71results about How to "Improve target recognition accuracy" patented technology

The invention discloses an obstacle identification distance measuring system and a method oriented to the construction process of a road roller, The system comprises an embedded industrial control computer, an industrial camera is arranged around the top of the cockpit of the road roller, a lidar is respectively arranged on the front side and the rear side of the road roller, the industrial cameraand the lidar are connected with the industrial control computer, and communication between the industrial camera and the lidar is established through a ROS system. The system fuses the image and point cloud data to realize the obstacle recognition and measurement in the construction process of road roller. The system has the characteristics of full coverage of construction environment, high accuracy of target recognition, high accuracy of obstacle position measurement, good real-time performance and strong anti-interference ability, and is suitable for unmanned road roller construction.

The invention relates to a method and a system for automatic positioning and identifying a target based on high definition video images. The system comprises a high definition video image enhancement processing module, a target detecting and positioning module, a grayscale image binarization module, a target contour image inclination correction module, a target contour image dividing and feature extracting module, and a target contour image identifying and post-processing module. The system has the advantages that automatic target detecting and fast target positioning and identifying effect of high definition (above 2 million pixels) video images (moving) under complicated conditions is fine, high definition video image enhancement processing effect under uneven illumination conditions is fine, dynamic range is large, automatic target positioning and identifying precision are high and the like. In the meantime, reliability and stability of software is improved greatly by the design of an embedded software function, and construction mounting requirement of a front end embedded system (intelligent camera) is lowered.

The invention, which belongs to the technical field of multi-sensor information fusion, discloses an evidence-synthesis-based information-fusion target recognition method. A plurality of sensors are used for carrying out attribute information collection on a to-be-identified target and a feature attribute is extracted from the collected attribute information; data having the feature attribute aredivided into training data and testing data, wherein the training data are used for constructing a neural network model and the testing data are used for obtaining a basic probability assignment value; and then evidences are synthesized based on an improved evidence synthesis method and the synthesized result is used as the target recognition basis. According to the invention, the basic probability assignment values of evidences are obtained accurately and a synthesis problem of high-conflict evidences is solved. The basic probability assignment values of evidences are obtained by using the neural network and the neural network has the high nonlinear mapping capability and is capable of mapping the intrinsic relationship between the target feature data, so that the accuracy of the basic probability assignment values is ensured, the conformance to the real scene is realized, and the practical significance is good.

The present invention discloses a neighborhood Gaussian structure and video feature fusion motion identification method. The method comprises: a 3D LARK operator is employed to extract the local structure feature of the video, and a neighborhood structure assessment algorithm based on the multivariate Gaussian is provided for expressing the integrated structure; the neighborhood structure assessment algorithm and the 3D LARK features are subjected to the local matching and statistical process of the multi-scale template and the video to be detected to obtain the statistic probability matrixes of two object motions; and the two statistic probability matrixes are fused to extract the object, and the double restraints improve the accuracy of the object motion. The neighborhood relation restraint integration idea is provided on the traditional LARK operator, and a new motion identification model is provided. Compared to the current method, the neighborhood Gaussian structure and video feature fusion motion identification method is more accurate in extraction of the object motion, high in identification accuracy and is applied to the visible light and the infrared videos at complex scenes.

The invention discloses a DHOG and discrete cosine transform-based overhead line system equilibrium line fault detection method. The method includes an equilibrium line and equilibrium line supporting seat recognition and positioning process and an equilibrium line default fault detection process. The equilibrium line and equilibrium line supporting seat recognition and positioning process includes the following steps that: DHOG feature extraction is performed on a positive sample and a negative sample; and an AdaBoost classifier is trained; and a classifier obtained through training is utilized to accurately judge whether a target area is contained in a slide detection window. The equilibrium line default fault detection process includes the following steps that: the target feature information of the equilibrium line and a supporting seat thereof is enhanced through utilizing discrete cosine transform; a two-step method which performs circle detection based on arc piecewise fitting is adopted to perform statistics on circles in an image; and an overhead line system equilibrium line default fault can be judged according the number of the circles. The DHOG and discrete cosine transform-based overhead line system equilibrium line fault detection method has the advantages of high detection accuracy and excellent robustness under a situation in which rotation and scale variation occur on the target image, and can provide a certain technical reference for intelligent inspection of high-speed rail overhead line systems.

