Search results

The purpose of this study is to propose a pulse width based, in-pixel, arbitrary size kernel convolution processor. When image sensors are used in machine vision tasks, large amount of data need to be transferred to the output and fed to a processor. Basic and low-level image processing functions such as kernel convolution is used extensively in the early stages of most machine vision tasks. These low-level functions are usually computationally extensive and if the computation is performed inside every pixel, the burden on the external processor will be greatly reduced.

In the proposed architecture, digital pulse width processing is used to perform kernel convolution on the image sensor data. With this approach, while the photocurrent fluctuations are expressed with changes in the pulse width of an output signal, the small processor incorporated in each pixel receives the output signal of the corresponding pixel and its neighbors and produces a binary coded output result for that specific pixel. The process is commenced in parallel among all pixels of the image sensor.

It is shown that using the proposed architecture, not only kernel convolution can be performed in the digital domain inside smart image sensors but also arbitrary kernel coefficients are obtainable simply by adjusting the sampling frequency at different phases of the processing.

Although in-pixel digital kernel convolution has been previously reported however with the presented approach no in-pixel analog to binary coded digital converter is required. Furthermore, arbitrary kernel coefficients and scaling can be deployed in the processing. The given architecture is a suitable choice for smart image sensors which are to be used in high-speed machine vision tasks.

The purpose of this paper is to design a kernel convolution processor. High-speed image processing is a challenging task for real-time applications such as product quality control of manufacturing lines. Smart image sensors use an array of in-pixel processors to facilitate high-speed real-time image processing. These sensors are usually used to perform the initial low-level bulk image filtering and enhancement.

In this paper, using pulse-width modulated signals and regular nearest neighbor interconnections, a convolution image processor is presented. The presented processor is not only capable of processing arbitrary size kernels but also the kernel coefficients can be any arbitrary positive or negative floating number.

The performance of the proposed architecture is evaluated on a Xilinx Virtex-7 field programmable gate array platform. The peak signal-to-noise ratio metric is used to measure the computation error for different images, filters and illuminations. Finally, the power consumption of the circuit in different operating conditions is presented.

The presented processor array can be used for high-speed kernel convolution image processing tasks including arbitrary size edge detection and sharpening functions, which require negative and fractional kernel values.

This paper primarily aims to focus on a review of convolutional neural network (CNN)-based eye control systems. The performance of CNNs in big data has led to the development of eye control systems. Therefore, a review of eye control systems based on CNNs is helpful for future research.

In this paper, first, it covers the fundamentals of the eye control system as well as the fundamentals of CNNs. Second, the standard CNN model and the target detection model are summarized. The eye control system’s CNN gaze estimation approach and model are next described and summarized. Finally, the progress of the gaze estimation of the eye control system is discussed and anticipated.

The eye control system accomplishes the control effect using gaze estimation technology, which focuses on the features and information of the eyeball, eye movement and gaze, among other things. The traditional eye control system adopts pupil monitoring, pupil positioning, Hough algorithm and other methods. This study will focus on a CNN-based eye control system. First of all, the authors present the CNN model, which is effective in image identification, target detection and tracking. Furthermore, the CNN-based eye control system is separated into three categories: semantic information, monocular/binocular and full-face. Finally, three challenges linked to the development of an eye control system based on a CNN are discussed, along with possible solutions.

This research can provide theoretical and engineering basis for the eye control system platform. In addition, it also summarizes the ideas of predecessors to support the development of future research.

In the field of steel defect detection, the existing detection algorithms struggle to achieve a satisfactory balance between detection accuracy, computational cost and inference speed due to the interference from complex background information, the variety of defect types and significant variations in defect morphology. To solve this problem, this paper aims to propose an efficient detector based on multi-scale information extraction (MSI-YOLO), which uses YOLOv8s as the baseline model.

