Conventional computers based on von Neumann architectures have limited performance due to the physical separation of memory and processor. To achieve sustainable development of advanced neuromorphic computing technologies in the post-Moore era, neuromorphic devices with an integrated memory-computer architecture have emerged as an alternative solution. Here, we demonstrate an optical modulation-based artificial synapse that achieves phototunable coordination of conductance by virtue of its excellent photoresponsive properties, and exhibits excellent optoelectronic synaptic behaviours, successfully simulating key behavioural features of biological synapses including stimulus reinforcement, training facilitation, and memory consolidation, and evaluating the predictability of device forgetting. This study has important implications for the implementation of hardware artificial neural networks (ANNs) and the simulation of perceptual systems.

In order to achieve high-sensitivity monitoring of chemical materials and drugs with low refractive index, a new type of high-sensitivity photonic crystal fiber (PCF) sensor with elliptical side core is designed based on the surface plasmon resonance (SPR). A circular hole and three different sizes of elliptical holes constitute the sensor air hole, and the gold nano film is coated in the left elliptical hole. When light enters the sensor, it will generate resonance at the gold thin film. The refractive index of the liquid to be measured can be obtained by analyzing the spectrum obtained from this physical phenomenon. The finite element method is adopted to systematically investigate the effects of ellipticity of elliptical holes and the thickness of gold thin film on the sensitivity of the proposed sensor. The results reveal that the maximum sensitivity of the sensor reaches 95 000 nm/RIU (RIU is refractive index unit) when the refractive index of the liquid to be measured ranges from 1.41 to 1.44. In view of the existing researches, the sensitivity is 3.7—47.5 times of that of typical available PCF-SPR sensors within the same range of refractive index. Therefore, the proposed PCF-SPR sensor can be widely applied in fields, such as material monitoring, food safety, and biomedicine.

Plasma induced transparency (PIT) is a novel optical response characteristic, which has unique advantages in applications such as sensors. In this paper, we present an optical model which is coupled by a double-ring resonator and a straight waveguide, and explain the cause of the multi-band PIT effect. The influence of the main structural parameters on the transmission characteristics of the structure is analyzed by using the finite difference time domain (FDTD) method. The effects of inner and outer radii Rin,Rout and the width WR of ring resonators on the red shift of transmission characteristics are studied. it is found that Rin and Rout are too large and too small to be conducive to the phenomenon of PIT. After optimizing the parameters, a nano-sensor with a sensitivity of 1 014.10 nm/RIU and a high quality factor of 90.65 is designed, which has the advantages of simple structure, easy integration, convenient fabrication and low cost, it can be well used in the field of high integration and high precision sensing detection.

To address the problems of incomplete feature extraction of architectural elements and difficulties in the recognition of similar architectural styles, we propose a salient region suppression and multi-scale feature fusion (SRSMSFF) architectural style recognition method. First, the improved Resnet18 is used to extract the initial architectural features. Next, the salient region suppression module (SRSM) is designed, which guides the network to learn the features of potential regions by hiding the most discriminative regions. And multi-scale feature fusion (MSFF) is designed, which combines multi-scale structure with salient region suppression to obtain a more complete feature of architectural elements. Then, channel attention is used to assign corresponding weights to each channel, which can highlight important channel information. Finally, the large-margin Softmax loss function (L-Softmax) is introduced through maximizing the decision boundary distance of the feature embedding space, which improves the performance of similar architectural style recognition. The experimental results show that our model achieves 64.44% and 80.21% accuracy on the 25-class and 10-class public architectural style datasets. It achieves an accuracy of 88.21% on the dataset of ancient Chinese architectural styles. Its performance is superior to current advanced method.

Aiming at the problem that the existing cross-modal retrieval methods are difficult to measure the weight of data at each node, and there are limitations in mining local consistency within modalities, a cross-modal image and text retrieval method based on multi-head attention mechanism is proposed. Firstly, a single image and text sample serves as an independent node when constructing the modal diagram, and graph convolution is used to extract the interaction information between each sample to improve the local consistency in different modal data. Then, attention mechanism is introduced into graph convolution to adaptively learn the weight coefficients of each neighboring node, thereby distinguishing the influence of different neighboring nodes on the central node. Finally, a multi-head attention layer with weight parameters is constructed to fully learn multiple sets of related features between nodes. Compared with the existing 8 methods, the mAP values obtained by this method in experiments on the Wikipedia dataset and Pascal Sentence dataset increase by 2.6% to 42.5% and 3.3% to 54.3%, respectively.

