In the design of offshore jacket foundations, fatigue life is crucial. Post-weld treatment has been proposed to enhance the fatigue performance of welded joints, where particularly high-frequency mechanical impact (HFMI) treatment has been shown to improve fatigue performance significantly. Automated HFMI treatment has enhanced the quality assurance of post-weld treatment and can lead to cost-effective design when combined with accurate fatigue life prediction. To accurately predict the fatigue life in complex or multi-axial joints, the finite element method (FEM) is commonly used but relies on a basic CAD depiction of the weld, which fails to consider the actual weld geometry and defects. Including the actual weld geometry in the FE model improves fatigue life prediction and possible crack location prediction but requires a digital reconstruction of the weld. Current digital reconstruction methods are time-consuming or require specialised scanning equipment and potential component relocation. Instead, the proposed framework uses an industrial manipulator paired with a line scanner to integrate digital reconstruction into the automated HFMI treatment setup. This approach applies prevalent image processing, simple filtering techniques, and non-linear optimisation for aligning and merging overlapping scans. A screened Poisson surface reconstruction finalises the 3D model to create a meshed surface. The outcome is a generic, cost-effective, flexible, and rapid method that enables the digital reconstruction of welded parts, aiding in component design, overall quality assurance, and documentation of the HFMI treatment.
In many real-world scenarios, obtaining large amounts of labeled data can be a daunting task. Weakly supervised learning techniques have gained significant attention in recent years as an alternative to traditional supervised learning, as they enable training models using only a limited amount of labeled data. In this paper, the performance of a weakly supervised classifier to its fully supervised counterpart is compared on the task of defect detection. Experiments are conducted on a dataset of images containing defects, and evaluate the two classifiers based on their accuracy, precision, and recall. Our results show that the weakly supervised classifier achieves comparable performance to the supervised classifier, while requiring significantly less labeled data.
The explosion of logistics is consolidating a positive impact on modern society by offering a critical contribution to its economic development. Traditional inventory methods tend to be transformed based on the rapid development of artificial intelligence technology. In particular, changing the way inventories are recorded, through the evolution of machine learning methods in contemporary logistics, improves industrial efficiency, reduces costs and promotes the long-term development of businesses. The introduction of systems banking on artificial intelligence allows for smarter and more efficient methods than the traditional warehouse inventory counting. This paper presents a new approach (content and architecture) being included in semantic logistics technologies for specific warehouses or distribution centres that stores a wide range of general or specialized machines. Warehouse staff should utilise a smart device to grab images of the stored items on the pallet racks. The images are automatically processed and associated with the database such that the inventory items are recognized and counted, resulting in achieving the goals of proper inventory counting and evaluating stored entries' existence. The proposed inventory counting system is expected to enhance physical inventory recording, thus contributing to the efficiency of the entire supply chain and collaborating with other components of Warehouse 4.0.
Automatic detection of traffic accidents is an important emerging topic in traffic monitoring systems. Nowadays many urban intersections are equipped with surveillance cameras connected to traffic management systems. Therefore, computer vision techniques can be viable tools for automatic accident detection. This paper presents a new efficient framework for accident detection at intersections for traffic surveillance applications. The proposed framework consists of three hierarchical steps, including efficient and accurate object detection based on the state-of-the-art YOLOv4 method, object tracking based on Kalman filter coupled with the Hungarian algorithm for association, and accident detection by trajectory conflict analysis. A new cost function is applied for object association to accommodate for occlusion, overlapping objects, and shape changes in the object tracking step. The object trajectories are analyzed in terms of velocity, angle, and distance in order to detect different types of trajectory conflicts including vehicle-to-vehicle, vehicle-to-pedestrian, and vehicle-to-bicycle. Experimental results using real traffic video data show the feasibility of the proposed method in real-time applications of traffic surveillance. In particular, trajec-tory conflicts, including near-accidents and accidents occurring at urban intersections are detected with a low false alarm rate and a high detection rate. The robustness of the proposed framework is evaluated using video sequences collected from YouTube with diverse illumination conditions. The dataset is publicly available at: http://github.com/hadi-ghnd/AccidentDetection.
