The purpose of the image captioning task is to generate syntactically accurate and semantically coherent sentences to express the content of a given image, which has a great practical value. The transformer model has a significant advantage in accomplishing this task. In order to solve the problems of Transformer with high attention complexity and internal covariance shift during training, the image captioning model based on external attention and double layer normalization is proposed. On the one hand, external attention is used on the encoding side, it adopts a learnable, external and shared memory unit to reduce the complexity from the second power to the first power, and learns prior knowledge based on the whole dataset, the potential correlation between the samples is mined, which makes the captions generated by the model more accurate. Meanwhile, the row and column of attention matrix are normalized to eliminate the influence of input feature size on attention. On the other hand, the concept of double layer normalization is proposed and used in the Transformer model, and it improves the data expression ability while maintaining the stability of the input data’s distribution. Compared with Up-Down, SRT, M2 and other representative models, simulation experiments on the MS COCO data set show that the improved model achieved scores of 29.3, 58.6, 131.7 and 22.7 in METEOR, ROUGE, CIDEr and SPICE respectively. Experimental results show that the semantic expression of the improved model is more sufficient, the description is more accurate, and the improvement is effective..

Aiming at the shortcomings that crow search algorithm is easy to fall into local optimization and location update strategy is blind, an adaptive crow search algorithm based on Levy flight is proposed, which dynamically adjusts the perception probability and flight distance of the standard crow search algorithm, and optimizes the location update strategy of crow individuals in the second case. The proposed algorithm introduces Levy flight, experience factor and adaptive adjustment parameter mechanism, dynamically increase the global search ability in the early stage and the local optimization ability in the later stage of the algorithm. When the algorithm individual finds that he is tracked by other individuals, it adopts the update strategy of combining empirical factors and Levy flight to guide other individuals, enhance the efficiency of individual location update, avoid individual oscillation around the optimal solution, and make the algorithm reach the extreme point quickly and accurately, it effectively makes up for the blindness of location update and slow convergence speed of the original crow search algorithm. Through the experimental comparison with other new intelligent optimization algorithms in 8 benchmark functions and 1 engineering application problem, the effectiveness of the algorithm is tested. The simulation results show that the optimization average result, standard deviation, convergence and robustness of the proposed algorithm are better than other algorithms, which effectively avoids the blindness of location update and enhances the performance efficiency of the algorithm.

A low-light image enhancement method based on deep learning, called ADLIE (Attention-directed Low-light Image Enhancement) algorithm, is proposed for the problems of uneven illumination and lack of clear recovery details produced in the light images captured in dark scenes. First, a dual-channel attention mechanism is used to guide the attention network, which is regionally graded based on the brightness and darkness of the input image, while the dual-channel attention mechanism can also extract more better local feature information. Then, the attention map is fed into the enhancement network together with the input image. Different levels of light intensity are used for different regions to enhance the image contrast and achieve the effect of uniform exposure of the overall image. Finally, the enhancement module is added to recover the image details using the multi-layer convolutional connections in the enhancement module to obtain a more clearer high quality image. In addition, the experimental validation using open LOL dataset and LSRW dataset collected in real scenes compares classical methods such as Retinex and R2RNet on common evaluation metrics such as PSNR (peak signal-to-noise ratio), SSIM (structural similarity), CTRS (contrast ratio) and Information Entropy (information entropy). The experimental results show a significant improvement. The network restores the details of the image while improving the overall brightness of the low-light image, reduces color distortion, avoids global overexposure, and obtains a clearer and more natural image.

In order to improve the performance of visible light object detection in complex scenes, a visible light object detection algorithm adaptively fused with infrared features is proposed by combining deep convolutional neural network with multi-source information fusion technology. The algorithm takes infrared and visible images as input, extracts infrared and visible features by means of convolution and activation combined with residual structure, and uses spatial and channel attention mechanisms to improve the category of the target and the feature weight of the image region where the target resides. Secondly, the extracted infrared features are incorporated into the visible features of the corresponding dimension in the way of adaptive weighting, which fully makes up for the limitation of the object in the single-mode model; Finally, for multi-scale objects, a pyramid sampling structure is designed. By alternating up-sampling and down-sampling, the global and local features of the feature are fully integrated to enhance the scale invariance of the network. Experiments show that the proposed attention mechanism, feature adaptive fusion and pyramid sampling structure can effectively improve the effect of object detection. Compared with the same type of infrared visible light object detection method, this method can fully integrate the multi-modal features of the object, effectively reduce noise interference, and make the network have higher detection performance. At the same time, in the actual power grid equipment detection, this method also shows high generalization ability and robustness, and can accurately and efficiently achieve the identification and positioning of object equipment.

