Abstract:Computer-vision-based surface defect detection of metal planar materials is a research hotspot in the field of metallurgical industry. The high standard of planar surface quality in metal manufacturing industry requires that the performance of automated visual inspection system and its algorithms is constantly improved. This paper attempts to present a comprehensive survey on surface defect detection technologies based on two-dimensional and three-dimensional machine vision by reviewing over 110 publications for some typical metal planar materials products of steel-, aluminium-, copper-plate and strips. According to the nature of algorithms as well as image features, the existing 2-D methodologies are categorized into four groups: statistical, spectral, model-based and machine learning. The 3-D defect detection technologies are divided into photometric stereo, laser scanner and structured light measurement methods on the basis of the way of 3-D data acquisition. These classical algorithms and emerging methods are introduced, analyzed and compared in this review. Finally, the remaining challenges and future research trends of visual defect detection are discussed and forecasted in an abstract level.

Abstract:Aiming at the problems of excessive cycles and latitude disasters in the traditional obstacle hazard division, the system uses a symmetric FIR filter. In this way, using symmetry to optimize the filter coefficients only requires N/ 2 multiplications and N additions. It can generate adaptive obstacle data and shorten the response time of the filter; in the problem of dimensional disaster, the system uses principal component analysis to achieve dimensionality reduction, considering the danger degree, distance, movement state of the obstacle, etc. In the case of a feature parameter, the two features are unified into one data set for processing, so that while improving the flexibility of the classification hyperplane, it also shortens the time required for the classification of obstacles. The result proves that the calculation time of the algorithm of the system is shortened to 0. 25 s, and the accuracy of the recognition of high-risk obstacles reaches 96. 80%.

Abstract:In order to improve the robustness of simultaneous localization and mapping (SLAM) in different indoor scenes and deal with the challenges in extreme environments such as less texture and poor light. The visual odometry based on ORB algorithm is improved by using the improved fast automatic color enhancement (ACE) image enhancement technique. At the same time, the original image data, the image data enhanced by contrast limited adaptive histogram equalization (CLAHE), single scale retinex ( SSR), and the improved fast ACE were applied to different real scenes, such as stairwells, subterranean parking lots, and two comparison experiments based on image quality and feature extraction matching are done. The experimental results show that the quality of the image enhanced by the improved fast ACE is better than the other algorithms. After enhancement, the number of feature points of visual odometry (VO) is increased by a multiple order, the matching number is increased by 7% ~ 25% in extreme environments, and the robustness is improved.

Abstract:To improve the speed and accuracy of salient object detection in complex scenes, a two-stage salient object detection method based on deep reinforcement learning is proposed. It is composed of the salient region location network (SRLN) and the salient object segmentation network (SOSN), corresponding to the salient region location stage and the salient object segmentation stage respectively. In the salient region location stage, deep reinforcement learning (DRL) is used to train an agent to locate to the salient region gradually through sequence actions. To the best of our knowledge, it is the first to use DRL-based method for salient object detection. Then the fine salient object segmentation result can be obtained via a simple segmentation network. To simplify the network structure and reduce the number of parameters, SRLN and SOSN share feature extraction network. At the same time, a divide-and-conquer training strategy is proposed for the two-stage salient object detection framework. Experimental results on public datasets show that no matter for simple or complex scene images, the proposed algorithm can eliminate the disturbing information effectively and rapidly, achieving accurate and real-time performance. The extended experiments on pedestrian detection dataset show the generalization ability of our method in other application problems.

Abstract:In pharmaceutical encapsulation production, the existing positioning method of Penicillin bottle mouth are easily affected by the interference of bottle mouth edge. It leads to inaccurate positioning of the bottle mouth circle center. In this paper, a multi bottle mouth positioning algorithm based on DBSCAN random circle detection is proposed. Firstly, the canny edge detection algorithm is used to get all the contours in the image. The density based on DBSCAN clustering algorithm is used to segment the interested bottle mouth edge sets. Then, for each individual bottle mouth edge image, the least square method and radial scanning are used to obtain the outer edge points of the bottle mouth. Then, a large number of candidate center sets are obtained by repeated random circle detection. Finally, the truth is obtained by clustering based on DBSCAN algorithm. The mean center of the real circle center set is the center of the bottle mouth. Compared with four typical algorithms, the experimental results show that the average positioning error of the proposed circle positioning algorithm is 0. 553 pixels, which is better than other algorithms. And the average execution speed of the algorithm is 1. 359 ms. The algorithm meets the requirements of accuracy and real-time of the pharmaceutical potting production line.