Dynamic imaging of flowing cells, especially for flow cytometry, is of great significance for the research in medicine and biology. Compared to traditional optical imaging methods, electrical impedance tomography (EIT) has the advantages of label-free and low cost. It can determine not only the shape and size of cells but also the distribution of its electric properties that cannot be obtained by other methods, presenting a significant advance in the traditional impedance-based detection method. However, the scale of traditional EIT sensor limits single-cell EIT imaging, which is usually around or larger than 1 mm, but the size of most cells is around 10 µm. Therefore, a micro scale sensor for dynamic imaging is designed and fabricated on Printed Circuit Board (PCB), with eight equally-spaced electrodes surrounding a microchannel of 150 µm in diameter. To verify the performance of the micro scale EIT sensor, static sensors with a similar structure were designed and used to reconstruct the grits located in the imaging area, with the shape and position of the grit well reconstructed. Thereafter in experiment, the dynamic sensor was used to image the flowing SMG-C6 cells, and the reconstructed image is similar to the microscopic observations, providing the basis for future EIT -based cell identification, classification, and enumeration.
Residents in some major US cities can report their drinking water quality complaints using 311, a local government-sponsored call line and online portal. The objectives of this research are therefore to: (1) perform spatiotemporal analysis of observed water quality complaints and identify hotspots; (2) develop deep learning-based time series models to forecast water quality complaints data, and (3) compare the performance of the models to subsequently develop performance improvements. A classic time series modeling, Autoregressive Integrated Moving Average (ARIMA) along with Machine Learning (ML) techniques such as Long Short-Term Memory (LSTM) network and Support Vector Machine (SVM) were used for the time series analysis. This study will provide an in-depth insight into the effective management of drinking water quality complaints, leading to sustained strategies for enhanced drinking water infrastructure management and operations in large urban cities.
Deep Neural Networks (DNNs) have been effective in providing solutions to most prevailing problems in the field Computer Vision. Modern Object Detector architecture families such as Single Shot Detector, You Only Look Once (YOLO), Region Based Convolutional Neural Networks (RCNN) and others are powerful models with strong capabilities of distinguishing many common object classes in images with unprecedented speed and accuracy. However, in most of their applications the use of deep networks conceals the inference of object class and location from the input data by introducing many layers of complex black-box functionality. This leads to sometimes bizarre and inexplicable detection outputs. A potential use of these networks in risk-prone applications motivates the development of techniques for visually interpreting their outputs, which could lead to more controlled training and deployment. Layer-wise Relevance Propagation (LRP) is a popular "eXplainable AI" (XAI) technique for constructing such explanations, frequently used in investigating the output of DNNs. For object detection, LRP could provide us with a useful heatmap of the input's relevance to the output class, localized within an estimate of a bounding box. In this work, we propose a new and efficient method for conducting relevance propagation in object detector DNNs that leads to visually informative results. In our experiments, we make use of the YOLOv5 model showing meaningful visual explanations of its output. Our implementation philosophy exemplifies how relevance propagation can be regulated for complex modular architectures.
Aiming at precise sub-category classification of images, fine-grained image recognition requires the algorithms to enjoy a remarkable ability of subtle feature extraction. Recently, the architecture of Transformer has been successfully applied in vision tasks, bringing a novel approach to improve feature extraction performance of fine-grained image recognition algorithms. However, fine-grained image datasets are usually quite limited in capacity, which are unfavorable for the data-consuming training process of Transformers. In order to increase the available amount of data for training, in this paper we firstly introduce a stochastic image data augmentation method for Vision Transformer (ViT), which uses a Dense-DETR model to extract feature regions and performs random insertion and removal for the transformed patch sequence. To select the most informative sequence elements in the forward propagation pro-cess, we implement a feature patch selection strategy by applying an additional convolutional network structure to ViT encoders. Inspired from active learning, a contrastive loss utilizing the posterior information of paired images is also introduced as a penalty item of ViT's cross-entropy loss objective. Such strategies can make the ViT extract the most discriminative feature information from its input. Extensive experiments have supported that the proposed sequence-selective Vision Transformer reaches the highest recognition accuracies on several frequently-used fine-grained image datasets.