The invention discloses an unmanned aerial vehicle target recognition and classification method based on deep learning, and the method comprises the specific processing steps: (1), carrying out the collection through a radio monitoring means, and obtaining a link signal of a small civil unmanned aerial vehicle; (2) sending the link signal to a data processing center, and carrying out signal preprocessing to obtain a time domain characteristic and a frequency domain characteristic of the signal; (3) inputting the time domain characteristics and the frequency domain characteristics of the signalinto a deep learning network for calculation, and outputting to obtain a target identification classification result of the signal; and (4) storing the acquired signal data and the identification result in a database, and if the acquired signal data and the identification result are newly discovered target categories, adding the newly discovered target categories to a signal identification library after manual confirmation. According to the unmanned aerial vehicle target identification classification method provided by the invention, the deep learning algorithm is applied to the field of unmanned aerial vehicle radio signal monitoring, so that the application bottleneck of a traditional radio signal analysis method is broken through, and the unmanned aerial vehicle signal monitoring and identification problems in low-altitude safety protection are solved practically.

The invention relates to a two-wheel self-balancing follow service robot based on depth information and panoramic information. The two-wheel self-balancing follow service robot comprises a mobile phone client, a robot main body, a storage box cover, a touch control display screen, wheels, a side end cover, a depth camera and fisheye cameras, wherein the client is installed on a smart phone; the robot main body comprises a control module, a power module and a voice recognition module; the depth camera is positioned on the front part of the robot main body; the fisheye cameras are positioned on the front part and the two sides of the robot main body. The two-wheel self-balancing follow service robot is mainly applied to outdoor environments such as shopping malls, stations and airports, and the functions of target recognition, self-balancing, carrier following, automatic barrier avoidance, dynamic map building, smart path planning, multi-robot cooperative operation, GPS (Global Positioning System) positioning, automatic driving and the like are realized. By adopting the two-wheel self-balancing follow service robot, the depth information and the panoramic information are combined, and a followed target does not need to wear a positioning device; the GPS is embedded in the robot main body to assist the robot in navigating, positioning and implementing other functions.

The invention discloses an electrocerebral-drive high-reliability control system based on a dual-mode check mechanism. The system comprises a signal acquisition module, an impedance detection module, a parameter setting module, a signal processing module, a target check module and a human-machine interaction module, wherein the impedance detection module and the parameter setting module are used for initializing the system; the signal processing module comprises a time-domain component signal processing module and a frequency-domain component signal processing module; the target check module comprises a check code encoder and a check code decoder and is used for performing error check and correction operations on the target key preliminary judgment result deduced from the dual-mode joint characteristics of the electrocerebral time-domain component calculated by the signal processing module; the human-machine interaction module comprises a control state display module and a visual stimulation module; the control state display module is used for receiving the target key coordinates output by the target check module, and displaying the system operation state after the operation of the target key so as to form visual feedback; and the visual stimulation module is used for continuously applying stimulus and inputting the next control command. The system disclosed by the invention has the advantages of simple structure and the like, is convenient to operate and can increase the control speed as well as the control reliability and accuracy.

The invention discloses a rail transit obstacle avoidance system based on image processing and target recognition. The rail transit obstacle avoidance system comprises a data acquisition module, an artificial intelligence processing module and a display control module; a collected image is segmented through the artificial intelligence processing module, so that the train running track can be marked quickly and accurately, and the warning area can be divided more accurately; obstacle detection is carried out through a millimeter-wave radar to obtain orientation and distance information of the target; targets such as personnel, vehicles and trains in the collected images are recognized through the artificial intelligence processing module, target image coordinates are obtained and fused withradar data, whether obstacles exist or not is comprehensively judged, and an alarm is given if the obstacles are in a warning area. The video data can be uploaded to the display control module in real time through the Ethernet. Image segmentation is carried out on the rail, so that warning area division is more accurate; and through image target identification and radar detection, the target identification accuracy can be effectively improved, the false alarm rate is reduced, and the alarm accuracy is improved.