First, the authors introduce an efficient multi-scale convolution with different-sized convolution kernels, which enables the feature extraction network to accommodate significant variations in defect morphology. Furthermore, the authors introduce the channel prior convolutional attention mechanism, which allows the network to focus on defect areas and ignore complex background interference. Considering the lightweight design and accuracy improvement, the authors introduce a more lightweight feature fusion network (Slim-neck) to improve the fusion effect of feature maps.

MSI-YOLO achieves 79.9% mean average precision on the public data set Northeastern University (NEU)-DET, with a model size of only 19.0 MB and an frames per second of 62.5. Compared with other state-of-the-art detectors, MSI-YOLO greatly improves the recognition accuracy and has significant advantages in computational cost and inference speed. Additionally, the strong generalization ability of MSI-YOLO is verified on the collected industrial site steel data set.

This paper proposes an efficient steel defect detector with high accuracy, low computational cost, excellent detection speed and strong generalization ability, which is more valuable for practical applications in resource-limited industrial production.

The purpose of this paper is to describe the implementation of discrete singular convolution (DSC) method to steady seepage flow while presenting one of the possible uses of DSC method in geotechnical engineering. It also aims to present the implementation of DSC to the problems with mixed boundary conditions.

Second order spatial derivatives of potential and stream functions in Laplace's equation are discretized using the DSC method in which the regularized Shannon's delta kernel is used as an approximation to delta distribution. After implementation of boundary conditions, the system of equations is solved for the unknown terms.

The results are compared with those obtained from the finite element method and the finite difference method.

The method is applied to the flow problem through porous medium for the first time.

In recent years, knowledge graph completion has gained increasing research focus and shown significant improvements. However, most existing models only use the structures of knowledge graph triples when obtaining the entity and relationship representations. In contrast, the integration of the entity description and the knowledge graph network structure has been ignored. This paper aims to investigate how to leverage both the entity description and the network structure to enhance the knowledge graph completion with a high generalization ability among different datasets.

The authors propose an entity-description augmented knowledge graph completion model (EDA-KGC), which incorporates the entity description and network structure. It consists of three modules, i.e. representation initialization, deep interaction and reasoning. The representation initialization module utilizes entity descriptions to obtain the pre-trained representation of entities. The deep interaction module acquires the features of the deep interaction between entities and relationships. The reasoning component performs matrix manipulations with the deep interaction feature vector and entity representation matrix, thus obtaining the probability distribution of target entities. The authors conduct intensive experiments on the FB15K, WN18, FB15K-237 and WN18RR data sets to validate the effect of the proposed model.

The experiments demonstrate that the proposed model outperforms the traditional structure-based knowledge graph completion model and the entity-description-enhanced knowledge graph completion model. The experiments also suggest that the model has greater feasibility in different scenarios such as sparse data, dynamic entities and limited training epochs. The study shows that the integration of entity description and network structure can significantly increase the effect of the knowledge graph completion task.

The research has a significant reference for completing the missing information in the knowledge graph and improving the application effect of the knowledge graph in information retrieval, question answering and other fields.

The purpose of this study is to achieve an accurate intelligent fault diagnosis of rolling bearing.

To extract deep features of the original vibration signal and improve the generalization ability and robustness of the fault diagnosis model, this paper proposes a fault diagnosis method of rolling bearing based on multiscale convolutional neural network (MCNN) and decision fusion. The original vibration signals are normalized and matrixed to form grayscale image samples. In addition, multiscale samples can be achieved by convoluting these samples with different convolution kernels. Subsequently, MCNN is constructed for fault diagnosis. The results of MCNN are put into a data fusion model to obtain comprehensive fault diagnosis results.

The bearing data sets with multiple multivariate time series are used to testify the effectiveness of the proposed method. The proposed model can achieve 99.8% accuracy of fault diagnosis. Based on MCNN and decision fusion, the accuracy can be improved by 0.7%–3.4% compared with other models.

The proposed model can extract deep general features of vibration signals by MCNN and obtained robust fault diagnosis results based on the decision fusion model. For a long time series of vibration signals with noise, the proposed model can still achieve accurate fault diagnosis.

