Search

Cloud computing has recently gained widespread attention owing to its use in applications involving the Internet of Things (IoT). However, the transmission of massive volumes of data to a cloud server often results in overhead. Fog computing has emerged as a viable solution to address this issue. This study implements an Artificial Intelligence of Things (AIoT) system based on fog computing on a smart farm. Three experiments are conducted to evaluate the performance of the AIoT system. First, network traffic volumes between systems employing and not employing fog computing are compared. Second, the performance of the communication protocols-hypertext transport protocol (HTTP), message queuing telemetry transport protocol (MQTT), and constrained application protocol (CoAP)-commonly used in IoT applications is assessed. Finally, a convolutional neural network-based algorithm is introduced to determine the maturity level of coffee tree images. Experimental data are collected over ten days from a coffee tree farm in the Republic of Korea. Notably, the fog computing system demonstrates a 26% reduction in the cumulative data volume compared with a non-fog system. MQTT exhibits stable results in terms of the data volume and loss rate. Additionally, the maturity level determination algorithm performed on coffee fruits provides reliable results.

Deep learning technology is generally applied to analyze periodic data, such as the data of electromyography (EMG) and acoustic signals. Conversely, its accuracy is compromised when applied to the anomalous and irregular nature of the data obtained using a magneto-impedance (MI) sensor. Thus, we propose and analyze a deep learning model based on recurrent neural networks (RNNs) optimized for the MI sensor, such that it can detect and classify data that are relatively irregular and diverse compared to the EMG and acoustic signals. Our proposed method combines the long short-term memory (LSTM) and gated recurrent unit (GRU) models to detect and classify metal objects from signals acquired by an MI sensor. First, we configured various layers used in RNN with a basic model structure and tested the performance of each layer type. In addition, we succeeded in increasing the accuracy by processing the sequence length of the input data and performing additional work in the prediction process. An MI sensor acquires data in a non-contact mode; therefore, the proposed deep learning approach can be applied to drone control, electronic maps, geomagnetic measurement, autonomous driving, and foreign object detection.

Treatment of facial palsy is essential because neglecting this disorder can lead to serious sequelae and further damage. For an objective evaluation and consistent rehabilitation training program of facial palsy patients, a clinician's evaluation must be simultaneously performed alongside quantitative evaluation. Recent research has evaluated facial palsy using 68 facial landmarks as features. However, facial palsy has numerous features, whereas existing studies use relatively few landmarks; moreover, they do not confirm the degree of improvement in the patient. In addition, as the face of a normal person is not perfectly symmetrical, it must be compared with previous images taken at a different time. Therefore, we introduce three methods to numerically approach measuring the degree of facial palsy after extracting 478 3D facial landmarks from 2D RGB images taken at different times. The proposed numerical approach performs registration to compare the same facial palsy patients at different times. We scale landmarks by performing scale matching before global registration. After scale matching, coarse registration is performed with global registration. Point-to-plane ICP is performed using the transformation matrix obtained from global registration as the initial matrix. After registration, the distance symmetry, angular symmetry, and amount of landmark movement are calculated for the left and right sides of the face. The degree of facial palsy at a certain point in time can be approached numerically and can be compared with the degree of palsy at other times. For the same facial expressions, the degree of facial palsy at different times can be measured through distance and angle symmetry. For different facial expressions, the simultaneous degree of facial palsy in the left and right sides can be compared through the amount of landmark movement. Through experiments, the proposed method was tested using the facial palsy patient database at different times. The experiments involved clinicians and confirmed that using the proposed numerical approach can help assess the progression of facial palsy.