The invention belongs to the technical field of video processing, and specifically relates to a cross-camera video concentration method based on pedestrian re-identification. The specific steps are asfollows: (1) using multiple segments of obtained videos shot by different cameras as input, and finally outputting multiple segments of concentrated videos; and (2) based on the multiple segments ofconcentrated videos obtained in a video concentration phase, in view of a specific target object, firstly searching for specific locations of the target object appearing in different videos in the concentrated videos according to a corresponding matching measurement method, and extracting concentrated video segments containing the target object to obtain the description of a coherent motion trajectory of the target object in the multiple videos. By adoption of the method of the invention, a lot of manpower can not only be saved, and the identification accuracy of the target object is also improved to some extent. The method proposed by the invention has an important value for practical application.

The invention provides a method for identifying a space target, which comprises the following steps of: performing envelope alignment preprocessing on HRRP sample data, and eliminating redundancy of the HRRP sample data subjected to the envelope alignment preprocessing; performing envelope amplitude preprocessing on the HRRP sample data after redundancy elimination, and forming a training set by the data set and a tag set corresponding to the data set; Setting network parameters, and constructing a GRU neural network; using the training set to train the GRU neural network, and reserving a partof data of the training set for verification to obtain a trained GRU neural network model; And preprocessing the HRRP data of the space target to be identified, inputting the preprocessed HRRP data into the GRU neural network model, and taking the highest probability as a space target identification result. According to the method, the deep essential characteristics of the radar HRRP data can beautomatically extracted by constructing an RNN neural network, so that the identification of a space target is completed. The influence of manual feature selection on the recognition result is eliminated, the time and energy of feature extraction are greatly reduced, and the recognition precision of the space target is improved.

The invention discloses a multiband image fusion recognition method. According to the method, a medium-wave infrared image and a long-wave infrared image are fused and used to supplement a training sample; and data processing is performed on the basis of the supplemented training sample. In this way, classified division can be performed on a visible light image with abundant details and color information and a high recognition by use of respective advantages of the medium-wave infrared image and the long-wave infrared image, data with respective maximum advantages is fused, and therefore information loss can be lowered to the maximum. Through the method, three types of classifiers are constructed first, and a fusion image of the medium-wave infrared image and the long-wave infrared image serves as the supplemented training sample; and since a feature early fusion technology is applied to the fusion image, the original single training sample has the outstanding characteristics of respective advantages of the medium-wave infrared image and the long-wave infrared image, and the probability of respective advantage information of medium-wave infrared and long-wave infrared is increased.

The invention belongs to the technical field of biological radar, and discloses a multi-human-body target recognition and positioning method based on multi-base IR-UWB biological radar signals. The method is based on a one-transmitting three-receiving three-channel biological radar. Echo signals are subjected to preprocessing and inflection point extraction, secondary inflection point signals of three channels and the first three largest values of the secondary inflection point signals are obtained, the peak-background ratio and the correlation coefficient mean value of the first three largestvalues are calculated, and recognition and positioning of multiple human body targets are achieved through judgment of the peak-background ratio and the correlation coefficient mean value of the three largest values. According to signal characteristics of human respiration and the like, a set of target discrimination program combining two characteristic parameters of an energy-to-noise ratio anda correlation coefficient mean value is adopted, recognition and distance direction positioning of multiple human targets are realized, and the target recognition accuracy is improved.

The invention relates to a novel submarine cable water area AIS target detection method. The method comprises the steps that AIS data sent by a ship around a harbor area is received, the water area ofthe harbor area is subjected to hierarchical gridding, and a ship ID is associated with coordinate positions of different levels; the AIS data is screened according to a time interval and a selectedtarget area, grid processing is carried out on the target area, and the frequency of the ship at different levels of grids is calculated to obtain a ship distribution frequency histogram; an image model is built by utilizing ship position information, the ship sailing trend is predicted by using a discrete random distribution Markov prediction model, and a prediction curve of the navigation trackof the target ship is formed; the information of the ship near the high-grade warning water area is extracted, and the information of the ship is identified in a graded grid chart of the harbor area,wherein the information of the ship includes the ship ID, the course and the navigational speed; a VHF radio station is controlled to broadcast to the ship near the high-grade warning water area. According to the method, the safety of the submarine cable water area is improved.