The current popular image processing technologies based on convolutional neural network have the characteristics of large computation, high storage cost and low accuracy for tiny defect detection, which is contrary to the high real-time and accuracy, limited computing resources and storage required by industrial applications. Therefore, an improved YOLOv4 named as YOLOv4-Defect is proposed aim to solve the above problems.

On the one hand, this study performs multi-dimensional compression processing on the feature extraction network of YOLOv4 to simplify the model and improve the feature extraction ability of the model through knowledge distillation. On the other hand, a prediction scale with more detailed receptive field is added to optimize the model structure, which can improve the detection performance for tiny defects.

The effectiveness of the method is verified by public data sets NEU-CLS and DAGM 2007, and the steel ingot data set collected in the actual industrial field. The experimental results demonstrated that the proposed YOLOv4-Defect method can greatly improve the recognition efficiency and accuracy and reduce the size and computation consumption of the model.

This paper proposed an improved YOLOv4 named as YOLOv4-Defect for the detection of surface defect, which is conducive to application in various industrial scenarios with limited storage and computing resources, and meets the requirements of high real-time and precision.

Thin cracks on the surface, such as those found in nuclear power plant concrete structures, are difficult to identify because they tend to be thin. This paper aims to design a novel segmentation network, called U-shaped contextual aggregation network (UCAN), for better recognition of weak cracks.

UCAN uses dilated convolutional layers with exponentially changing dilation rates to extract additional contextual features of thin cracks while preserving resolution. Furthermore, this paper has developed a topology-based loss function, called ℓ_cl Dice, which enhances the crack segmentation’s connectivity.

This paper generated five data sets with varying crack widths to evaluate the performance of multiple algorithms. The results show that the UCAN network proposed in this study achieves the highest F1-Score on thinner cracks. Additionally, training the UCAN network with the ℓ_cl Dice improves the F1-Scores compared to using the cross-entropy function alone. These findings demonstrate the effectiveness of the UCAN network and the value of incorporating the ℓ_cl Dice in crack segmentation tasks.

In this paper, an exponentially dilated convolutional layer is constructed to replace the commonly used pooling layer to improve the model receptive field. To address the challenge of preserving fracture connectivity segmentation, this paper introduces ℓ_cl Dice. This design enables UCAN to extract more contextual features while maintaining resolution, thus improving the crack segmentation performance. The proposed method is evaluated using extensive experiments where the results demonstrate the effectiveness of the algorithm.

The research on structure function recognition mainly concentrates on identifying a specific part of academic literature and its applicability in the multidiscipline perspective. A specific part of academic literature, such as sentences, paragraphs and chapter contents are also called a level of academic literature in this paper. There are a few comparative research works on the relationship between models, disciplines and levels in the process of structure function recognition. In view of this, comparative research on structure function recognition based on deep learning has been conducted in this paper.

An experimental corpus, including the academic literature of traditional Chinese medicine, library and information science, computer science, environmental science and phytology, was constructed. Meanwhile, deep learning models such as convolutional neural networks (CNN), long and short-term memory (LSTM) and bidirectional encoder representation from transformers (BERT) were used. The comparative experiments of structure function recognition were conducted with the help of the deep learning models from the multilevel perspective.

The experimental results showed that (1) the BERT model performed best, with F1 values of 78.02, 89.41 and 94.88%, respectively at the level of sentence, paragraph and chapter content. (2) The deep learning models performed better on the academic literature of traditional Chinese medicine than on other disciplines in most cases, e.g. F1 values of CNN, LSTM and BERT, respectively arrived at 71.14, 69.96 and 78.02% at the level of sentence. (3) The deep learning models performed better at the level of chapter content than other levels, the maximum F1 values of CNN, LSTM and BERT at 91.92, 74.90 and 94.88%, respectively. Furthermore, the confusion matrix of recognition results on the academic literature was introduced to find out the reason for misrecognition.

This paper may inspire other research on structure function recognition, and provide a valuable reference for the analysis of influencing factors.