This paper proposes a machine vision based sorghum grain adulteration identification method for Baijiu brewing applications. A case study with respect to glutinous sorghum mixed with japonica sorghum is conducted. The designed sampling box with vibrating table is used to capture images with mobile phone. Image preprocessing, including brightness equalization, binarization, morphological operations and connected component calculation, are used to segment different sorghum particles and thus extract particle size, shape and color features including area, major and minor axis length, eccentricity, roundness, shape factor, aspect ratio, RGB, HSV and CIELAB parameters. With principal component analysis (PCA) for data dimensionality reduction, the support vector machine (SVM) is used to classify sorghum varieties and discriminate whether sorghum grains are adulterated. The preliminary experimental results show that the proposed machine vision based method and SVM model with radial basis function (RBF) kernel achieved an accuracy of 84.71% for the classification of glutinous sorghum and japonica sorghum. Moreover, the model identifies and labels impurity grains precisely on the constructed manually constructed samples for adulterated sorghums, which demonstrates the promising for the nondestructive and low-cost adulteration identification of sorghum raw material in Baijiu brewing process and may promote the digital transformation of the industry and improve production efficiency
Mitochondria are the organelles that generate energy for the cells. Many studies have suggested that mitochondrial dysfunction or impairment may be related to cancer and other neurodegenerative disorders such as Alzheimer's and Parkinson's diseases. Therefore, morphologically detailed alterations in mitochondria and 3D reconstruction of mitochondria are highly demanded research problems in the performance of clinical diagnosis. Nevertheless, manual mitochondria segmentation over 3D electron microscopy volumes is not a trivial task. This study proposes a two-stage cascaded CNN architecture to achieve automated 3D mitochondria segmentation, combining the merits of top-down and bottom-up approaches. For top-down approaches, the segmentation is conducted on objects' localization so that the delineations of objects' contours can be more precise. However, the combinations of 2D segmentation from the top-down approaches are inadequate to perform proper 3D segmentation without the information on connectivity among frames. On the other hand, the bottom-up approach finds coherent groups of pixels and takes the information of 3D connectivity into account in segmentation to avoid the drawbacks of the 2D top-down approach. However, many small areas that share similar pixel properties with mitochondria become false positives due to insufficient information on objects' localization. In the proposed method, the detection of mitochondria is carried out with multi-slice fusion in the first stage, forming the segmentation cues. Subsequently, the second stage is to perform 3D CNN segmentation that learns the pixel properties and the information of 3D connectivity under the supervision of cues from the detection stage. Experimental results show that the proposed structure alleviates the problems in both the top-down and bottom-up approaches, which significantly accomplishes better performance in segmentation and expedites clinical analysis.
How younger students, such as primary and secondary school students, can improve their concentration when studying alone has been the subject of research by education experts. In this paper, we develop a learning-state-based tomato-clock system that can help improve concentration. We propose a joint model to detect students' learning states and thus control whether the tasks of the tomato-clock can be completed properly, which in turn motivates students to focus on their learning. In the joint model, the temporal action detection model SlowFast detects the video and identifies the base action category and the state action category. In cases where the actions are similar, such as the states of reading and watching video on mobile phone, we calculate the student's head pose information to determine his or her field of view and use the FPN-Faster RCNN model to detect the key items within his or her field of view to detect the real action. Finally, their learning state was identified based on the duration of the action, with mAP of 85.28%.
Our product seeks to provide a teacher-driven computer programming education platform that allows users total anonymity in communication and grading. The purpose of this software is to provide educators the ability to assign students both in class programming contests that are graded on a time-to-completion basis and to facilitate both guided and collaborative communication about programming and computer software. This product was initially designed to be used in university's Intro to Programming classes where the professor recognized that students, especially females, were hesitant to participate due to a perceived lack of knowledge of the topic. In any situation, asking questions can be beneficial, and this platform will provide students the ability to ask their peers and professors questions without the fear of negative reflection on their knowledge or understanding. The platform is built on a custom serverless architecture utilizing Amazon Web Services (AWS). The platform hosts a publicly accessible web portal, API layers for integration and data manipulation, and database and object storage solutions for data management and storage. Our choice in using AWS gave us the ability to implement pre-built and managed security solutions for our project. During the course of the project, we enhanced our time management skills and learnt how to collaborate and communicate within a team. Ultimately the research project will be considered a success if the application promotes better communication and learning within the classroom.