Gaze is an excellent indicator and has utility in that it can express interest or intention and the condition of an object. Recent deep-learning methods are mainly appearance-based methods that estimate gaze based on a simple regression from entire face and eye images. However, sometimes, this method does not give satisfactory results for gaze estimations in low-resolution and noisy images obtained in unconstrained real-world settings (e.g., places with severe lighting changes). In this study, we propose a method that estimates gaze by detecting eye region landmarks through a single eye image; and this approach is shown to be competitive with recent appearance-based methods. Our approach acquires rich information by extracting more landmarks and including iris and eye edges, similar to the existing feature-based methods. To acquire strong features even at low resolutions, we used the HRNet backbone network to learn representations of images at various resolutions. Furthermore, we used the self-attention module CBAM to obtain a refined feature map with better spatial information, which enhanced the robustness to noisy inputs, thereby yielding a performance of a 3.18% landmark localization error, a 4% improvement over the existing error and A large number of landmarks were acquired and used as inputs for a lightweight neural network to estimate the gaze. We conducted a within-datasets evaluation on the MPIIGaze, which was obtained in a natural environment and achieved a state-of-the-art performance of 4.32 degrees, a 6% improvement over the existing performance.

Digital pathology analysis using deep learning has been the subject of several studies. As with other medical data, pathological data are not easily obtained. Because deep learning-based image analysis requires large amounts of data, augmentation techniques are used to increase the size of pathological datasets. This study proposes a novel method for synthesizing brain tumor pathology data using a generative model. For image synthesis, we used embedding features extracted from a segmentation module in a general generative model. We also introduce a simple solution for training a segmentation model in an environment in which the masked label of the training dataset is not supplied. As a result of this experiment, the proposed method did not make great progress in quantitative metrics but showed improved results in the confusion rate of more than 70 subjects and the quality of the visual output.

The adoption of low-crested and submerged structures (LCS) reduces the wave behind a structure, depending on the changes in the freeboard, and induces stable waves in the offshore. We aimed to estimate the wave transmission coefficient behind LCS structures to determine the feasible characteristics of wave mitigation. In addition, various empirical formulas based on regression analysis were proposed to quantitatively predict wave attenuation characteristics for field applications. However, inherent variability of wave attenuation causes the limitation of linear statistical approaches, such as linear regression analysis. Herein, to develop an optimization model for the hydrodynamic behavior of the LCS, we performed a comprehensive analysis of 10 types of machine learning models, which were compared and reviewed on the prediction accuracy with existing empirical formulas. We found that, among the 10 models, the gradient boosting model showed the highest prediction accuracy with MSE of 1.0 × 10 -3 , an index of agreement of 0.996, a scatter index of 0.065, and a correlation coefficient of 0.983, which indicates a performance improvement over the existing empirical formulas. In addition, based on a variable importance analysis using explainable artificial intelligence, we determined the significant importance of the input variable for the relative freeboard (R C /H 0 ) and the relative freeboard to water depth ratio (R C /h), which confirms that the relative freeboard was the most dominant factor for influencing wave attenuation in the hydraulic behavior around the LCS. Thus, we concluded that the performance prediction method using a machine learning model can be applied to various predictive studies in the field of coastal engineering, deviating from existing empirical-based research.

Multi-object tracking is a significant field in computer vision since it provides essential information for video surveillance and analysis. Several different deep learning-based approaches have been developed to improve the performance of multi-object tracking by applying the most accurate and efficient combinations of object detection models and appearance embedding extraction models. However, two-stage methods show a low inference speed since the embedding extraction can only be performed at the end of the object detection. To alleviate this problem, single-shot methods, which simultaneously perform object detection and embedding extraction, have been developed and have drastically improved the inference speed. However, there is a trade-off between accuracy and efficiency. Therefore, this study proposes an enhanced single-shot multi-object tracking system that displays improved accuracy while maintaining a high inference speed. With a strong feature extraction and fusion, the object detection of our model achieves an AP score of 69.93% on the UA-DETRAC dataset and outperforms previous state-of-the-art methods, such as FairMOT and JDE. Based on the improved object detection performance, our multi-object tracking system achieves a MOTA score of 68.5% and a PR-MOTA score of 24.5% on the same dataset, also surpassing the previous state-of-the-art trackers.

