Bottom-up image detection of water channel slope damages based on superpixel segmentation and support vector machine

The operation of water supply channels is threatened by the occasionally occurred slope damages. Timely detection of their occurrence is critical for the rapid enforcement of mitigation measures. However, current practices based on routine inspection and structural heath monitoring are inefficient, laborious and tend to be biased. As an attempt to address the limitations, this paper proposes a bottom-up image detection approach for slope damages, which includes four steps, i.e. superpixel segmentation, feature handcrafting, superpixel classification based on support vector machine (SVM), and slope damage recognition. The approach employs a bottom-up strategy to infer the upper-level slope condition from the classification results of individual superpixels in the bottom level. Experiments were conducted to demonstrate the effectiveness of the approach. The handcrafted feature “LBP + HSV” was demonstrated to be effective in characterizing the image features of slope damages. An SVM model with “LBP + HSV” as input can reliably identify the slope condition in superpixels. Based on the SVM model, the bottom-up strategy achieved high recognition performance, of which the overall accuracy can be up to 91.7%. The proposed approach has potential to facilitate the early and comprehensive awareness of slope damages along the entire route of water channel by the integration with unmanned aerial vehicles.     

面向航拍图像中工程车辆检测与识别的改进胶囊网络

目的 利用无人机(unmanned aerial vehicle, UAV)巡检识别航拍图像中的工程车辆对于减少电力安全事故的发生具有重要意义。采用人工提取特征的经典模式识别方法或YOLOv5(you only look once v5)等深度学习算法识别无人机电力巡检航拍图像中的工程车辆,存在识别准确率低、模型参数规模大等问题。针对上述问题,提出一种改进的胶囊网络识别航拍图像中的工程车辆。方法 采用多层密集连接型方法改进原始胶囊网络结构,以提取图像中工程车辆更多的特征;改进了胶囊网络的动态路由方法,以提高胶囊网络的抗干扰能力;探索了网络层数和动态路由算法中关键参数对识别准确率的影响,以找到识别准确率最高时的参数。结果 实验结果表明：1)在所采用的算法模型中,本文方法的平均识别率(mean average precision, mAP)达到94.56%,明显高于其他方法。2)网络层数对识别准确率有很大影响,但二者之间并非单调线性关系。在本文的应用场景中,5层胶囊网络的识别准确率最高;此外,动态路由算法改进与否并不会影响识别准确率跟随网络层数的变化趋势。3)胶囊网络层数增加会降低识别效率...     

The nested k‐means method: A new approach for detecting lost persons in aerial images acquired by unmanned aerial vehicles

A new method, named as the nested k‐means, for detecting a person captured in aerial images acquired by an unmanned aerial vehicle (UAV), is presented. The nested k‐means method is used in a newly built system that supports search and rescue (SAR) activities through processing of aerial photographs taken in visible light spectra (red‐green‐blue channels, RGB). First, the k‐means classification is utilized to identify clusters of colors in a three‐dimensional space (RGB). Second, the k‐means method is used to verify if the automatically selected class of colors is concurrently spatially clustered in a two‐dimensional space (easting‐northing, EN), and has human‐size area. The UAV images were acquired during the field campaign carried out in the Izerskie Mountains (SW Poland). The experiment aimed to observe several persons using an RGB camera, in spring and winter, during various periods of day, in uncovered terrain and sparse forest. It was found that the nested k‐means method has a considerable potential for detecting a person lost in the wilderness and allows to reduce area to be searched to 4.4 and 7.3% in spring and winter, respectively. In winter, land cover influences the performance of the nested k‐means method, with better skills in sparse forest than in the uncovered terrain. In spring, such a relationship does not hold. The nested k‐means method may provide the SAR teams with a tool for near real‐time detection of a person and, as a consequence, to reduce search area to approximately 0.5–7.3% of total terrain to be visited, depending on season and land cover.     

