Abstract:To prevent accidents caused by crane arms touching transmission lines, a transmission line antiexternal breakage system based on binocular vision is designed. First, by installing the binocular vision antiexternal breakage device on the crane arm to obtain image data in real time; then, an improved SGBM stereo matching algorithm is proposed for the light interference, weak texture area existing in the actual environment and edge hole filling in the parallax map, and the algorithm is optimized by using least squares fitting interpolation and bilateral filtering, so as to accurately acquire the threedimensional information of the transmission line; finally, 10 kV transmission line test environment and ranging environment are built according to the principle of binocular visual and space coupling capacitance voltage division and tested in the field. The results show that the average relative error of the improved SGBM algorithm is only 1345% and 1229% under the interference environment, which is 2047% and 1851% less than the traditional SGBM algorithm and the field strength ranging method, respectively, and the algorithm running time is 817552 ms. The ranging system takes into account the needs of realtime and precision, and meets the use of the transmission line crane against external breakage in the actual environment.

Abstract:With the rapid development of social economy, the proportion of heavy metal pollution in soil pollution is increasing, posing a huge threat to the ecological environment and people′s life and health. Aiming at the above problems, this paper proposes a soil heavy metal pollution prediction model based on the improved mixing strategy, that is, the optimal subset of characteristics is selected by random forest, and then the LightGBM parameters are optimized by random search, and finally the Nemero comprehensive pollution index of the soil is predicted by the trained LightGBM model, so as to obtain the soil heavy metal pollution status. A certain area of the North China Plain in China is used as the research area, and the prediction results of RSLightGBM, LightGBM and SVR models are compared. The results show that the mean squared error and mean absolute error of the proposed model are reduced by 6909% and 3909% respectively compared with the LightGBM model. The coefficient of determination is 611% higher than the LightGBM model. The above results show that the proposed model can be effectively applied to the prediction of soil heavy metal pollution.

Abstract:The research of specific emitter identification based on deep learning mainly focuses on the improvement of recognition accuracy, but often ignores the threat of adversarial samples in the recognition process. To solve the above problems, the experiment not only increases the category of emitter and improves the accuracy of model recognition, but also analyzes the impact of adversarial samples on deep learning recognition network with high recognition rate. In experiments, small samples of ADSB signals were obtained, and the data were sliced randomly. Then fine tune the original network and add convolutional attention module to improve the recognition rate of the model. Finally, generate adversarial samples were created by using four adversarial attack algorithms and tested on the network which was trained in advance. Additionally, images of signal examples before and after the attack were compared to maintain a balance between the attack success rate and the attack concealment. The results show that the model with high recognition rate is also vulnerable to adversarial samples, the momentum iteration method has the best performance among four algorithms, and the attack performance of momentum iteration method is more than 10% higher than the fast gradient sign method.

Abstract:In order to make better use of the relevant features in EEG signals and improve the classification performance of motor imagery EEG, a multilayer convolutional network (MTACNet) based on mixed features and parallel multiscale TCN modules was constructed. First, build a multilayer convolutional neural network based on mixed features, and embed an efficient channel attention mechanism in it, and select PReLU as the activation function to extract the temporal and spatial information in the EEG signal; then improve the TCN module, build a parallel multiscale timedomain feature extraction module, connect to a multilayer convolutional network, and further mine feature information at different time scales. Tested on the public dataset BCI_IV_2a and the selfcollected dataset SCU_MI_EEG, the average classification accuracy rates are 8615%, 7710%, and the standard deviations are 917%, 1358%, respectively. And for the selfcollected data set, a preprocessing method was designed to fuse multifrequency domain EEG signals for threechannel input. After preprocessing, the average classification accuracy rate increased by 329%. The experimental results show that: Compared with other methods, the classification network constructed in this paper has achieved relatively good classification results, and the designed preprocessing method can reduce the impact of complex environments and irrelevant interference factors on the classification results.