The Monnig Meteorite Gallery provides educational and outreach experiences to more than 10,000 visitors and students each year. Within the gallery, lighting and design issues can cause complications for visually impaired visitors. The Monnig Meteorite Gallery aims to provide accommodations for visually impaired and non-English speaking visitors by providing a customizable experience through a tablet application, provided to visitors to the gallery. The application uses an array of features, including Bluetooth Beacons, Language Translation, Text-to-Speech, and UI customization to provide the contents of the gallery in an accessible manner to every visitor.
As the aged population is rapidly increased, the need for efficient and low-cost ambient systems becomes vital. The effectiveness of such systems lies upon the accurate and fast motion analysis in order to predict the elderly's action and develop systems to act in need. To achieve that, the precise estimation of the entire human body pose is often exploited, providing the required motion-related information. Yet, the exploitation of the entire human pose can present several limitations. The paper at hand exploits state-of-the-art data-driven classifiers and compares their efficiency in action recognition based on a specific set of joints or coordinates, i.e., the x, y and z-axis. The above rests upon the notion that each action in real life can be effectively perceived by observing only a specific set of joints. Considering that, we aim to investigate the capacity of such a joint analysis and its ability to deliver an enhanced pose-based action recognition system. To that end, we correlate specific joints with each action, indicating the joints that contribute the most. We evaluate our findings on two different senior subjects using two different classifiers, viz., support vector machine (SVM) and convolutional neural network (CNN), showing that the above strategy can improve recognition rates.
Ensuring smart safety of workers in large construction sites calls for centralized ecosystems, supported transversely by security, privacy, and trust enablers, to facilitate data sharing and protect the growing attack surface. At the same time, human-centricity has to be a core part of the process. Mixed Reality (MR) applications allow the execution of context-aware applications with advanced interaction interfaces improving decision making, data gathering, interoperability with other services, accessibility, and real-time data sharing. We leverage the advanced features of MR interfaces and interoperability with Internet of Things (IoT) systems to contribute towards a more usable, functional, and perceptive human-centric Next Generation ecosystem in large construction plants aiming to contribute to workers' health and safety (H&S) during inspection processes. In this study, an MR enabler is designed and developed to collect data from various sources of such an ecosystem. The main objectives are to identify workers in large construction sites and their medical and training records, get real-time information regarding their stress and health levels, and alert when dangerous activity is being performed, or a worker reaches an unauthorized location. At the same time, the MR enabler allows the visualization of BIM models, their digital representations and data, as well as the submission of reports. The MR enabler, connected to other components within the Next-Generation IoT ecosystem (NGIoT), will be applied to various construction sites in Poland.
Lung is a major part of our body. Any disease in the lung can affect the whole respiratory system. Lung cancer is considered one of the most dangerous types of cancer where the death rate is relatively more than other types of cancer. It can be diagnosed in an appropriate manner only if lung nodules are classified at an early stage of its growth. A Computer-Aided Diagnosis (CAD) system can improve the accuracy of properly classifying the cancerous cells which will surely increase the survival rate of patients. This research study has proposed a transfer learning approach for the classification of lung tumors into four classes, which are Small-cell-carcinoma, Adenocarcinoma, Squamous-cell-carcinoma, and Large cell-carcinoma. A large-scale CT and PET/CT DICOM images containing datasets have been used for this research work. This research article has proposed a novel approach that includes a combination of four different advance deep Convolutional Neural Networks: VGG-16, VGG-19, Inception-V3, and Mobilenet-V2 with the K- Nearest Neighbor (KNN) algorithm for the classification of lung cancer into four classes. This research experiment consists of four stages namely image preprocessing, feature generation, feature extraction, and classification. VGG-16 with the KNN algorithm has given the highest validation accuracy as compared to other models for the classification of lung nodules. The proposed model has been evaluated by using a total of 4105 medical images belonging to four categories.
