Aiming at the problems of YOLOv4 target detection algorithm in some application scenarios with too many parameters, complex network and low accuracy, an improved lightweight target detection algorithm GD-YOLO was proposed. Firstly, the main feature extraction network of YOLOv4, CSPDarknet, is replaced by the lightweight network GhostNet, which greatly reduces the number of parameters and computation of the algorithm and makes the algorithm more lightweight. Secondly, the double attention mechanism (DATM) is proposed, which not only strengthens the spatial and channel features of the model, but also has a small number of structural parameters. The double attention mechanism is added to the three effective feature layers extracted from the backbone network to make the model more effective for feature extraction. Finally, ACON activation function was added to replace ReLU activation function in GhostNet network to further improve the detection accuracy of the algorithm. Experimental results on VOC2007+2012 data set show that the GD-YOLO algorithm has an average accuracy (mAP) of 84.28%, which is 4 percentage points higher than YOLOv4 algorithm and about 1 percentage point lower than YOLOv5 algorithm. Compared with YOLOv4 algorithm, the number of model parameters is reduced by 11M, and 3M compared with YOLOv5 algorithm. Compared with YOLOv4, the proposed GD-YOLO algorithm not only reduces the number of model parameters, but also preserves a higher average accuracy, indicating that the algorithm is more lightweight and has higher accuracy.
With the development of multi-core embedded real-time systems, the DAG task synchronization problem has received extensive attention. Most of the current task synchronization methods use the lock mechanism, but there are many problems in the lock mechanism, such as the spin lock in the task busy waiting state, which wastes CPU resources; If a task using a mutex cannot obtain a shared resource, it will be blocked, resulting in context switching overhead; Sequential locks allow write tasks to have higher priority, but write tasks cannot update data frequently, otherwise read tasks will starve to death. If the above lock mechanism is applied to the synchronization of DAG tasks on a multi-core platform, it will not only affect the overall execution efficiency of the system, but also cause subsequent tasks to fail to execute, and in severe cases, will lead to deadlocks and system crashes. Therefore, the DCAS lock-free mechanism is used in the DAG task synchronization process, and the while condition is used to avoid the ABA problem. Under the LITMUSRT multi-core platform, taking multitasking to apply, fill and release Vxworks network buffers at the same time as an example, the triplet mBlk, clBlk, and cluster in the buffer pool respectively use the DCAS lock-free mechanism. The experimental results show that compared with the traditional lock mechanism, the DCAS lock-free mechanism has a better effect in DAG task synchronization, the response time is reduced by 10.4%, and the overall execution efficiency of the system is improved by 4.2%.
In the field of natural scene text detection, the existing deep learning network has the situation of text false detection, missed detection, and inaccurate positioning. To solve this problem, a text detection algorithm based on Large Receptive Field Feature Network (LFN) is designed. First, ShuffleNet V2 is selected as a lightweight backbone network with better speed and accuracy, and a fine-grained feature fusion module is added to obtain more hidden text feature information. Then the double fusion feature extraction module is constructed to extract multi-scale features from the input image by analyzing the different receptive fields of different scale feature maps and the influence of the size of the feature map obtained after normalization of feature maps with different scales on the results is compared, thereby reduced feature loss and increased the receptive field. Finally, Dice Loss is introduced into the differentiable binary module to deal with the imbalance between positive and negative samples and increase the accuracy of text location. Experiments result on the ICDAR2015 and CTW1500 datasets show that the network has significantly improved text detection in both performance and speed. The F1 on ICDAR2015 dataset is 86.1%, which is 0.4% higher than PSENet method with the best performance and the speed is 50.3fps, which is about 1.92 times higher than DBNet method with the fastest speed. The F1 on CTW1500 dataset is 83.2%, which is 1% higher than PSENet method and the speed is 35fps, which is about 1.65 times higher than EAST method.