Abstract:Aiming at the problem of multi-person wearing masks in the classroom and gesture recognition in COVID-19, this paper presents a multi-person state detection algorithm, based on the YOLO and OpenPose models. The Efficient-YOLO model proposed in this paper uses the classical CBAM attention and SPNET-NEW modules to deal with the problems of multi-person occlusion and irregular targets. In addition, this paper presents a lightweight Class-OpenPose model to detect the students’ posture. Based on the OpenPose model, our proposed algorithm uses ShuffleNetV2-NEW to improve the traditional model in terms of low-level feature extraction, and extracts correct key posture points in complex environments and in real-time. Experiments show that in the multi-person abnormal event detection task, the average accuracy of the Class-OpenPose model is 79. 0% that is higher than that of the traditional model, and the detection speed reaches 13. 5 F/ s; the Efficient-YOLO mask recognition model achieves an average accuracy of 83. 1%, and the detection time is only 31. 54 ms, which provides a good algorithm idea for classroom student status detection.

Abstract:With the development of deep learning in recent years, remote sensing image fusion methods based on convolutional neural network were proposed and presented with good performance. Because there is no high-resolution multispectral image as a reference, the convolutional neural network is trained in the degraded images. The trained network is used to predict high resolution multispectral images. However, the fusion process of degraded images cannot reflect the fusion process of original images. In order to improve fusion performance, a semi-supervised fusion method based on convolutional neural network is proposed. The same fusion network is trained in the degraded image and the original image simultaneously. Because degraded image fusion has the corresponding reference image, the supervised learning method is used to train the fusion network. Moreover, the spectral loss is added to preserve the spectral information. However, there is no high-resolution multispectral reference image in the original image fusion. Spectral degradation network and spatial degradation network are designed to train the fusion network. The experimental results show that the proposed method is better than the compared method in preserving the spectral and details.

Abstract:In order to overcome the block phenomenon and spectral distortion in the current remote sensing image fusion methods, by introducing the nonsampling contourlet transform, this paper designs a remote sensing image fusion method based on the restriction of energy similarity. HSV transform was applied to multispectral image for extracting its brightness factor from MS image. The different image coefficients of V factor and PAN image were calculated by NSCT transform. The regional energy function was used to measure the image energy information, and the fusion model of low-frequency coefficients was constructed to obtain the fusion low-frequency coefficients with better spectral characteristics. The spatial frequency function was used to measure the detail features of the image, so as to build the fusion model of high-frequency coefficients to obtain the fusion high-frequency coefficients with better definition. Finally, the fusion results were obtained under the inverse transform of HSV and NSCT. The experimental results show that compared with the existing fusion techniques, this algorithm has better fusion effect, and the output image has better spectral characteristics, which effectively reduces the blocking effect.

Abstract:In response to the actual needs of regular maintenance and maintenance of underwater pipe networks, a set of underwater pipeline network visual inspection system is designed. A pipeline tracking method is proposed according to the characteristics of underwater optical images and the characteristics of pipe network laying. The method uses adaptive histogram equalization to convert the unevenly illuminated underwater image into a uniform light image, and combines the color space transformation method to enhance the image and complete the pipeline region segmentation. The pattern recognition method of the support vector machine is used to classify the pipeline image according to the underwater robot tracking strategy, and the pipeline extraction strategy is established according to the classification characteristics of each pipeline. Finally, Kalman filtering is used to track pipelines to ensure system stability. The pipeline tracking detection experiment was carried out in the experimental pool, and the correct rate of pipeline network path identification reached 93. 7%, which verified the effectiveness and stability of the visual inspection system and can meet the actual needs of underwater pipeline network inspection.

Abstract:In order to solve the problems of low efficiency and poor accuracy in manual detection of water meter movement sensitivity, a set of water meter movement sensitivity detection system based on machine vision was developed. Designing an indirect detection algorithm based on the difference in the rotation angle of the pointer before and after ventilation. Using the least square method to find the center of the initial contour, dividing the pointer contour into four parts with the abscissa and ordinate of the center, calculating the standard deviation of the distance from the contour point of each part to the center of the circle and taking the smallest part of the standard deviation as the contour set to fit draw a precise circle, realizing the pointer center positioning, combining the Shi-Tomas corner detection algorithm and the distance feature from the needle point to the center of the circle to realize the needle point positioning; calculating the angle between the straight line and the horizontal line formed by the center of each sub-dial pointer and the needle point before and after ventilation, comparing the pulse number converted from the angle difference with the set pulse number threshold, and judging whether it is qualified. Experiments show that the system can efficiently complete the sensitivity detection of water meter movement while maintaining the accuracy of detection, and the accuracy of algorithm verification is as high as 99. 7%.