Robust fault and icing diagnosis in unmanned aerial vehicles using LPV interval observers

This paper proposes a linear parameter varying (LPV) interval unknown input observer for the robust fault diagnosis of actuator faults and ice accretion in unmanned aerial vehicles (UAVs) described by an uncertain model. The proposed interval observer evaluates the set of values for the state, which are compatible with the nominal fault‐free and icing‐free operation and can be designed in such a way that some information about the nature of the unknown inputs affecting the system can be obtained, thus allowing the diagnosis to be performed. The proposed strategy has several advantages. First, the LPV paradigm allows taking into account operating point variations. Second, the noise rejection properties are enhanced by the presence of the integral term. Third, the interval estimation property guarantees the absence of false alarms. Linear matrix inequality–based conditions for the analysis/design of these observers are provided in order to guarantee the interval estimation of the state and the boundedness of the estimation. The developed theory is supported by simulation results, obtained with the uncertain model of a Zagi Flying Wing UAV, which illustrate the strong appeal of the methodology for identifying correctly unexpected changes in the system dynamics due to actuator faults or icing.     

Convolutional neural networks for object detection in aerial imagery for disaster response and recovery

Accurate and timely access to data describing disaster impact and extent of damage is key to successful disaster management (a process that includes prevention, mitigation, preparedness, response, and recovery). Airborne data acquisition using helicopter and unmanned aerial vehicle (UAV) helps obtain a bird’s-eye view of disaster-affected areas. However, a major challenge to this approach is robustly processing a large amount of data to identify and map objects of interest on the ground in real-time. The current process is resource-intensive (must be carried out manually) and requires offline computing (through post-processing of aerial videos). This research introduces and evaluates a series of convolutional neural network (CNN) models for ground object detection from aerial views of disaster’s aftermath. These models are capable of recognizing critical ground assets including building roofs (both damaged and undamaged), vehicles, vegetation, debris, and flooded areas. The CNN models are trained on an in-house aerial video dataset (named Volan2018) that is created using web mining techniques. Volan2018 contains eight annotated aerial videos (65,580 frames) collected by drone or helicopter from eight different locations in various hurricanes that struck the United States in 2017–2018. Eight CNN models based on You-Only-Look-Once (YOLO) algorithm are trained by transfer learning, i.e., pre-trained on the COCO/VOC dataset and re-trained on Volan2018 dataset, and achieve 80.69% mAP for high altitude (helicopter footage) and 74.48% for low altitude (drone footage), respectively. This paper also presents a thorough investigation of the effect of camera altitude, data balance, and pre-trained weights on model performance, and finds that models trained and tested on videos taken from similar altitude outperform those trained and tested on videos taken from different altitudes. Moreover, the CNN model pre-trained on the VOC dataset and re-trained on balanced drone video yields the best result in significantly shorter training time.     

Mining network data for intrusion detection through combining SVMs with ant colony networks

In this paper, we introduce a new machine-learning-based data classification algorithm that is applied to network intrusion detection. The basic task is to classify network activities (in the network log as connection records) as normal or abnormal while minimizing misclassification. Although different classification models have been developed for network intrusion detection, each of them has its strengths and weaknesses, including the most commonly applied Support Vector Machine (SVM) method and the Clustering based on Self-Organized Ant Colony Network (CSOACN). Our new approach combines the SVM method with CSOACNs to take the advantages of both while avoiding their weaknesses. Our algorithm is implemented and evaluated using a standard benchmark KDD99 data set. Experiments show that CSVAC (Combining Support Vectors with Ant Colony) outperforms SVM alone or CSOACN alone in terms of both classification rate and run-time efficiency.     

Robust text detection in natural scenes using text geometry and visual appearance

This paper proposes a new two-phase approach to robust text detection by integrating the visual appearance and the geometric reasoning rules. In the first phase, geometric rules are used to achieve a higher recall rate. Specifically, a robust stroke width transform (RSWT) feature is proposed to better recover the stroke width by additionally considering the cross of two strokes and the continuousness of the letter border. In the second phase, a classification scheme based on visual appearance features is used to reject the false alarms while keeping the recall rate. To learn a better classifier from multiple visual appearance features, a novel classification method called double soft multiple kernel learning (DS-MKL) is proposed. DS-MKL is motivated by a novel kernel margin perspective for multiple kernel learning and can effectively suppress the influence of noisy base kernels. Comprehensive experiments on the benchmark ICDAR2005 competition dataset demonstrate the effectiveness of the proposed two-phase text detection approach over the state-of-the-art approaches by a performance gain up to 4.4% in terms of F-measure.     