Most of the current image denoising algorit hms usually cause the loss of edge detail information of the image while removing noise. Aiming at this problem, an image denoising method based on edge feature fusion is proposed. First,the edge information of the image is extracted by the edge extraction network based on Canny operator. Because the Canny operator does not need to be trained, the denoising time is shortened to a large extent. Secondly, the initial denoising network based on dense residual connections is used to ensure the stability of the training and avoid the disappearance of the gradient to achieve the initial denoising of the image. Finally, through the fusion network based on channel and spatial attention mechanism, the extracted edge information image is fully fused with the preliminary denoised image, and the relatively important edge information is adaptively allocated with more weight, and the edge details of the image are enhanced, so as to get a clear image with more edge information. Experimental results show that on BSD68 and Set12 datasets, compared with the common denoising methods such as DnCNN and BM3D, the average PSNR of the proposed denoising method is 0.13 dB and 0.29 dB higher than DnCNN, and 0.76 dB and 0.82 dB higher than BM3D, respectively. In terms of visual effects, more image details are retained, and the denoising rate is also greatly improved.
The desert areas of Xinjiang are prone to drought disasters and agricultural and animal husbandry production under the dual influence of climate and environment, which is not conducive to the sustainable economy of Xinjiang, the identification of desert plants in Xinjiang is the basis for various plant researchers to understand the growth status of plants, as well as a prerequisite for ecological conservation research and implementation of management measures. At the same time, the study is difficult due to the similarity of Xinjiang desert plant images between classes, complex image background and unbalanced data samples. In order to improve recognition accuracy, accurately locate locally important features and comprehensively consider complex global information, a plant image recognition method that combines convolutional neural network (CNN) and Swin Transformer network is proposed. The method combines the advantages of CNN network which is good at extracting local features and Swin Transformer which is good at capturing global representation, and embeds an improved Convolutional Block Attention Module (CBAM) in the CNN branch to fully extract the local key features with differentiation, and the Focal Loss function is used to solve the problem of data sample imbalance. The experimental results show that the proposed fused method can extract the features of the images more adequately than the single-branch network on the Xinjiang desert plant dataset, and its recognition accuracy can reach 97.99%, and the precision, recall and
Developing deep learning methods for intelligently diagnosing COVID-19 based on medical lung imaging can ease the work of a large number of healthcare workers and provide reliable accuracy, but the high accuracy of deep learning methods often depends on the quality of data samples. The source and processing process of medical image data that exists in nature is not single, and the large difference and poor quality of data samples will increase the difficulty of deep learning models to extract key features, and effective data pre-processing and appropriate model design are critical. Based on lung CT images, this paper proposes a pixel-segmentation combined dual-branching model ReSWNet to assist in the diagnosis of COVID-19 infection. This method first trains the pixel segmentation model for segmentation preprocessing to achieve the rejection of the irrelevant background of lung CT images, and then performs pneumonia diagnosis by combining the advantages and disadvantages of convolutional neural network and self-attention model. The method was validated on the COVID-CT dataset and showed that the method improved by 8.6%, 16.05 and 7.71% compared with the baseline model in terms of diagnostic accuracy, recall rate and F1 score, respectively, and finally the visualization of the results heat map provided interpretability for diagnosis.
To address the problem of insufficient utilization of image visual features and channel feature information in the process of text-to-image synthesis task, a text-to-image synthesis method based on Feature-enhanced Generative Adversarial Network (FE-GAN) was proposed. Firstly, a Memory on Memory (MoM) module was designed to pay attention to and fuse the generated intermediate features during dynamic memory reading. The attention mechanism was used to enhance the first visual features when memory was read, and then the obtained attention results were fused with the image features generated by the previous generator to achieve the second image visual feature enhancement. Then, channel attention was introduced into the residual block to obtain different semantics in image features, enhance the correlation between similar semantic channels, and achieve channel feature enhancement. Finally, the Instance Normalization Upsampling Block and the Batch Normalization Upsampling Block were combined to improve the image resolution, while mitigating the influence of the batch size on the generation effect and improving the style diversity ability of the generated image. Simulation experiments showed that the Inception Score (IS) of the proposed method reaches 4.83 and 4.13 respectively on the datasets of Caltech-UCSD Birds-200-2011 (CUB-200-2011) and 102 category flower dataset (Oxford-102), which are 1.68% and 5.62% higher than those of DM-GAN, respectively. Experimental results show that the images generated by FE-GAN are better in detail processing and more consistent with text semantics.