Abstract:YOLO algorithm loses part of the effective information of large-size feature maps in the downsampling process, which leads to the problem of insufficient target location in the detection task, which affects the overall detection accuracy of the model. This paper proposes the use of multi-scale feature fusion to solve the problem of inaccurate location of YOLO; First, modify the network model of YOLO algorithm, use different size feature maps in the YOLO network model with different feature attributes, and merge different size feature maps to improve the location accuracy of the detection network to the target; second, based on the pre-training model Re-train the modified network model on the last; finally, test the trained model in the computer. Experimental results show that the YOLO target detection algorithm based on multi-scale features improves the Accuracy rate by 3. 02% and improves mAP by 1. 53% compared with YOLO target detection algorithm.

Abstract:Compared with ordinary images, high-resolution remote sensing images have the characteristics of diverse directions and large scale changes. Aiming at the problem of remote sensing image object detection, this paper proposes an R-CenterNet remote sensing image object detection algorithm. First, redesign the CenterNet network and add a rotation factor to the network structure to provide angle information for the detection frame; secondly, increase the network depth and improve the network detection performance; finally, to aggregate the information of different regions, further extract the multi-scale information of the object. This paper proposes an attention pyramid pooling module that combines the object feature attention information with multi-scale pooling information. The experimental results show that R-CenterNet has a better detection effect, and the mAP value is increased by 8% compared with the original CenterNet detection results.

Abstract:A camera of all-weather vehicle driver monitoring system is designed to detect the driver′s abnormal driving behavior and improve driving safety. The RGB-IR CMOS image sensor, dual band-pass filter lens, hardware circuit, and active infrared lighting technology of the camera are studied in this work. Firstly, infrared enhancement characteristics of the RGB-IR CMOS image sensor are analyzed. Secondly, parameters of the lens are chosen to design according to the application scene and the suitable CMOS image sensor. The optical structure of the lens is optimized and a dual band-pass filter is also selected. Thirdly, the overall design of the camera hardware circuit is completed. Finally, the imaging performance of the camera and the effectiveness of the algorithm are evaluated. The experimental results show that the camera can achieve daytime high-definition color imaging and high-definition infrared imaging at night. The camera can meet the standards of the vehicle specification level and the image can be transmitted over a long distance of 10~ 15 m. In addition, the camera also has the advantages of low cost and high stability.

Abstract:The real-time tracking of badminton in 3d space is realized by using binocular vision system in this paper. This paper mainly studies binocular images acquisition, binocular calibration, image correction, stereo matching and so on, to identify the badminton in sport as the target direction and realize the real-time tracking and positioning of the badminton in sport. In order to increase the stability of the system, the background subtraction algorithm and Camshift algorithm are combined to achieve the tracking and monitoring of badminton in a relatively complex background, the position information of the target badminton is obtained through binocular, and the position of badminton is fed back to the robot motion control system in real time. Through the constant testing, the tracking detection and the calculation of three-dimensional information of the flight path of badminton are realized, and the positioning error is within the range of ±10 mm, and the real-time tracking and positioning of badminton is successfully realized.

Abstract:In order to solve the existing problems that the hand gestures recognition is easily to be interfered by background noise and the algorithm is complex, a digital gesture semantic recognition method based on 3D vision is proposed. First of all, RGB and depth images of hand area were collected by Realsense 3D camera, and segmentation results of hand gesture were obtained by combining depth information and skin color information. Secondly, after morphological filtering of gesture images, the feature parameters of gesture region such as area ratio of contour to convex hull, number of convex defects, angle between fingers and the length ratio of key points connection were obtained. Finally, analyzed the unique characteristic parameters of different gestures to achieve accurate gesture recognition. The digital gesture recognition experiments of 0- 9 were carried out 50 times, the accuracy of gesture segmentation was 100%, and the accuracy of gesture recognition was 98. 5%. The experiments show that this method is accurate and reliable, and the effect of digital gesture recognition is ideal.

Abstract:Aiming at the problem that the traditional Camshift algorithm is susceptible to the sudden acceleration and deceleration of the target and the background or the target interference of the similar colors when the intelligent vehicle is visually tracking the pilot vehicle in front, a pilot vehicle tracking algorithm that combines the improved Camshift and Kalman filter was proposed. The algorithm tracked the back projection of the three-dimensional histogram established by the target template hue, saturation, and edge gradient amplitude feature components. The Bhattachayya coefficient was used as the basis for determining the accuracy of target tracking. If the coefficient was greater than the set threshold, the target tracking would be judged to be inaccurate. At this time, the LBP cascade classifier was used to detect and recognize the pilot vehicle, and finally the Kalman filter was introduced to predict the position of the pilot vehicle in the next frame. The experimental results demonstrate that the proposed algorithm can accurately track the pilot vehicle in real time in a complex background.