Aiming at the problems that the feature map is sensitive to illumination changes and the reconstruction is incomplete when using depth neural network for multi-view image 3D reconstruction, a multi view reconstruction method integrating gradient and Gaussian process regression is proposed. a multi-view reconstruction method integrating gradient and Gaussian process regression is proposed. Firstly, aiming at the problem that the illumination change affects the extraction of features, a feature extraction network integrating gradient is designed. Through the independent gradient calculation of the image and the convolution neural network is used to extract features based on the gradient and the original image, the influence of gradient information in the feature map is improved and the inhibition of the influence of illumination change factors is enhanced. Secondly, aiming at the problem that the feature extraction step in multi-view reconstruction only focuses on the current view without considering the potential spatial relationship between views, a view feature enhancement module integrating Gaussian process regression algorithm is proposed, which effectively increases the influence of relevant information between views on the multi-view stereo vision reconstruction task and improves the completeness of the multi-view stereo vision reconstruction results. Finally, the contribution of different views to CostVolume is calculated by measuring the degree of matching between the reference image and the adjacent image features, reconstructing CostVolume that conforms to visual perception. Experiments on the DTU and Tanks and Temples datasets show that compared with the mainstream multi-vision stereoscopic vision reconstruction method, the method has a great improvement in the completeness of three-dimensional reconstruction and has good generalization.

As the number of patients infected with the novel coronavirus increases, a large amount of epidemiological investigation data associated with them has been generated. Based on the data, the semantic association features among patients can be analyzed to express the disease transmission process at the individual level and to explore the distribution of patient characteristics and the transmission paths among patients. Firstly, the semantic relationship of patients is defined based on the analysis of flow modulation data, and the pattern layer of the patient relationship graph is designed accordingly. Then, the data layer is constructed by identifying patients and place entities and extracting "patient-relation-patient" and "patient-residence-place" triplets. Finally, the Neo4j graph database is used to visualize and analyze the patient relationship graph. The results show that the patient relationship graph can explore the intrinsic association of patients, effectively integrate the semantic relationship of patients, and express the process of disease transmission among patients by verifying the super spreader analysis and route of transmission.

To solve the problem of irregular crack trend and difficult to extract the characteristics of small cracks in concrete buildings, an improved crack detection algorithm based on YOLOv4 was proposed. Based on YOLOv4 framework, RFB module with wider receptive field is introduced in the feature extraction network to capture feature images. Based on the multi-scale path fusion structure of PANet, a new multi-scale feature fusion method sl-PANET is proposed. Firstly, the shallow network feature information is added to improve the accuracy of the model in identifying fine cracks. Secondly, the upper sampling module of DUpsampling is adopted to fully restore the image feature information. The CBAM attentional mechanism module was incorporated in the up-sampling and down-sampling processes to highlight the fracture feature information and remove the interference of background redundant information, so as to enhance the expression ability of fracture feature. The AdamW optimizer is also used to accelerate the convergence of network training. Experimental results show that the detection accuracy of the improved algorithm is as high as 94.47%, which is 6.44% higher than the original YOLOv4 algorithm, and can meet the current crack detection requirements of concrete buildings.