Aiming at the different sizes of the disease manifestations of citrus leaves, the model of improved CenterNet is proposed to solve the problems of missing detection, false detection and low accuracy in the detection process. The feature enhancement improved atrous spatial pyramid pooling (IASPP) module was introduced into a series of residual structures of the first two residual layers of the feature extraction network RestNet50 to expand the shallow receptive field, obtain more detailed information of small target leaf diseases, and enhance the significance of shallow features. The bidirectional feature pyramid network (BiFPN) module was introduced to effectively integrate the shallow and deep leaf disease information of the feature extraction network. In order to improve the overall detection effect, the multi-scale channel attention module (MS-CAM) was introduced. The trained model was used to detect citrus disease leaves. The experimental results show that, compared with the original model CenterNet, the R-value of the proposed model is increased by 8.32%, mAP is increased by 4.53%, and AP0.5:0.95 is increased by 27.3%. It can achieve accurate detection of small target, medium target and large target leaf diseases in citrus planting.

Aiming at the problem of missing detection in the process of small target detection due to insufficient semantic information of shallow features, a multi-layer feature fusion improved single shot multi-box detector (SSD) method is proposed. Firstly, deepwise separable convolution (DSC) is added to the shallow network, and the shallow semantic information is strengthened by channel-by-channel convolution and point-by-point convolution. Then the features of deep network and shallow network are refined utilizing deconvolution and dilation convolution. Finally, the attention mechanism is added to the deep network to enhance the detection ability of small targets. Verified on VOC2007 and VOC2012 data sets, the average detection accuracy is improved by 5.56% compared with the benchmark algorithm and 4.25% compared with other advanced algorithms. The experimental results show that the proposed refined semantics and enhanced perception methods can achieve the purpose of improving the detection accuracy of small targets.

A fusion algorithm for infrared and visible light images combining visible light image enhancement and multiscale decomposition is proposed to address the problems of detail information loss and low contrast in infrared and visible light image fusion in recent years. Firstly, an adaptive visible light image enhancement method is proposed to improve the overall contrast of visible light images and enhance the detail information in visible light images. Then, a multiscale decomposition algorithm based on Gaussian filtering and rolling guided filtering is proposed to decompose the source image into small-scale layers, large-scale layers, and base layers. The small-scale layer fusion adopts a fusion rule based on maximum absolute value, the large-scale layer fusion injects the infrared spectral features into the visible light image using nonlinear weight coefficients, and the base layer adopts a fusion rule based on visual saliency mapping to avoid contrast loss. Finally, each scale layer is reconstructed to generate the fusion image. Experimental results show that compared with other algorithms, the proposed method has improved the objective evaluation indicators such as edge preservation, human-inspired perceptual metrics, spatial frequency, standard deviation, and edge intensity by an average of 23.50%, 30.38%, 46.67%, 50.41%, 20.17%, and 54.19%, respectively, and the generated fusion image also performs well in subjective evaluation.

In order to overcome the problem of phantom shadow at the edge of the object in the process of infrared and visible image fusion, an image fusion method based on divide region universe algorithm (DRUA) is proposed. Firstly, the space model of the universe is established, the universe in the core region is communicated with each other bilaterally, and the universe in the non-core region is only communicated with each other unilaterally. Secondly, the connected universe connecting the core region and the non-core region forms a neighborhood relationship between the core region and the non-core region. Thirdly, the detail layer, the basic layer and the saliency layer of the infrared and visible images are obtained respectively, and different fusion strategies are used for different layers. Finally, the fusion process of universe optimization is given. Experimental results show that the proposed method has more thermal target and scene information, and the visual effect is consistent with the visual characteristics of the human eye. Compared with other algorithms, the mutual information, standard deviation, spatial frequency, and average gradient index are increased by 17.57%, 14.80%, 34.16% and 21.01%, respectively, so that it is better than other algorithms.

To address the problem of unsatisfactory classification results due to the limited number of labeled samples and insufficient extraction of diverse features in hyperspectral image classification tasks, this paper proposes a hyperspectral image classification method based on three-dimensional dilated convolution and graph convolution. Firstly, we introduce different scales of dilated convolution (DC) to build a three-dimensional dilated convolution network model to extract multi-scale deep spatial-spectral features. Secondly, we build a graph convolution neural network model by aggregating the neighborhood feature information of graph nodes to obtain the contextual features with spatial structure. Finally, to improve the representation capability of diverse features, we fuse deep spatial-spectral features with spatial contextual features and use Softmax to achieve classification. The proposed method can make full use of the diverse features of hyperspectral images and has a strong feature learning capability, which can effectively improve the classification accuracy. The proposed method is experimentally compared with seven related methods on the hyperspectral datasets of Indian Pines and Pavia University, and the results show that the proposed method could obtain optimal results with an overall classification accuracy of 99.33% and 99.41%.