Abstract:The complexity of traditional zero-forcing (ZF) precoding increases due to large-scale antenna technology. We propose a low complexity precoding technique. Firstly, the complexity of the matrix inversion problem is solved by using semi-iteration symmetric successive overrelaxation method (SSOR) technique, then the convergence rate of SSOR technique is accelerated by Chebyshev semiiterative ( SI) method, which makes the zero forcing precoding technology converge quickly. The experimental results show that for Chebyshev semi-iterative accelerated SSOR (SI-SSOR) precoding technique, it can approximate the 95% performance of ZF precoding with 2 iterations, and it takes less computing resources than ZF precoding. Therefore, The SI-SSOR precoding can play one significant technique to resist the influence of the communication interference.

Abstract:A background calibration algorithm based on ref. ADC is proposed for the mismatch in time interleaved ADC (TIADC). The difference between the output of the ref. ADC and sub-ADC at the same sampling time was used to estimate the mismatch, then subtracts the mismatch from the system output to realize adaptive calibration. Furthermore, randomization technology is adopted to solve the problem of the interference between the ref. ADC and TI-ADC, coupled through the input network when there is no separate input buffer at the front end of TI-ADC system by reducing residual interleaving spurs. The proposed algorithm can achieve effective calibration of the three main mismatch errors at the same time, and there is no limitation on the input signal bandwidth. Applied to a 12-bit 1 GS / s TIADC, when the input signal frequency is 470 MHz, FPGA verification results show that after calibration, the spurious-free dynamic range (SFDR) increase by 44. 14 dB to 76. 16 dB.

Abstract:Aiming at the yaw problem of the flat tire vehicle, the model predictive control (MPC) based front wheel steering controller is proposed to correct the vehicle yaw and ensure the vehicle to run on the safe path. Based on the model predictive control, the simplified double track vehicle model is selected, and the vehicle motion state is described by differential equation. The prediction equation is derived by linear discretization. Considering the vehicle stability factor, it is added as a constraint condition and transformed into the standard quadratic form to calculate the optimal solution. The closed-loop system is constructed for simulation. The results show that the MPC based design has good performance: The front wheel active steering controller can keep the vehicle running stably under two driving conditions of straight driving and double shifting driving, and effectively control the running track of the flat tire vehicle. The center of mass side slip angle, tire side slip angle and yaw angle speed are all within the stability requirements, so as to ensure the vehicle to drive in the safe path and achieve the trajectory control effect of the tire burst vehicle.

Abstract:An online calibration method of a six-axis force sensor for the industrial robot was proposed based on the requirement of realtime detecting of end effector working force of industrial robots. First, establish a calculation model of a sensor data relationship, and on this basis, focus on the interference caused by the sensor’s “zero drift”. Then, the least square method was used to identify the base angle,end effector weight and barycenter coordinate parameters of the robot. Finally, according to the real-time kinematic parameters of the robot, the time series method was utilized to compensate for the weight of the end effector and the zero of the sensor online. Experiments show that the online calibration method can eliminate the interference caused by the “zero drift” of the sensor compared with the static calibration method, and have a better online compensation effect. Through real-time tracking detection of the external constant and variable forces received by the end-effector of the industrial robot, the validity and practicability of the algorithm are verified.

Abstract:Image reconstruction algorithm is the key to the industrial application of electrical capacitance tomography. The Landweber iterative algorithm has achieved a good compromise in the accuracy and speed of image reconstruction, but its convergence speed is slow, and the late iteration is unstable, there is no clear iteration stop criterion. In this paper, an improved gradient accelerated Landweber iterative algorithm is proposed, based on the series theory, the gradient accelerated Landweber iterative algorithm is deeply analyzed, a new iterative formula is obtained by constructing the residual matrix and adding constraint factors, and it is applied to ECT Image reconstruction. Numerical simulation results show that the relative error and correlation coefficient of the improved algorithm can converge to a certain value after less iteration steps, and has a clear iteration stop criterion. Moreover, for most flow patterns, the reconstructed images are closer to the original flow patterns in subjective quality, which verifies the stability and effectiveness of the improved algorithm.

