Yu Zhang

This paper proposes a lightweight image super-resolution reconstruction algorithm (Lightweight NAS Super-Resolution model, LNSR) based on neural network architecture search. The search space of the algorithm is divided into Cell level and Network level. The design of the Cell-level search space is based on the RFDB structure, focusing on searching for combinations of lightweight operators, aiming to build a more lightweight and efficient structure. The Network-level search space focuses on searching the feature connections between Cells, aiming to find the information flow that is most beneficial to improve performance. In order to reduce the time of network architecture search, this paper extends the search strategy of MiLeNAS. At the same time, this paper uses a new type of loss function, which comprehensively considers image distortion, high-frequency detail reconstruction, and model size to promote the model to search for a lightweight and high-performance network structure. The experimental results show that LNSR only needs to spend about 2 days on a 2080 Ti to search for the optimal network structure, and the hardware resources and time consumed are much less than that of FALSR-B. In terms of network scale and performance, LNSR surpasses the artificial design and NAS-based SOTA lightweight method with a lower amount of parameters and calculation.

In metallographic examination, spherular pearlite gradation, an important step in a metallographic examination, is the main indicator used to assess the reliability of heat-resistant steel. Recognition of pearlite spheroidization via the manual way mainly depends on the subjective perceptions and experience of each inspector. Deep learning-based methods can eliminate the effects of the subjective factors that affect manual recognition. However, images with incorrect labels, known as noisy images, challenge successful application of image recognition of deep learning models to spherular pearlite gradation. A deep-learning-based label noise method for metallographic image recognition is thus proposed to solve this problem. We use a filtering process to pretreat the raw datasets and append a retraining process for deep learning models. The presented method was applied to image recognition for spherular pearlite gradation on a metallographic image dataset which contains 422 images. Meanwhile, three classic deep learning models were also used for image recognition, individually and coupled with the proposed method. Results showed that accuracy of image recognition by a deep learning model solely is lower than the one coupled with our method. Particularly, accuracy of ResNet18 was improved from 72.27% to 77.01%.

The precise identification of micro-features on 2.25Cr1Mo0.25V steel is of great significance for understanding the mechanism of hydrogen embrittlement (HE) and evaluating the alloy’s properties of HE resistance. Presently, the convolution neural network (CNN) of deep learning is widely applied in the micro-features identification of alloy. However, with the development of the transformer in image recognition, the transformer-based neural network performs better on the learning of global and long-range semantic information than CNN and achieves higher prediction accuracy. In this work, a new transformer-based neural network model Swin–UNet++ was proposed. Specifically, the architecture of the decoder was redesigned to more precisely detect and identify the micro-feature with complex morphology (i.e., dimples) of 2.25Cr1Mo0.25V steel fracture surface. Swin–UNet++ and other segmentation models performed state-of-the-art (SOTA) were compared on the dimple dataset constructed in this work, which consists of 830 dimple scanning electron microscopy (SEM) images on 2.25Cr1Mo0.25V steel fracture surface. The segmentation results show Swin–UNet++ not only realizes the accurate identification of dimples but displays a much higher prediction accuracy and stronger robustness than Swin–Unet and UNet. Moreover, efforts from this work will also provide an important reference value to the identification of other micro-features with complex morphologies.

Image inpainting is one of the most important problems in the field of image algorithm, and it is also an effective preprocessing method for many other image applications. In this paper, we suggest an image decontamination method which is mainly used to remove mesh stains and also provide a data set for this task. To our knowledge, this work is the first attempt to solve this kind of problem. Specifically, the proposed method is composed of two phases: we first remove the mesh stains with an Improved Conditional Generative Adversarial Network, and then utilize a Semantic Network to fine tune the details. Experiments demonstrated that this two-stage model can remove the mesh stains. Results show that our method significantly out-performs existing methods and achieves superior inpainting results on challenging cases.

Mobile Edge Computing (MEC) is providing cloud computing capabilities within the radio access networks and offering a new paradigm to liberate the mobile devices from heavy computational workloads. Importantly, MEC can effectively reduce latency, avoid congestion, and prolong the battery lifetime of mobile devices by offloading the computation tasks from the mobile devices to a physically proximal MEC servers. Particularly, Virtual Machines (VMs) scheduling is a critical issue for tasks offloading and computation in MEC. Regarding to the VMs scheduling problem in MEC environmnet, this paper pioneers the use of Formal Concept Analysis (FCA) methodology for identifying the mapping from tasks to VMs. Specifically, the VMs profile and tasks descriptions are initially characterized as the formal contexts, respectively. With the constructed formal contexts, the corresponding formal concepts which refer to the rules set, are then generated. To better infuse the rules set of VMs and tasks, this paper defines a similarity measurement between formal concepts of VMs and tasks. Consequently, the matching problem from a given task to a virtual machine is to return the expected virtual machine according to the principle of maximum similarity degree between formal concepts of virtual machine and task. Extensive simulations are conducted with a real dataset for the validation of feasibility and effectiveness of the proposed approach. Specifically, the proposed approach can significantly reduce the energy consumption around 28 percent comparing to the approach without consideration of energy consumption. Overall, It is demonstrated that FCA-based VMs scheduling is a novel solution for a sustainable VMs scheduling in MEC environment.

Recently, deep convolutional neural networks (CNN) have been attracting considerable attention in single image superresolution. Some CNN-based methods, such as VDSR verified that residual learning can speed up the training and significantly improve the performance of accuracy. However, with very deep networks, convergence speed is still a critical issue in training due to the cost of requiring enormous parameters. In order to deal with this issue, we redesign the residual networks based on dilated networks. In this paper, we propose symmetrical dilated residual convolution networks (FDSR) to tackle image superresolution problems. Our network is on the basis of the dilated convolutions supported exponential expansion of the receptive field without loss of resolution and coverage. This means that FDSR can speed up the training and improve the performance of accuracy without increasing the model’s depth or complexity. Meanwhile, we attempt to combine the image pre-processing approach of VGG-net with Mean Squared error (MSE) to enhance the performance. The experimental results demonstrate that the training time-consuming proposed model achieves nearly a half with even superior restoration quality. Further, we present a novel network with less layers and parameters can achieve realtime performance on a generic CPU and still maintain superior performance.

Traditionally, the spatial–numerical association of response codes (SNARC) effect was presented in two-choice condition, in which only one individual reacted to both even (small) and odd (large) numbers. Few studies explored SNARC effect in a social situation. Moreover, there are many reference frames involved in SNARC effect, and it has not yet been investigated which reference frame is dominated when two participants perform the go-nogo task together. In the present study, we investigated which reference frame plays a primary role in SNARC effect when allocentric and egocentric reference frames were consistent or inconsistent in social settings. Furthermore, we explored how two actors corepresent number–space mapping interactively. Results of the two experiments demonstrated that egocentric reference frame was at work primarily when two reference frames were consistent and inconsistent. This shows that body-centered coordinate frames influence number–space mapping in social settings, and one actor may represent another actor's action and tasks.