Abstract:Aiming at the detection problem of hovering UAV in complex clutter environment, an improved Kalmus filterresidual echo timedomain mean cancelationadaptive CFAR joint processing algorithm is proposed to detect microDoppler of UAV and realize the purpose of air traffic control monitoring. The improved Kalmus filter is used for frequency domain filtering, and the high frequency signal and zero frequency signal of target echo are suppressed at the same time, and the micro Doppler signal gain near zero frequency is improved. The residual echo mean cancellation was used for secondary filtering to improve the signal to noise ratio of Doppler characteristic signals of the UAV highspeed rotor. The shorttime Fourier algorithm was used to detect Doppler changes in the target region. Finally, the constant false alarm processing was used to further suppress clutter and extract microDoppler information. The experimental results show that the proposed algorithm can effectively detect the rotor Doppler characteristics of hovering UAV, and the amplitude of the target Doppler signal is increased by about 20 dB to achieve the purpose of low altitude monitoring and control.

Abstract:In the process of location and tracking of unmanned bus, the sampling signal is affected by noise variance, bandwidth and sampling rate, which is prone to signal loss or discontinuity. In addition, the related filtering algorithm lacks asynchronous sampling and smoothing capabilities, leading to location failure. In order to improve the positioning accuracy and supplement missing data, this paper proposes an improved sensor asynchronous sampling fusion smoothing algorithm based on asynchronous extended Kalman filtering and non causal filtering smoothing. First, asynchronous extended Kalman filter is used to exponentially discretize the continuous time stochastic differential equation to process the measured value at any time. After the state value at the next time is predicted and updated, the non causal filter is introduced to smooth the given available initial variance information, so that the noise variance impact is smaller and the estimation performance is better. The algorithm is verified by physical experiments on an unmanned bus. The results show that this multisensor asynchronous fusion smoothing algorithm has a good effect in vehicle driving. Compared with the results of asynchronous Kalman filtering algorithm, it can achieve a positioning accuracy better than 05 m. The data prediction error is significantly reduced, and the positioning accuracy is improved and missing data is supplemented.

Abstract:The core link in the construction of smart ecological forestry is the monitoring and prevention of forest fires. In order to put out the fire source to prevent the spread of fire and eliminate the hidden danger of forest fires in the first time, two forest fire detection models, YOLO_MC and YOLO_MCLite, are proposed for UAV aerial inspection. Among them, YOLO_MC can detect open flame and smoke in standard images. Based on the lightweight of YOLO_MC model, a detection model YOLO_MCLite suitable for high temperature region in thermal images is proposed. In the design of the network structure, the Transformer model is firstly integrated into the conventional convolutional neural network, which improves the perception ability of the backbone network for global feature information; at the same time, lightweight design of the Transformer model is performed. First, reduce the number of tokens in the form of group computation on the network structure to reduce the amount of tokens. Second, deredundant and weighted the number of channels in feature blocks through channel attention mechanism to reduce the dimension parameters of tokens to reduce the computational complexity. And the distillation algorithm is also used to extract the ultralightweight network from the designed network and apply it to the detection of forest high temperature points in the infrared image of the UAV to prevent the occurrence of forest fires. After experiments, the following data are obtained: the detection accuracy of the designed detection model for open flame and smoke can reach 948%, and the detection accuracy for high temperature points can reach 972%. And the test frame rate on the NVIDIA JETSON TX2 embedded device reached 225 and 324 respectively. The experimental results show that the network designed in this paper can effectively detect forest fires and prevent fires in time by detecting high temperature points.

Abstract:Aiming at the technical difficulties in the quality inspection process of the textile industry for the defect detection of complexly textured fabrics, an image detection model based on a deep convolutional neural network is proposed. Firstly, the YOLOv7tiny model was selected as the reference frame of the algorithm, and then the optimization was carried out, including using the SimAM module to reconstruct the feature fusion layer so as to improve the model′s ability to extract local features of defects and suppress background features. SIoU was used to optimize the coordinate positioning loss function to speed up the regression efficiency of bounding boxes. The FReLU activation function is introduced to enhance the utilization of spatial information in the nonlinear activation layer and improve the spatial sensitivity of the activation function. The experimental results show that the accuracy and recall ratio of this model are better than those of other existing algorithms in the detection tasks of five typical defects for complex texture fabrics. The mAP reaches the maximum value of 805%, the size of the model is only 92 M, and the detection of a single frame image on the PC is only 2113 ms.