航空遥感图像深度学习目标检测技术研究进展

航空遥感图像目标检测旨在定位和识别遥感图像中感兴趣的目标,是航空遥感图像智能解译的关键技术,在情报侦察、灾害救援和资源勘探等领域具有重要应用价值。然而由于航空遥感图像具有尺寸大、目标小且密集、目标呈任意角度分布、目标易被遮挡、目标类别不均衡以及背景复杂等诸多特点,航空遥感图像目标检测目前仍然是极具挑战的任务。基于深度卷积神经网络的航空遥感图像目标检测方法因具有精度高、处理速度快等优点,受到了越来越多的关注。为推进基于深度学习的航空遥感图像目标检测技术的发展,本文对当前主流遥感图像目标检测方法,特别是2020—2022年提出的检测方法,进行了系统梳理和总结。首先梳理了基于深度学习目标检测方法的研究发展演化过程,然后对基于卷积神经网络和基于Transformer目标检测方法中的代表性算法进行分析总结,再后针对不同遥感图象应用场景的改进方法思路进行归纳,分析了典型算法的思路和特点,介绍了现有的公开航空遥感图像目标检测数据集,给出了典型算法的实验比较结果,最后给出现阶段航空遥感图像目标检测研究中所存在的问题,并对未来研究及发展趋势进行了展望。     

Early fault detection in gearboxes based on support vector machines and multilayer perceptron with a continuous wavelet transform

One of the research problems investigated these days is early fault detection. To this end, advanced signal processing algorithms are employed. The present paper makes an attempt at early fault detection in a gearbox. In order to evaluate its technical condition, artificial neural networks were used. Early fault detection based on support vector machines is a relatively new and rarely employed method for evaluating condition of machines, particularly gearboxes. The available literature offers very promising results of using this method. In order to compare the obtained results, a multilayer perceptron network was created. Such standard neural network ensures high effectiveness. The vibration signal obtained from a sensor is seldom a material for direct analysis. First, it needs to be processed to bring out the informative part of the signal. To this end, a wavelet transform was used. The presented results concern both a “raw” vibration signal and processed one, investigated for two neural networks. The wavelet transform has proved to improve significantly the accuracy of condition evaluation and the results obtained by the two networks are consistent with one another.     

A statistically based fault detection and diagnosis approach for non-residential building water distribution systems

Large non-residential buildings can contain complex and often inefficient water distribution systems. As requirements for water increase due to water scarcity and industrialization, it has become increasingly important to effectively detect and diagnose faults in water distribution systems in large buildings. In many cases, if water supply is not impacted, faults in water distribution systems can go unnoticed. This can lead to unnecessary increases in water usage and associated energy due to pumping, treating, and heating water. The majority of fault detection and diagnosis studies in the water sector are limited to municipal water supply and leakage detection. The application of detection and diagnosis for faults in building water networks remains largely unexplored and the ability to identify and distinguish between routine and non-routine water usage at this scale remains a challenge. This study using case-study data, presents the application of principal component analysis and a multi-class support vector machine to detect and classify faults for non-residential building water networks. In the absence of a process model (which is typical for such water distribution systems), principal component analysis is proposed as a data-driven fault detection technique for building water distribution systems for the first time herein. Hotelling T²-statistics and Q-statistics were employed to detect abnormality within incoming data, and a multi-class support vector machine was trained for fault classification. Despite the relatively limited training data available from the case-study (which would reflect the situation in many buildings), meaningful faults were detected, and the technique proved successful in discriminating between various types of faults in the water distribution system. The effectiveness of the proposed approach is compared to a univariate threshold technique by comparison of their respective performance in the detection of faults that occurred in the case-study site. The results demonstrate the promising capabilities of the proposed fault detection and diagnosis approach. Such a strategy could provide a robust methodology that can be applied to buildings to reduce inefficient water use, reducing their life-cycle carbon footprint.     

A prototype classification method and its use in a hybrid solution for multiclass pattern recognition

In this paper, we propose a prototype classification method that employs a learning process to determine both the number and the location of prototypes. This learning process decides whether to stop adding prototypes according to a certain termination condition, and also adjusts the location of prototypes using either the K-means (KM) or the fuzzy c-means (FCM) clustering algorithms. When the prototype classification method is applied, the support vector machine (SVM) method can be used to post-process the top-rank candidates obtained during the prototype learning or matching process. We apply this hybrid solution to handwriting recognition and address the convergence behavior and runtime consumption of the prototype construction process, and discuss how to combine our prototype classifier with SVM classifiers to form an effective hybrid classifier.