The invention discloses a kernel sparse representation-based fast remote sensing target detection and recognition method. The method includes the following steps that: S1, four RGB characteristic channels are created; S2, the four-phase Fourier transformation of the four characteristic channels of a given image is calculated, a phase spectrum is extracted, the images of the four characteristic channels are reestablished through inverse Fourier transformation, and a saliency map can be generated; S3, binaryzation division is performed on the saliency map obtained in the step S2, and candidate regions of interest are extracted; S4, a search box is scanned through an effective sub-window search algorithm, so that image blocks to be detected are obtained, so that a remote sensing target image block training set is obtained; S5, SIFT features are extracted from the remote sensing target image block training set, and a sparse dictionary is generated; S6, a spatial pyramid is adopted to map the SIFT features; S7, kernel sparse representation is obtained; S8, the kernel sparse representation is solved; S9, the space pyramid vector representation of a target is performed; and S10, a linear support vector machine classification algorithm is used in combination to complete a recognition task.

The invention provides a rapid segmentation method for a remote sensing image port water surface target. The rapid segmentation method relates to the field of remote sensing image processing, and solves the problem that the requirement of real-time performance cannot be satisfied due to high computation complexity, poor robustness of the existing image preprocessing method. The rapid segmentation method comprises the steps of: preprocessing an image; performing binarization processing on the original image to realize sea-land segmentation; extracting angular points of the image after binarization processing, and acquiring an image after sea-land segmentation; performing connected domain processing on the image after sea-land segmentation, acquiring a total number of connected domains, calculating the circumference of the connected domains, and determining the connected domains with a total number of connected domain pixel points greater than a connected domain circumference threshold value to be a land portion; and extracting a port target. The rapid segmentation method realizes sea-land segmentation, can extract the coastal target quickly and precisely, can reduce the calculation amount of subsequent target identification, accelerates the target identification speed, and can improve the target identification precision since the vast majority of land image interference is eliminated.

The invention relates to an SAR image target recognition method based on a multi-discriminator generative adversarial network, and the method comprises the steps: obtaining an original real sample, and generating a training sample data set based on a multi-discriminator generative adversarial network model; training the convolutional neural network by using the training sample data set to obtain atrained target recognition model; and inputting a to-be-detected SAR image to the target recognition model, and outputting a target recognition result corresponding to the to-be-detected SAR image through feature extraction and feature matching. Compared with the prior art, the invention provides a multi-discriminator generative adversarial network model for generating a high-quality and high-stability training sample data set. A training mode of multi-discriminator joint feedback is adopted, a dynamic adjustment selection function is utilized to fuse output results of discriminators to complete update training of generators, training stability and generated sample quality are ensured, the reliability of subsequent target recognition model training is further improved, and accuracy of SARimage target recognition is improved.

The invention discloses a method for making electroencephalogram target judgment with assistance of machine vision. With the method, a problem that the single-trial P300 detection algorithm precision is reduced because of changing of a P300 component incubation period due to specific implication of a target image can be solved. The method comprises: (1), an electroencephalogram signal collection device collects a tested electroencephalogram signal; (2), a changing error of a tested P300 component incubation period is determined; (3), a position of a target image in an image sequence is localized based on the P300 component; (4), a target candidate image is determined by combining the incubation period changing error determined at the step (2) and the target image position determined by the single-trail detection algorithm at the step (3); and (5), identification and classification are carried out on the target candidate image by using machine vision, and an image type with the highest occurrence frequency is used as a concerned target image type on trial. The method has advantages of high accuracy and low false alarm rate.

The invention provides a target identification method and device based on laser scanning, and the method comprises the steps: obtaining the data of each point obtained through the laser scanning reflection of a to-be-identified scene through a laser radar, and calculating the three-dimensional coordinate value of each point according to the data of each point, wherein the data comprises a linear distance value between a sampling point scanned by the laser radar and the laser radar; obtaining a two-dimensional coordinate value of each point according to the three-dimensional coordinate value ofeach point, and clustering each point according to the linear distance value of each point and the two-dimensional coordinate value of each point to generate a plurality of categories; calculating each piece of target information according to the two-dimensional coordinate values of each category of points formed by clustering; and determining each target in the to-be-identified scene according to each piece of target information. In order to solve the problem that the horizontal resolution and the vertical resolution of the vehicle-mounted laser radar are different, the density clustering algorithm is adopted for clustering, the clustering radius is self-adaption about the linear distance, the obtained clustering result is good, and therefore the target identification precision is improved.