Abstract:In order to improve the detection and recognition of small targets and obscured targets in infrared images, the Efficientnetbased infrared target detection algorithm is proposed for the problem of low accuracy and low recall of infrared target detection in complex scenes. First, the efficient and lightweight Efficientnet is used as the feature extraction backbone of the model to reduce the number of parameters of the model and improve the training speed. In the last output layer of Efficientnet backbone, SPP module is introduced to enrich the expression capability of feature map, perform multiscale fusion and expand the perceptual field of feature map; in the feature fusion part of the model, FPN feature pyramid network is used, and CSPNet module and ECA attention mechanism are added after feature fusion to enhance feature extraction. The detection part uses YOLO Head to classify and regress the targets, and uses CIoU Loss as the bounding box regression loss function to improve the recognition ability of the obscured targets. The experimental results show that the Efficientnetbased model is only 188% of the size of YOLOv3, and the mAP reaches 8074% on the FLIR dataset, which is 1012% better than the YOLOv3 algorithm, and the model improves the detection accuracy while reducing the number of model parameters. The model has good generalization ability on the FLIR dataset and improves the detection of small and occluded targets.

Abstract:Under complex background, it is difficult to extract the contour of automobile plate spring. Therefore, an improved Kmeans background segmentation algorithm based on sparrow search optimization is proposed, and the feature points to be measured are extracted by beam laser. Firstly, by traversing the global pixels, the optimal direction is determined according to the gradient threshold, and the step size moving to the optimal direction is reduced. So as to improve the sparrow search optimization algorithm, which can overcome the problem that the algorithm has weak global search ability and is easy to fall into the local optimal. Secondly, the pixel points of interest searched by sparrows were taken as the initial center point of Kmeans algorithm, and the pixels with similar characteristics were grouped into one group, so that the spring could be separated from the complex background environment and the contour of the spring could be obtained. Finally, ray laser is applied to the surface of the spring for auxiliary marking, and the feature points to be measured are extracted by intersecting the contour of the spring. The results show that the proposed detection method of plate spring size based on background segmentation can extract the feature points, and the accuracy can reach 025 mm, forming online measurement data, which is conducive to improving the production process.

Abstract:This paper studies a human motion analysis system for limb status assessment and motion posture correction. Firstly, to address the problems such as occlusion that are prone to occur during human motion, this paper introduces deformable attention and generative adversarial networks based on Transformer for optimal human key point location detection. Secondly, using the proposed algorithm, this paper designs a motion analysis system by combining the limb space constraint relationship of human posture and knowledge related to body posture analysis. Finally, through testing on public datasets and in real scenarios, this paper evaluates the feasibility of the proposed algorithm and system from both qualitative and quantitative perspectives in experiments. The experimental results prove that the detection accuracy of the algorithm in this paper can reach up to 937% on public datasets; in the tests on real scenes, the algorithm and the motion analysis system designed in this paper can effectively solve the common problems such as occlusion in human posture recognition, and show the multidimensional analysis results of human motion posture through the visualization system.

Abstract:The traditional ORB-RANSAC algorithm has the problems of high error matching rate and poor stability of similar feature points, which can not meet the technological requirements of welding handle based on the direction of bottom label in aluminum pot production. An improved ORBRANSAC algorithm is proposed. Firstly, ORB algorithm was used to extract and match the feature points, and hamming distance threshold method was used to remove the matching points. Secondly, when RANSAC algorithm performs fine elimination of matching point pairs, kfold crossvalidation is added to achieve consistency prejudgment of the initial model. Finally, the classified point pairs in the previous iteration are removed during each iteration, and the sampling space is dynamically updated. The results show that the repeatability accuracy of the improved ORBRANSAC algorithm is improved by 6604% compared with the original algorithm in the interference environment, and the single frame calculation time is reduced by 613%. The Angle measurement error based on the improved ORBRANSAC algorithm is 0201° and the average detection time is 0255 s in the multitype pot bottom label measurement experiment, which meets the requirements of automatic production measurement accuracy and realtime.