The efficiency of the metal detection method using deep learning with data obtained from multiple magnetic impedance (MI) sensors was investigated. The MI sensor is a passive sensor that detects metal objects and magnetic field changes. However, when detecting a metal object, the amount of change in the magnetic field caused by the metal is small and unstable with noise. Consequently, there is a limit to the detectable distance. To effectively detect and analyze this distance, a method using deep learning was applied. The detection performances of a convolutional neural network (CNN) and a recurrent neural network (RNN) were compared from the data extracted from a self-impedance sensor. The RNN model showed better performance than the CNN model. However, in the shallow stage, the CNN model was superior compared to the RNN model. The performance of a deep-learning-based (DLB) metal detection network using multiple MI sensors was compared and analyzed. The network was detected using long short-term memory and CNN. The performance was compared according to the number of layers and the size of the metal sheet. The results are expected to contribute to sensor-based DLB detection technology.

In this paper, we propose a detection method for salient objects whose eyes are focused on gaze tracking; this method does not require a device in a single image. A network was constructed using Neg-Region Attention (NRA), which predicts objects with a concentrated line of sight using deep learning techniques. The existing deep learning-based method has an autoencoder structure, which causes feature loss during the encoding process of compressing and extracting features from the image and the decoding process of expanding and restoring. As a result, a feature loss occurs in the area of the object from the detection results, or another area is detected as an object. The proposed method, that is, NRA, can be used for reducing feature loss and emphasizing object areas with encoders. After separating positive and negative regions using the exponential linear unit activation function, converted attention was performed for each region. The attention method provided without using the backbone network emphasized the object area and suppressed the background area. In the experimental results, the proposed method showed higher detection results than the conventional methods.

RGB-D cameras have been commercialized, and many applications using them have been proposed. In this paper, we propose a robust registration method of multiple RGB-D cameras. We use a human body tracking system provided by Azure Kinect SDK to estimate a coarse global registration between cameras. As this coarse global registration has some error, we refine it using feature matching. However, the matched feature pairs include mismatches, hindering good performance. Therefore, we propose a registration refinement procedure that removes these mismatches and uses the global registration. In an experiment, the ratio of inliers among the matched features is greater than 95% for all tested feature matchers. Thus, we experimentally confirm that mismatches can be eliminated via the proposed method even in difficult situations and that a more precise global registration of RGB-D cameras can be obtained.

The use of human gesturing to interact with devices such as computers or smartphones has presented several problems. This form of interaction relies on gesture interaction technology such as Leap Motion from Leap Motion, Inc, which enables humans to use hand gestures to interact with a computer. The technology has excellent hand detection performance, and even allows simple games to be played using gestures. Another example is the contactless use of a smartphone to take a photograph by simply folding and opening the palm. Research on interaction with other devices via hand gestures is in progress. Similarly, studies on the creation of a hologram display from objects that actually exist are also underway. We propose a hand gesture recognition system that can control the Tabletop holographic display based on an actual object. The depth image obtained using the latest Time-of-Flight based depth camera Azure Kinect is used to obtain information about the hand and hand joints by using the deep-learning model CrossInfoNet. Using this information, we developed a real time system that defines and recognizes gestures indicating left, right, up, and down basic rotation, and zoom in, zoom out, and continuous rotation to the left and right.

Vehicle detection is an important research area that provides background information for the diversity of unmanned-aerial-vehicle (UAV) applications. In this paper, we propose a vehicle-detection method using a convolutional-neural-network (CNN)-based object detector. We design our method, DRFBNet300, with a Deeper Receptive Field Block (DRFB) module that enhances the expressiveness of feature maps to detect small objects in the UAV imagery. We also propose the UAV-cars dataset that includes the composition and angular distortion of vehicles in UAV imagery to train our DRFBNet300. Lastly, we propose a Split Image Processing (SIP) method to improve the accuracy of the detection model. Our DRFBNet300 achieves 21 mAP with 45 FPS in the MS COCO metric, which is the highest score compared to other lightweight single-stage methods running in real time. In addition, DRFBNet300, trained on the UAV-cars dataset, obtains the highest AP score at altitudes of 20-50 m. The gap of accuracy improvement by applying the SIP method became larger when the altitude increases. The DRFBNet300 trained on the UAV-cars dataset with SIP method operates at 33 FPS, enabling real-time vehicle detection.