The invention relates to an early warning method and device based on video analysis, electronic equipment and a storage medium, and the method comprises the steps: carrying out the target detection ofa collected video stream of a preset geographic region, and determining a first image frame, including a target, in the video stream; when the state of the video stream is an early warning silence state and the first image frame is detected for multiple times within a preset first duration, changing the state of the video stream into an early warning activation state, wherein the early warning silence state is used for indicating that no target exists in the preset geographic area, and the early warning activation state is used for indicating that the target exists in the preset geographic area; and when the video stream is in the early warning activation state, sending early warning information for the video stream. According to the embodiment of the invention, the target recognition accuracy can be improved, and reasonable early warning is effectively realized.

The invention relates to an early warning method and apparatus based on video analysis, an electronic device and a storage medium. The method comprises the steps of: performing target detection on a captured video stream of a preset geographic area, and determining a first image frame comprising a target in the video stream; when the state of the video stream is an early warning silence state, andthe first image frame is detected for multiple times within a preset first time length, changing the state of the video stream into an early warning activation state, wherein the early warning silence state is used for indicating that the target does not exist in the preset geographic area, and the early warning activation state is used for indicating that the target exists in the preset geographic area; and in the case that the first image frame is detected for multiple times within a preset second time length after the video stream enters the early warning activation state, sending early warning information for the video stream. By adopting the early warning method and apparatus provided by the embodiment of the invention, the target recognition accuracy can be improved, reasonable early warning is effectively realized, and the trouble caused by excessive early warning to users is reduced.

The invention discloses an unmanned aerial vehicle tracking method based on deep learning image processing technology, and belongs to the technical field of unmanned aerial vehicles. The tracking method comprises the following specific steps: (1) collecting a to-be-recognized target image; (2) preprocessing the image ; (3) recognizing a target; (4) fusion positioning the target; (5) measuring the speed of the target; and (6) performing tracking control. According to the method, a YOLOv3 algorithm model is used to carry out learning training on the target image, and image enhancement processing is carried out based on a CLAHE image enhancement preprocessing algorithm, so that the target recognition precision of an unmanned aerial vehicle in complex environments with environment occlusion or illumination condition change can be improved, and the situation that the unmanned aerial vehicle cannot recognize the target is avoided. In addition, a laser radar and a millimeter-wave radar are used to synchronously measure the position and the moving speed of the target in real time, and the Kalman filtering algorithm fusion is utilized to calculate the optimal solution, so that the positioning and tracking performance of the unmanned aerial vehicle in complex environments with heavy snow and heavy fog can be improved, and the target can be prevented from being lost.

The invention provides an SAR target identification method combining few-sample learning and target attribute characteristics. The SAR target identification method comprises the following steps: acquiring a source domain data set R, a target domain data set E, a target domain support set ES, a target domain query set EQ and a target attribute characteristic set A of R and ES; constructing a visualclassification network F1; performing iterative training on the visual classification network F1 by using R; constructing a visual attribute classification network F; performing iterative training onthe visual attribute classification network F by using the source domain data set R, the target domain support set ES and the target attribute feature set A thereof; and obtaining a target recognition result of the SAR image by using the target domain support set ES, the target domain query set EQ and the target attribute feature set A'o thereof. According to the SAR target identification method,the SAR target recognition performance under the condition that training samples of known categories are limited is improved by combining few-sample learning and target attribute characteristics.

The invention discloses a personnel target identification method for an unattended sensor system. The personnel target identification method comprises the following steps of S1, preprocessing the data; converting the original data acquired by the unattended ground sensor equipment into two forms of data of the time sequence data and the power spectral density spectrum data; S2, training a recurrent neural network; taking the two types of data obtained in the step S1 as the training samples respectively, and inputting the training samples into two different recurrent neural networks respectively for training; S3, connecting the recurrent neural network identification signals in parallel; connecting the two models generated through training in the step S2 in parallel, wherein one model judges whether a person walks or not, the other model judges whether the person runs or not, carrying out the OR operation on the results of the two models, and finally judging whether the person walks ornot. The method has the advantages that the principle is simple, the personnel can be detected in real time, and the recognition accuracy can be remarkably improved, etc.