To obtain the optimal radiation characteristics of the electron under the collision of a circularly polarized laser pulse, several transverse initial positions of the electron are selected with the method of controlling variables by using MATLAB programs. The effect of the transverse initial positions of the electron on the trajectory and radiation characteristics of high-energy electrons is discussed. The results reveal that when the transverse initial position of the electron moves in the positive x-axis direction, the circularly polarized intense laser pulse has a considerable effect on the electron trajectory and radiation properties. According to the analysis of the radiation power of the electron, when the transverse initial position is 0.15 0 (0=1 m), the maximum radiation power generated by the spiral motion of the electron will reach the optimal state. By simulating the model of the interaction between circularly polarized intense laser pulses and electrons, the distribution of electron radiation power in the entire space is found, which provides a theoretical basis for accurate experiments in practice.

GaN with a wide band gap, high quantum efficiency, excellent thermal stability, and radiation resistance is important role in high frequency, high power electronics, and UV photoelectron devices. In this study, we present a novel approach by utilizing high-energy N plasma as the N source for synthesizing the GaN films with higher crystalline quality. This process occurs at a relatively low temperature of 850 °C, utilizing the economical and eco-friendly plasma-enhanced chemical vapor deposition (PECVD) method. Furthermore, the effects of N2 flux on the crystalline quality of the films, growth rate and optical characteristics are investigated. The results reveal that an increase in N2 flux enhances both the film growth rate and crystalline quality by boosting the kinetic energy of reacting atoms. Nevertheless, a further increase in N2 flux results in excessive nucleation rate, preventing atoms adsorbed on the substrate from migrating to appropriate positions. Consequently, the films grow in random directions, leading to a decline in crystalline quality. The GaN films prepared in this study achieve a carrier concentration of 2.19×1018 cm-3 and mobility of 5.17 cm2·V-1·s-1, demonstrating significant potential for optoelectronic device applications.

Existing lightweight networks for classifying thoracic diseases have a large number of parameters and require significant hardware resources. This paper proposes a lightweight algorithm for classifying thoracic diseases based on mixed knowledge distillation (KD) training strategy. Firstly, the algorithm incorporates an optimized residual shrinkage module into the MobileViT base network and employs soft thresholding to filter background noise in X-ray images. Then a mixed knowledge distillation training strategy is proposed, utilizing multi-level attention maps and similarity activation matrices as supervisory signals to enhance the ability of lightweight networks to recognize thoracic diseases. Finally, the focal loss function is employed to address the imbalance between positive and negative samples in the dataset. Experimental results on the ChestX-Ray14 dataset demonstrate that the average AUC value for the RMSNet student model trained with distilled knowledge to recognize 14 types of thoracic diseases is 0.836. The number of parameters and FLOPs are only 0.96 M and 0.27 G, respectively. These results indicate that the proposed algorithm improves classification accuracy while maintaining lightweight, enabling the network to run with less hardware.

In order to solve nonnegligible problems in brain tumor magnetic resonance imaging (MRI) segmentation, such as few samples, class imbalance and low accuracy of small districts, this essay proposes a new multi-scale and multitask deep-learning algorithm called TDDU-Net based on 3D No-New U-Net. Firstly, this paper applies the structure with an encoder and three different decoders to the network. Next, the ConvXt module is a pre-processor of the original decoder with a reverse bottleneck structure in order to overcome the underutilization of high-level semantics when some core regions decode. Then, InConvXt is a generalized feature processing module at the bottom layer between the encoder and decoder to ensure the accuracy of the generalized features and enhance the stability of the network. Finally, the deepwise convolution is used to reduce the calculation amount of the network parameter at the appropriate location while ensuring accuracy. The experiments show that the Dice similarity coefficients (DSCs) of the predicted segmentation in the BraTS18 dataset reaching 0.907, 0.847, 0.807 in the whole tumor region (WT), the tumor core region (TC) and the enhancing tumor region (ET). The method performs better, which is helpful in segmenting the smaller tumor area in MRI.