Abstract:The accurate prediction of greenhouse humidity is of great significance to the formulation of disease control strategies and automatic irrigation of water and fertilizer. In this paper, a prediction method based on multimodal data driven for full or Chinese names is studied. In order to decouple the complex relationship of environmental variables in greenhouse environmental control and improve the prediction efficiency of the model, this paper uses LASSO regression to screen the strongly related environmental factors of greenhouse air humidity changes from multiple greenhouse environmental parameters. Combining the advantages of CNN in extracting image spatial characteristics, based on GAF theory, the greenhouse time series are converted into two dimensional images of Gram angle summation field and Gram angle difference field, further enhancing effective information and suppressing environmental noise, The low complexity double convolution layer is introduced to fully extract the potential features of the image, identify the humidity change trend, and construct for the time series of different humidity change trends one by one Bayesian_ LSTM prediction model, increase smooth input to improve prediction accuracy. In this paper, the historical time series of indoor temperature, humidity and light intensity are converted into twodimensional images as input for cucumber greenhouse, and the prediction performance of the model is analyzed and verified. The experimental data shows that when the time sliding window size is 15, Gram angular difference field, Bayesian_ When the number of LSTM hidden nodes is 100, the average absolute error, average absolute percentage error, and root mean square error reach 258%, 456%, and 480% respectively, which is the best performance of the model. Compared with four mainstream prediction models, RNN, GRU, BiGRU and 1DCNN, the test results show good prediction performance.

Abstract:The wearable autonomous positioning system composed by array inertial devices can significantly improve the positioning accuracy of the wearer, but the array inertial devices in the wearable autonomous positioning system are difficult to avoid failure in the process of operation. To address the phenomenon of array accelerometer noise increasing fault in the autonomous positioning system worn by emergency rescue personnel, a Convolutional Neural Networks (CNN) based array accelerometer fault detection method is proposed, using the Generalized Likelihood Ratio (GLR) test to compare the array gyroscope with the array accelerometer, The GLR test is used to compare the array gyroscope control data, and then the CNN calculates the mapping result between the accelerometer data and the gyroscope control data to achieve fast detection of the array accelerometer growth increase fault. Through the twelve IMU array data fusion and fault detection test results show that the detection method can quickly and effectively detect the typical fault of accelerometer noise increase in the array inertial device, the fault detection rate ≥ 98%, the effect is obvious.

Abstract:In view of the difficulties in labeling the data samples of oil immersed transformers, the small amount of labeled samples and the low accuracy of traditional fault diagnosis methods, a twolayer fault diagnosis model for oil immersed transformers with few labels based on GBDT and Kmeans gain clustering is proposed. Firstly, a stacked autoencoder is used to reduce the dimension of the highdimensional characteristic gas characterizing the transformer state, remove redundant information, and obtain the lowdimensional characteristic vector containing the transformer operating state as the input of the subsequent classifier. Secondly, a twolayer fault diagnosis model is constructed; For unlabeled samples, the GBDT method is introduced as the first layer of the proposed model to obtain the false labels of unlabeled samples. In order to further improve the diagnosis accuracy, the Kmeans clustering gain based on the false label of unlabeled samples is proposed as a new feature vector, which is input into the end layer model Kmeans to realize the secondary diagnosis. Experimental analysis shows that the proposed method can effectively improve the accuracy of transformer fault diagnosis under the condition of few tags, and the diagnosis accuracy is improved by at least 6% compared with other methods. It provides a new idea for fault diagnosis of oil immersed transformer with few labels.