How to detect the fire source efficiently and accurately locate it is an important prerequisite for effectively controlling the deterioration of fire situation and making fire control plan in time. At present, the main problem faced by fire source detection and location is that the dual tasks of fire source detection and location are separated from each other, which seriously restricts the real-time performance of fire early warning. To overcome the above problems, YOLO V5 is proposed as the basic model of fire source detection, and CIOU (Complete intersection over union) loss function is used to accurately frame anchor (anchor-boxes) and GT (Ground Truth) to further improve the annotation accuracy of the model. The leaky RELU activation function is replaced by GELU (Gaussian Error Linear Unit), which combines regularization and activation function. In addition, while accurately detecting the fire source, the parallel binocular location algorithm is used to locate the fire source in space, to realize the intelligent integration of fire source detection and location. The experimental results show that the fire source detection map value of the proposed method is 9.8% higher than the original algorithm, which can accurately locate the fire source while ensuring the accuracy of fire source detection.

Mixed-criticality system is one of the main trends in the development of modern embedded systems. The high critical task represents the practical task with high urgency or importance, which usually needs to be guaranteed first. In order to ensure the execution of high-level critical tasks, the current mixed critical task scheduling algorithms often discard or schedule low-level critical tasks in a timely manner. As a result, the task loss time limit rate is large and the system utilization rate is low during the critical level conversion. Therefore, the EDF-os semi-partition scheduling algorithm is improved in hybrid critical systems with dual critical levels. First, in the division phase, the tasks at high critical levels are treated as fixed tasks and the tasks at low critical levels are divided by utilization using the Worst-Fit policy. Secondly, in the execution phase, the form of job boundary migration is used, and the strategies for determining the priority of tasks at different critical levels under different system critical levels are discussed in detail, and tasks are scheduled according to the priorities. Finally, a multi-processor hybrid critical system with dual critical levels is simulated, and task sets are randomly generated for simulation experiments. The results show that the proposed method increases the executable ratio of low-critical level tasks by 14.8% on average, and decreases the task loss time rate by 19.7%.

In order to solve the problems of long response speed and slow convergence speed of traditional PTAT current source during startup, a new PTAT current source with fast startup and stability is proposed. Based on the highly reliable bipolar structure, an operational amplifier structure with weak mismatch mechanism is introduced. By using local negative feedback and shared load capacitance, the chip area is optimized, and at the same time, the circuit startup speed and loop convergence stability speed are greatly accelerated. The circuit performance has been greatly improved in three aspects: power-on start-up, stable convergence and working mode. The results show that under the standard UMC 180 nm CMOS process, the circuit can achieve a wide temperature range of −40 C to 85℃, a wide swing voltage of 2.5-5 V, an average current of 104.26 μA, a startup speed of 4 μs and a high PSRR rejection ratio of 108 dB. It has a good application prospect in system chips with high requirements for startup and convergence speed.

In this paper, a high performance temperature protection for GaN HEMT gate driver chip is designed, which can accurately respond and output protection signal to ensure the circuit safety. The over temperature protection uses two temperature detection circuits to collect the voltage value of the temperature signal and amplify the voltage difference, then after comparative filtering, a Schmitt trigger with hysteresis function outputs the shaping protection signal, which can overcome the influence of common mode noise and temperature stress. The design is based on CSMC 0.18 μm BCD process, and the circuit design verification and testing are completed. The results show that the circuit function is correct and can meet the application requirements of GaN HEMT gate driver.

CVSD coding is widely used in the field of speech communication in military wireless, satellite, underwater and other complex environments for its simple implementation and strong error correction ability. However, the hybrid double integral delta modulation system with standard CVSD coding is unstable. In the face of channel burst error, the lack of anti-interference ability easily leads to the degradation of speech quality. An improved CVSD coder and decoder is designed. Based on the delay of speech coding and the correlation of speech data, the error measurement is carried out by continuously tracking multiple paths, and the optimal output sequence is found from a variety of possible coding outputs by using appropriate distortion criteria. The instability of the hybrid double integral delta modulation system can be solved. At the same time, the improved digital diffuser is used to improve the signal-to-noise ratio of coding and decoding, solve the error accumulation problem introduced by table lookup method, and further improve the adaptability of CVSD coding in special channels. Finally, the usability of the improved CVSD compiler is tested and evaluated based on an underwater voice communication terminal. The experimental results show that the improved CVSD encoder and decoder has stronger error resistance and better speech quality than the standard CVSD encoder in terms of stability, and has higher application and promotion value.