Mapping out a space or area is widely used in the construction and interior design industry. Technological advances in auto-mapping such as drones, lasers, and robots allow for an analysis of the work area at hand. In a construction site, this can be used to inspect the conditions of the area and find its size, measurements, safety, and security. In interior designing, mapping out an area is more limited due to the availability of tools because of its high costs. Because of this, measurements are usually done by hand in order to find the area of a room. In order to improve efficiency and provide more cost-effective options in auto-mapping tools, we have designed an RC auto-mapper, i.e., a remote controlled robotic vehicle for autonomous area exploration and mapping in this senior capstone design project. The RC car will be able to autonomously explore and measure the area of a space and output that data onto a website using Bluetooth. We have created a user-friendly interface that is compatible with any device using React. Our RC car will be utilizing ultrasonic sensors in order to measure the distance to an object using ultrasonic sensors sound waves. Because our product is a more cost effective yet accurate auto mapping tool, interior design firms and smaller construction site projects will have more resources available to complete their tasks at hand in safe conditions.
Magnetic Resonance Imaging (MRI) plays a major role in the diagnosis of several diseases. However, the acquisition of measurements in the k-space domain, which is the basis for image reconstruction, takes a long time compared to other imaging modalities and is comparatively costly. In this context, undersampled MRI reconstruction is an approach for decreasing the acquisition duration and the exam's final cost. With this objective, both Compressed Sensing (CS) and Deep Learning (DL) provide techniques for generating good quality MRI images from undersampled measurements. In this paper, we combine CS and DL methods in order to investigate the potential increase in image quality over each isolated approach. We use reconstructions from from highly undersampled MRI signals using two CS approaches, the L1 and total minimizations, as inputs to a U-Net. We also use, for comparison, the reconstructions from the same undersampled signals using L2 minimization, and also test them as inputs to a U-net. The goal is to evaluate whether the the U-Net can improve the results of the CS reconstructions after learning from degraded and original image pairs. Our experimental results suggest that the combination of L1 or TV minimization with U-Nets can improve reconstruction, in terms of objective image quality, over each technique used in isolation.
Foreground detection is an important first step in video analytics. While the stationary cameras facilitate the foreground detection due to the apparent motion between the moving foreground and the still background, the moving cameras make such a task more challenging because both the foreground and the background appear in motion in the video. To tackle this challenging problem, an innovative real-time foreground detection method is presented, that models the foreground and the background simultaneously and works for both moving and stationary cameras. In particular, first, each input video frame is partitioned into a number of blocks. Then, assuming the background takes the majority of each video frame, the iterative pyramidal implementation of the Lucas-Kanade optical flow approach is applied on the centers of the background blocks in order to estimate the global motion and compensate for the camera movements. Subsequently, each block in the background is modeled by a mixture of Gaussian distributions and a separate Gaussian mixture model is constructed for the foreground in order to enhance the classification. However, the errors in motion compensation can contaminate the foreground model with background values. The novel idea of the proposed method matches a set of background samples to their corresponding block for the most recent frames in order to avoid contaminating the foreground model with background samples. The input values that do not fit into either the statistical or the sample-based background models are used to update the foreground model. Finally, the foreground is detected by applying the Bayes classification technique to the major components in the background and foreground models, which removes the false positives caused by the hysteresis effect. Experimental evaluations demonstrate the feasibility of the proposed method in the foreground segmentation when applied to videos in public datasets.
This paper proposes a virtual lane prediction algorithm based on inverse perspective transformation and machine learning for lane departure warning in low-power embedded systems. The virtual lane prediction method helps in more accurate lane detection by predicting the opposite lane when only one lane is detected or checking whether the distance between lanes is correct when both lanes are detected. The inverse perspective transformation is used for obtaining a bird's-eye view of the scene from a perspective image to remove perspective effects for lane detection and virtual lane prediction. This method requires only the internal and external parameters of the camera without a homography matrix with 8 degrees of freedom (DoF) that maps the points in one image to the corresponding points in the other image. To improve the accuracy and speed of lane detection in complex road environments, we use a machine learning algorithm to accurately detect lanes in the region that passed the first classifier that roughly detects lanes. The system has been tested through the driving video of the vehicle in embedded system. The experimental results show that the proposed virtual lane prediction method works well in various road environments and meet the real-time requirements for low-power embedded systems.