Abstract:In the field application of ground penetrating radar, multiple waves are a common interference wave, which will affect the authenticity and reliability of radar data and bring difficulties to target interpretation. The Radon transform is widely used in the field of seismic processing for multiple suppression, but it is still not much studied and applied in the field of ground penetrating radar. This paper analyzes and expounds the Radon transform theory for the reflected wave of hyperbolic ground penetrating radar. In the case where the position of the weak target echo and the multiple wave of the strong target are highly overlapping, and the multiple wave intensity is close to the strength of the weak target echo, analyze the difference between the position and energy intensity of weak targets and multiples in the Radon domain, recover the multiple energy by selecting the filter window and performing polynomial fitting on it in the Radon domain, and then recover multiple times in the spacetime domain through Radon forward modeling. Then, the multiple waves are subtracted from the original radar data to suppress the multiple waves, and the radar data without false interference is obtained. The experimental results show that a better multiple suppression effect is achieved while retaining the echo strength of weak targets.

Abstract:To improve the camera calibration accuracy, this paper proposes a monocular camera calibration method based on the Modified Aquila Optimizer: The internal and external parameters of the monocular camera are calculated by Zhang Zheng You calibration method. Based on the obtained camera parameters, the average reprojection error of all corners in the calibration image is calculated and the objective function is established. The Aquila Optimizer which is improved by adaptive allocation mechanism, dynamic compensation strategy and nonlinear tide strategy is used for optimization to obtain the optimal internal parameters and distortion coefficient of camera calibration. Thus, the optimization accuracy of the camera nonlinear calibration process is improved. The experimental results show that the improved Aquila optimizer algorithm has superior optimization results on different benchmark test functions. The calibration results obtained by the camera optimization calibration method proposed in this paper are more accurate, and the reprojection error is 0006 pixels.

Abstract:Aiming at the problem that the classification and detection of colorectal polyps by common computeraided detection systems are not accurate and realtime, a YOLFCBAM model combined with spatial attention mechanism (CBAM) and improved feature fusion layer based on YOLOv4 is proposed, which can classify and detect hyperplastic polyps and adenomatous polyps in dual modal of white light and NBI endoscopic images in real time. In order to make the feature extraction of polyps more accurate, a CBAM module is integrated to the backbone of YOLOv4, so that the network feature extraction layer pays attention to more important spatial and channel information, and inhibits the downward transmission of unnecessary features. On this basis, the network structure is optimized by pruning the feature fusion layer PANet to reduce the amount of network parameters and further improve the detection speed of the model. In order to train and test the improved model, 2 988 white light and NBI endoscopic images are collected from the Affiliated Hospital of Hebei University, and are divided into training set and test set at a ratio of 9∶1. Experimental results show that our proposed YOLOFCBAM achieves a mAP of 8644%, recalls of 8962% and 8564% for identifying hyperplastic and adenomatous polyps respectively, accuracies of 9135% and 8519% for identifying hyperplastic and adenomatous polyps respectively, and a classification speed of 47 FPS on the test set, which proves that the proposed model has potential clinical application value.

Abstract:The hardware units of fixedpoint and floatingpoint calculation are designed and implemented after studying and improving the traditional CORDIC algorithm to calculate the elementary functions in image processing. Two micro rotation angles of CORDIC algorithm iteration are proposed to expand the definition domain of function calculation, and angle coding is used to reduce the number of iterations of trigonometric function calculation. Arc tangent and square root can be calculated in rotation mode, sine and cosine can be calculated in vector mode. The units of Fixed point and floating point are designed in pipeline structure, and the functions can be selected by mode configuration. The floatingpoint unit bases on the format of IEEE754 single precision floatingpoint number. The data path contains order matching, iteration and normalization, and can be calculated once in 24 cycles. The verification of SystemVerilog platform is realized and the worst accuracy of fixedpoint calculation is 10-3, and floating point is 10-7. The maximum working frequency of 32 bit fixedpoint calculation can reach 2439 MHz, which takes less resources than the traditional CORDIC algorithm when verificate on the FPGA. The improved fixedpoint CORDIC algorithm is applied to image edge detection based on Sobel, with clearer edges and faster imaging speed. FPGA platform for image data acquisition, processing and display system is built to complete the verification of the algorithm.