Aiming at the coverage control problem of the manipulator with strong nonlinear characteristics, based on Voronoi diagram theory, a region optimal coverage control algorithm for multi-manipulator systems is proposed. Firstly, the target region is divided into Voronoi regions by calculating the position of the end effectors of each manipulator; Secondly, according to the convex optimization theory, the optimal movement of the joints and the end effectors of the multi-manipulator systems are measured by the defined objective cost function that describes the effect of area coverage control; Finally, combined with the special dynamic characteristics of the manipulator system, the distributed area optimal coverage controller of the multi-manipulator systems is given. The Lyapunov stability theory is used to analyze the stability of the algorithm. Numerical simulation experiments show that the algorithm is effective, that is, the proposed algorithm can make the end effectors of the multi-manipulator systems reach the corresponding Voronoi region centroids with the minimum cost function value, and the speed gradually converges to zero, forming the optimal coverage of the target region. In particular, the algorithm takes the manipulator as the research object, which enriches the existing research on the coverage control agent model. In addition, based on the nonlinear structural characteristics of the manipulator, the coverage control law of the task space designed in the algorithm can also be applied to the coverage control research of the second order system agent, expanding the existing coverage control research based on the first order system.
The identification model based on fault signal frequency has low accuracy due to the influence of surrounding magnetic field. In order to improve the accuracy of real-time identification of transient fault signal of smart meter circuit, from the direction of fault signal processing, a real-time identification method of transient fault of smart meter circuit based on time domain rough detection method is studied. To sample the fault signal of the smart meter circuit and make the fault signal transmit according to the specified route, it is necessary to increase the receiving capacity of the channel. The complex electrical signal is preprocessed by using the wavelet analysis method. The signal after time-domain detection is uniformly redistributed and marked in a relatively simple way; The fault signal fusion model based on wavelet entropy measure adopts the information entropy method to fuse and separate similar fault signals, and determines the transient characteristics of fault circuit signals of all smart meters; A fault signal fusion model based on wavelet entropy measure is constructed to complete the signal screening and decomposition on the signal spectrum, and realize the real-time identification of the transient fault signal of the smart meter circuit. The experimental results show that the difference between the fault voltage curve value identified by this method and the actual value is less than 0.3 V, which has high identification accuracy and certain practical application significance for the four different types of transient faults of the smart meter circuit, i.e. grounding short circuit, two-phase grounding short circuit, phase to phase short circuit and three-phase grounding short circuit.
Under the background of automation upgrading, the coupling between equipments is constantly improving, and the fault manifestations are complex and diverse. Failure to deal with a single fault in a timely manner is likely to expand the scope of influence, which further escalates the accident. In order to ensure the normal operation of the equipment, new requirements are put forward for the traditional fault diagnosis methods based on case analysis, including low cost, long-term monitoring, small sample or zero sample fault identification. To meet these new requirements, this paper proposes to introduce the zero sample classification recognition idea used in the image processing field into the fault diagnosis field. By studying the characteristic parameters of the existing fault samples, the characteristics used for condition monitoring are determined through optimization. The fuzzy neural network is used to form a feature attribute descriptor, which describes the characteristics as equipment attributes. Based on the attribute description, the ART network conducts incremental learning for long-term monitoring of equipment at the same time. That is to build a monitoring and learning mechanism based on the analysis of a small number of equipment samples or similar samples, identify the original faults and learn and record new faults at the same time. In order to illustrate the feasibility and effectiveness of this method, this paper uses a small number of motor fault data sets as prior knowledge to build a system, and mixed unknown fault samples for system testing. The experimental results show that the application of zero sample classification is expected to solve the new challenges of equipment fault diagnosis under the current technical background.
The current/frequency converter is mainly used to measure the output current of accelerometers,the logic control method of typical extended width reset charge balance current /frequency converter has a great influence on the accuracy. To overcome the control logic shortage of typical logic control method, a new logic control method is given. The integrator output of the circuit realizes the analog-to-digital conversion output through the primary D trigger as the comparison circuit,then the counter start trigger pulse is generated by the self-locking circuit formed by the second D trigger and the and gate, finally the switch control logic is formed by counting of the counter circuit. The new method avoids the constraint of integration time in typical method ,and reduces the difference of the average value of the integrator output waveform between high current input and low current input, thus the circuit accuracy is effectively improved. The performance of the same hardware circuit is tested by writing different CPLD algorithmic programs, the results show that the new method can significantly improve the circuit accuracy compared with the typical method.
