Xiaohao Cai, Tammy Horton, Benjamin W. Price, Karolina M. Zarzyczny, Phillip B. Fenberg
Summary: The shell morphology of limpets can be difficult to identify visually, even for experts, due to its cryptic and variable nature. In this study, we demonstrate that computer vision models can assist with species identification. By analyzing digital images of limpet shells, the models were able to distinguish between different species and genera with slightly better accuracy and much faster speed compared to experts. The use of heatmaps further confirmed the focus on important morphological features for species and genus differentiation. This research highlights the significant potential of computer vision in enhancing species identification techniques and exploring the biology and ecology of limpets in greater detail.
FRONTIERS IN MARINE SCIENCE
(2023)
Article
Computer Science, Artificial Intelligence
Riccardo Caccavale, Mirko Ermini, Eugenio Fedeli, Alberto Finzi, Vincenzo Lippiello, Fabrizio Tavano
Summary: In this study, a multi-robot approach based on distributed deep Q-learning technique is proposed to sanitize railway stations. The approach utilizes anonymous data from existing WiFi networks to estimate crowded areas within the station and develops a prioritized heatmap for sanitation. A team of cleaning robots, each equipped with a robot-specific convolutional neural network, effectively cooperates to sanitize the station's areas according to the priorities. The approach is evaluated in a realistic simulation scenario at Rome Termini, the largest railway station in Italy, by considering different case studies and a real dataset from one-day data recording of the station's WiFi network.
APPLIED INTELLIGENCE
(2023)
Article
Computer Science, Information Systems
Hyungho Jung, Ryong Lee, Sang-Hwan Lee, Wonjun Hwang
Summary: The paper introduces an active weighted mapping method that infers proper weight values on the fly, successfully applied to various backbone architectures. Results show the method's superiority and generality on various datasets compared to the baseline.
MULTIMEDIA TOOLS AND APPLICATIONS
(2021)
Article
Food Science & Technology
Fanqianhui Yu, Tao Lu, Changhu Xue
Summary: In this study, different Convolutional Neural Network (CNN)-based models including series networks (AlexNet and VGG-19) and directed acyclic graph (DAG) networks (ResNet-18, ResNet-50, and ResNet-101) were used with transfer learning to identify and classify 13 classes of apples from 7439 images. The dataset configuration had a significant impact on the classification results, with VGG-19 achieving the highest accuracy. The study also used visualization techniques to improve the interpretability and credibility of the models.
Article
Environmental Sciences
Xiao-Yan Xu, Min Shao, Pu-Long Chen, Qin-Geng Wang
Summary: In this study, deep convolutional neural network models were constructed to predict tropical cyclone intensity, minimum central pressure, and maximum 2 min mean wind speed. Sensitivity experiments were also conducted to explore the interpretability of the model structure, and the results verified the validity and reliability of the model. The study found that the importance of predictors varies in different targets.
Article
Computer Science, Artificial Intelligence
Kaiwen Hu, Jing Gao, Fangyuan Mao, Xinhui Song, Lechao Cheng, Zunlei Feng, Mingli Song
Summary: In this paper, a disassembler framework called Disassembler is proposed to disassemble convolutional segmentation networks into category-aware convolution kernels for customizable tasks without additional training. It utilizes forward channel-wise activation attribution and backward gradient attribution to effectively disassemble the kernels. Extensive experiments show that the Disassembler can accomplish category-customizable tasks without extra training and outperform state-of-the-art methods with one epoch of fine-tuning.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2023)