Abstract:At present, most of the building scene classification methods in remote sensing images use manual annotation method, which requires a lot of time. To solve this problem, this paper proposes a high resolution remote sensing building scene classification method using automatic sample selection. Firstly, the feature space of multidimensional highresolution remote sensing image with spectral features, geometric features and depth features is established. Secondly, the buildings were initially extracted by decision tree, and the scene density histogram of buildings was constructed. Then, the natural discontinuity method was used to classify the building density, and the proportion method was used to extract some scene images from each type of scene as training samples. Finally, ResNet50 network is used to classify building scenes. Taking Hunnan District of Shenyang City, Liaoning Province as the study area and Google Earth remote sensing images as the experimental data, the experimental results show that the proposed method can achieve unsupervised scene classification, and the overall classification accuracy and Kappa coefficient are 089 and 082, respectively, which are improved by 3% and 8% compared with the original sample selection method.

Abstract:To promote the automatic realization of coaxial tire type discrimination in vehicle safety inspection, a tire pattern image verification algorithm based on siamese network was proposed. The algorithm is oriented to the tire pattern images of small data sets. On the infrastructure of the siamese network, an image preprocessing module of orientation correction is added to realize the alignment of tire patterns and eliminate the obvious orientation difference between tire images. The Gabor Orientation Filters are used in the lowlevel convolutional network of its subnetwork to improve the learning speed of the network on tire pattern texture features and the robustness of tire image recognition with different quality. Experimental results on CIIP_TPID and WTP datasets show that the accuracy of the proposed algorithm is 0926 and 0849 respectively.

Abstract:Under complex background, the characteristics of buildings at different scales are quite different, and the existing algorithms have problems such as uneven segmentation and misjudgment of multiscale building segmentation. To solve the above problems, we design a new network structure adapted to scale changes. Firstly, aiming at the problem of low segmentation accuracy in remote sensing image scenes, we introduce and embed a coordinate attention mechanism in the basic network to enhance the context information capture ability, eliminate noise and enhance the network′s ability to extract spatial features. we introduce A new recursive residual convolution module ed to deepen the network layer, reduce information loss, and improve the efficiency of feature extraction. Finally, we introduce a hollow space convolutional pooled pyramid in the hop connection to increase the network receptive field, enhance the effective features, and suppress the useless features. Design the system to verify the usefulness of the model. Experimental results show that the proposed method improves the accuracy, recall, F1score, and IoU indicators by 305%, 156%, 13%, and 308% compared with the UNet network, respectively.

Abstract:Aiming at the problem of low target detection accuracy caused by numerous small targets in remote sensing images, drastic changes in target scales and complex backgrounds, a target detection algorithm based on improved YOLOX is proposed. On the basis of YOLOX, firstly, an attention mechanism is added to the backbone network to improve the network′s ability to perceive small targets in remote sensing images and enrich the semantic information; secondly, the feature fusion part is added to the MSCER multiscale information fusion module in the feature fusion part to reduce the loss of image detail information caused by scale changes in remote sensing images through feature maps of different sizes; finally, the convergence speed of the network is accelerated by introducing CIoU loss function to make it meet the demand of realtime. In this paper, the proposed detection algorithm is experimented on the RSOD remote sensing dataset, and the average detection accuracy is 9512%, which is 869% higher than that of the unimproved YOLOX. The experimental results prove that the proposed method has higher detection accuracy.

Abstract:Due to the interference of complex background in UAV inspection image and the influence of external factors such as aerial shooting Angle, insulator object recognition will bring certain difficulty. When the commonly used Faster RCNN model is used for insulator object detection under complex background, there is the problem of missing detection of small object insulators that are distant or blocked. Therefore, this paper selects ResNet101 as the backbone network on the existing Faster RCNN model. The FPN structure is introduced to improve the detection accuracy of the occluded small object insulator, reduce the missed detection rate of occluded targets, and the channel attention mechanism SENet is added to enhance the insulator characteristics. The experimental results show that the improved model based on Faster RCNN achieves an accuracy of 932% in insulator object detection under complex background, which is 64% higher than that of the baseline model AP50, and is superior to some advanced object detection models at present, with high detection accuracy for insulators under complex background, and can solve the problem of false detection and missing detection of small object insulators.