In recent years, tightly coupled AI processors have received extensive attention in resource-constrained edge-side intelligent processor applications. But when doing early design space exploration for the main coprocessor in the pipeline coupling relationship. There are the characteristics of shared hardware resource relationship, complex and diverse data path structure, and heterogeneity of on-chip main-cooperative computing features. This makes the simulation evaluation modeling for AI processors facing challenges. This paper focuses on the structural characteristics of tightly coupled AI processors. Abstract the hardware structure into a software simulation model framework. By analyzing the basic hardware resources of the main coprocessor, the different data paths controlled by the instruction are decomposed, and the AI processor simulation model is designed. The main processor and the coprocessor are used in trace simulation and model analysis, respectively. Introduce hybrid trace record timestamps to count widget access information. Then, combined with the analytical performance evaluation algorithm, the performance evaluation of the AI processor is realized. The experimental results show that the AI processor model and evaluation analysis based on the hybrid trace can effectively solve the actual execution results of AI computing, and evaluate the performance of the hardware, including important parameters such as delay, energy and power.
In order to improve the modeling accuracy and control performance of switching power converters, a state-switching discrete time model (SSDM) is proposed. The model considers the state of each switching period instead of the average state of the converter. Therefore, its accuracy is higher than that of the traditional state average model at high frequency. Based on the full differential equation of the switching duration, the inductor current and the output voltage in a cycle are accurately calculated, and the SSDM is derived. In addition, the limiting voltage response (LVR) control strategy is deduced by SSDM. The LVR control strategy calculates the appropriate duty cycle by voltage prediction to adjust the output voltage to the reference value in the minimum switching period. With this strategy, the converter not only achieves very fast load/line transient response and reference tracking speed, but also exhibits high stability under deviated inductance. Finally, the accuracy of the proposed strategy and the stability of the system are verified by frequency response analysis and experiments. The experimental results show that the transient response time of the output voltage under the LVR control strategy is more than 70% shorter than that under the traditional control strategy under different working conditions. When the inductance value deviates by 23%, the output voltage remains stable under the LVR control strategy.
Random forest has been successfully applied in various fields due to its advantages of fast training speed, difficult over fitting and easy realization. In order to solve the problem that the power consumption test of different memory unit sizes, voltages and temperatures is needed in the post simulation stage of chip design, and the test time is very long, a power prediction method based on the combination of Sparrow Search Algorithm (SSA) and Random Forest (RF) is proposed. Firstly, the unit library after 14 nm SRAM is characterized to find out the appropriate feature variables and obtain the feature data to build the training test set. Secondly, the characteristic variables are analyzed by the characteristic importance, and sorted according to the characteristic importance. Finally, the random forest model is used for regression prediction, and the sparrow search algorithm is introduced to find the model parameters with the smallest root mean square error. Compared with linear regression model, support vector regression model and other models, SSA-RF has higher convergence accuracy, faster training speed. The R2 value of SSA-RF model is about 0.97. In addition, in the case of less data, the R2 value can also reach about 0.95. A better prediction model is constructed, which provides a feasible scheme for reducing the power consumption test time, and can leave more time for designers to optimize the circuit.
- 1Improved fusion method based on ambient illumination condition for multispectral pedestrian detection
- 2Research of Phoneme Recognition Based on Recurrent Neural Network
- 3Semantic segmentation algorithm based on separable dilated convolution and joint normalization method
- 4Design of RISC-V processor based on Chisel
- 5A new stage for microsystem integration-the integrated development of integrated circuits chips and system-level electronic packaging
- 6An overview of SRAM in-memory computing
- 7Restoration of minimum cascade mobile for wireless sensor networks
- 8Improved multi-scale edge detection method based on HED
- 9A Method of Ellipse Fitting Based on Total Least Squares
- 10The Apply of LOD Effects and WPE Effect in Nanometer Process PDK
- 1An improved particle swarm optimization algorithm for adaptive inertial weights
- 2Research Image Mosaic Algorithm Based on Improved SIFT Feature Matching
- 3Task Scheduling Algorithm Based on Load Balancing Ant Colony Optimization in Cloud Computing
- 4K-means Optimal Clustering Number Determination Method Based on Clustering Center Optimization
- 5A Method of Ellipse Fitting Based on Total Least Squares
- 6An Module Level Reusable Randomization Verification Platform Based On UVM
- 7Research of Vertical Reuse Based on UVM
- 8The Level of K-means Clustering Algorithm Based on the Minimum Spanning Tree
- 9Improved BP Neural Network Based on Simulated Annealing
- 10A review of feature selection methods