A DC/DC-controller-adaptable test and trim circuit is designed based on the commercial process of 0.25μm BCD. Based on the thought of pin multiplexed and by applying specific signals on specific pins of the circuit, it could readout internal signals, trim key parameters, and overcome parasitic effects of package on high voltage resolution and large current applications without effecting the common pin signals’ status in normal operation, featuring easy test operation, cheap test cost and extensive test range. As a result, the size problem that traditional PAD probe trim need to occupy more area when trimming more fuses would be relieved, the cost problem that traditional laser trim need expensive equipment and programs would be avoided, and the common problem that both PAD probe trim and laser trim can only be feasible at wafer level and unfit to be used in high voltage resolution and large current application would be improved. The circuit is comprised of register clock and data input circuit, test and trim enable circuit, test and trim array circuit, and test data output circuit. Simulation results show that under the circumstance of open loop, by configuring specific data bits, the design lends itself well to test internal signals such as oscillator’s output and parameters such as on-resistance, and trim critical parameters such as reference voltage etc.

The insulated gate bipolar transistor (IGBT) is an important part of power converter, and the prediction of its remaining service life is very important. In response to the remaining service life of IGBT, the method of optimizing the adaptive options of the Elman neural network implementation of the Elman neural network is optimized by using the slime mould algorithm (SMA), and it is used for the life prediction of IGBT. Firstly, the peak of the gate emitter turn-off voltage in the aging test data set of NASA research center is smoothed. Secondly, the time domain feature is extracted from the processed data. Thirdly, the kernel principle component analysis (KPCA) is used for optimization dimensionality reduction. Finally, the SMA-Elman neural network model is used to predict the lifetime of IGBT. The results show that the proposed SMA-Elman neural network has better performance than Elman and BP neural network and SVR, the mean square error is 0.021%, the root mean square error is 0.014, the fitting degree is 0.998, and it can better predict the remaining service life of IGBT.

Based on simulation and experimental methods, the design and research of 100VN trench MOSFET is carried out. Through the trench depth, body dose and gate oxide thickness experiment, the effects on breakdown voltage, threshold voltage and on-resistance are obtained and the mechanism is analyzed. The current path inside the device and the distribution of impact ionization rate can be seen through the simulation. As the trench depth increases, the breakdown voltage first increases and then decreases, and the on-resistance shows an opposite trend; the breakdown voltage has a weak correlation with the implant dose, and the threshold voltage increases with the increase of the implant dose; the breakdown voltage increases with the gate oxide thickness increases, but the change range is not large, and the threshold voltage has a strong correlation with the thickness changes. Through gradual optimization, the final structure and process parameters are obtained as trench depth is 1.5 um, body dose is 1.3E13, gate oxide thickness is 700 A and the device final electrical parameters are obtained by real tape-out. Finally, we can get the breakdown voltage is 105.6 V, threshold voltage is 2.67 V and on-resistance is 3.12 mR. Compared with the simulation results, the change rates are 98%, 94% and 75% respectively.

System-in-Packet (SiP) technology integrates multiple subsystems in one package, which has the advantages of flexible assembly method and short development cycle, and has broad development prospects in the process of miniaturization of electronic equipment. In the SiP design process, whether the schematic design is correct often determines the success or failure of the overall design. However, connectivity errors in schematic designs often require engineers to spend a lot of time looking for comparisons to pinpoint the location of the error. In order to improve the efficiency of schematic connectivity error checking, this paper proposes a connectivity rule checking error back-marking tool applied to the SiP system-in-package schematic design stage. The tool is integrated into the OrCAD Capture CIS tool in the form of a plug-in, which can cooperate with the existing schematic rule checking tool, so that users can obtain and analyze the valid error information generated by the rule checking tool through the graphical interface, and the error information is clear and intuitive. By performing an error back-labeling test on a test system consisting of 26 pages of schematics, the tool can back-label the connectivity error information in the schematic to the corresponding location on the schematic within seconds, allowing designers to quickly locate the wrong location, which effectively improves the efficiency of the connectivity check in the schematic design stage.