Abstract:To realize the rapid identification and analysis of aged wheat with different storage years, this paper puts forward one method to identify the age of wheat by employing a voltammetric electronic tongue (VE-Tongue) combined convolutional neural networks (CNN) and wasserstein generative adversarial networks (WGAN). VE-Tongue was used to detect six kinds of wheat with different storage time, then the corresponding electronic tongue signals were acquired. An automatic feature extraction and classification recognition model of electronic tongue signal based on CNN structure was designed to solve the problems of large amount of information and difficult feature extraction of sensor array response signals. To solve the problem of insufficient training samples, WGAN was utilized to generate electronic tongue signals to improve the generalization ability of CNN model and avoid over-fitting problem. The results showed that the proposed method exhibited better classification performance compared with the deep learning models like AlexNet, VGG16 and the traditional machine learning models like random forest (RM) and extreme learning machine (ELM). The test set accuracy, precision, recall rate and F1-Score of the proposed model reached 0. 98, 0. 98, 0. 977 and 0. 988 respectively. This study found that the VETongue combined with CNN and WGAN could be a sensitive, reliable and effective detection method for identifying the amount of storage year of wheat, it can provide a new research way of thinking for the sensory recognition technology based on artificial intelligence.

Abstract:This research designs a wireless measurement system using flexible piezoelectric composite film sensor. The system mainly includes piezoelectric film sensor, interface circuit module, AD acquisition module, data processing module and wireless transmission module. The piezoelectric film sensor converts structural vibration information into Piezoelectric signal, the system amplifies, collects, wirelessly transmits and receives the piezoelectric signals. A new method to invert the structural strain by collecting voltage is proposed, The transfer function of piezoelectric signal is derived, and the relationship between acquisition voltage and strain is established, therefore, dynamic strain measurement can be achieved since the frequency and the strain amplitude of the structural vibration can be calculated inversely based on the collected piezoelectric signal. Experimental results indicate that the system can accurately measure the vibration frequency and strain of the structure, and the distance for wireless strain measurement by this system can reach 40 meters, the proposed wireless strain measurement system can partially replace the conventional wired strain measurement system so that it has great potential in engineering applications.

Abstract:Aiming at optimal sensor placement in vibration testing of large structures, an improved particle swarm optimization was adopted to optimize placement of acceleration sensors in the vibration responses test of tracked semi-vehicles. First, a particle swarm optimization with nonlinear dynamic adjustment of inertial weight collaborative learning factor ( PSO-IWLF) was proposed,and it was used as a subsequent optimal sensor placement algorithm. Then, the finite element modal analysis was performed on tracked semivehicles to obtain different orders’ mode shape of the node. Finally, according to the modal assurance criterion ( MAC), optimal calculation of acceleration sensor placement in the vibration test of tracked semi-vehicles was carried out, and the validity of the optimization results was verified through the vibration response test and analysis. The optimization and experimental results show that compared with the standard PSO, the optimization accuracy of PSO-IWLF is improved by 17. 5%, and the selection of 2, 3, 4, 5, 9 and 10 among the 10 measuring points in the primary selection is more reasonable.

Abstract:Aiming at the current difficulty in nondestructive testing of pipeline gas pressure, combined with the principle of ultrasonic reflection pressure measurement, a deep belief network (DBN) extraction of ultrasonic echo amplitude characteristics is proposed, which is least square support vector machine ( LSSVM) pipeline gas pressure detection method. First, the features are extracted through unsupervised layer-by-layer learning of the restricted Boltzmann machine (RBM) in the DBN network. Secondly, the supervised error back propagation adjustment is performed through the label layer to optimize the RBM parameters of each layer of the DBN. Finally, input the characteristic signal extracted by the optimized DBN network into the trained LSSVM to complete the gas pressure recognition. Design related experiments to obtain ultrasonic data for model testing. The results show that the average relative error of pressure recognition of the DBN-LSSVM pressure recognition model proposed in this paper is 0. 635 7%, which is lower than the average relative error of the DBN-BP model (1. 802 6%), which is better complete the pressure detection of the pipeline gas.

Abstract:Aiming at the premise that the current mainstream modeling method is based on the reliable test results, it is difficult to obtain an accurate information description problem between fault and test. A new test-based modeling method based on probabilistic generalized stochastic Petri net is proposed. This paper first compares the mainstream modeling method with the Petri net modeling method, clarifies the problems existing in the mainstream modeling method, and the reasons for selecting the generalized stochastic Petri net model; then introduces the probability theory into the generalized stochastic Petri net model, using the conditional probability of the model nodes obtained by the Bayesian network, and it is also the first combination of Bayesian network and generalized stochastic Petri net. Finally, a missile engine system is modeled to obtain the probability correlation matrix of faults and tests, and the test indicators are tested. The feasibility and accuracy of the proposed method are verified by comparing the original correlation matrix with the probability correlation matrix obtained in this paper.