The invention relates to a training method and application of a head and shoulder image classifier, and the method comprises the steps: obtaining a human body posture data set which comprises a plurality of human body sample pictures and human body head and shoulder labeling data corresponding to each sample picture; Synchronously obtaining an aggregation channel feature map and a downsampling aggregation channel feature map corresponding to each human body sample picture through a parallel algorithm; Based on the feature map, determining a positive detection window group and a corresponding positive feature vector group, and a negative detection window group and a corresponding negative feature vector group; And on the basis of the positive feature vector group and the negative feature vector group, performing multi-stage training on the enhanced decision tree to obtain a head and shoulder image classifier. Through a parallel algorithm, the head and shoulder detection speed is improved, and head and shoulder detection can be carried out in real time in a video stream; A feature map containing a plurality of feature channels is used for training, and a classifier result obtained through training is more reliable; And multi-stage training is carried out on the enhanced decision tree, so that the identification precision of the head and shoulder image classifier is greatly improved.

The invention provides an integrated imaging identification system and method based on a dynamic vision sensor, and aims to solve the technical problems of high system power consumption and low targetidentification accuracy in the prior art. The imaging identification method comprises the following steps: a lens collects a plurality of images; the dynamic vision sensor senses the change of the image gray scale; a processor in the data acquisition module analyzes the signals output by the dynamic vision sensor to obtain an address event data stream; a denoising sub-module in the target recognition module is used for denoising the address event data flow; a segmentation sub-module in the target identification module segments the address event data stream; an identification sub-module in thetarget identification module identifies the address event data flow; and the data visualization module acquires an imaging identification result.

The invention discloses an intelligent transfer robot target recognition method and system. The method comprises the steps: obtaining a multi-angle image sequence of a to-be-recognized region collected by a robot at different positions, carrying out the preprocessing of the image sequence, and converting the image sequence into an LAB color space; respectively extracting target images corresponding to the same target in the image sequence; performing feature fusion and splicing on the target images corresponding to the same target in the image sequence to obtain a fused target image; calculating a brightness adjustment coefficient of each fused target image based on the LAB value of each fused target image, and performing brightness correction on each target image; and judging a color interval to which the LAB value of the corrected target image belongs, and recognizing a color corresponding to the target. According to the method, the to-be-recognized target is processed through image fusion and brightness correction. Maximum color distinguishing between different targets is realized on the basis of keeping original color features of the target, and the adaptability to illumination change and the accuracy of target recognition are improved.

The invention discloses an aqueduct personnel identification system, an identification method thereof and an aqueduct personnel identification system, the system comprises an acquisition front end, anexchanger and a signal emitter, the acquisition front end is in communication connection with the exchanger, and the signal emitter is also in communication connection with the exchanger; The exchanger is also connected with an external network; The aqueduct scene monitoring system is characterized in that the acquisition front end acquires a monitoring picture of an aqueduct scene, the system further comprises a video analyzer, the video analyzer is built by a deep learning method, and the video analyzer is in communication connection with the exchanger. In the aqueduct scene, monitoring images shot by the collection front end suffer from the conditions of small targets, fuzzy characteristics, high personnel mobility and more interference factors such as vehicles, ships and umbrellas inevitably, the method is used for personnel detection, interference of small targets and negative samples except people can be effectively identified, the identification speed is high, and the system identification accuracy is high.

The invention discloses an experimental animal image calibration method based on hue saturation. The method is realized on the basis of OpenCV; When the extracted target is preprocessed, the target isnot converted into a grayscale image, but is converted into a HSV (Hue, Saturation, Value) color model, and a warm-toned colorant such as picric acid is applied to the target animal to distinguish itfrom other colors in the HSV color model, and the target extraction is performed on this basis.The method has the advantages that the requirement on the recognition environment is reduced, the anti-interference performance of target extraction is improved through dyeing, particularly, the influence on high light reflected by the water surface and the shadow of the corner can be ignored almost, and the target recognition accuracy is greatly improved in a similar analysis environment. The invention relates to a method for identifying a target through dyeing, which has lower dependence on the identification background, extracts a specific hue area target through color space conversion, and can resist high light on the water surface and corner shadow interference.

The invention discloses a residual unit, a network, a target recognition method, system and device and a medium, and relates to the field of underwater acoustic recognition, deep learning and artificial intelligence, and the multi-scale residual unit comprises a residual network architecture composed of a batch normalization layer, an activation layer and a convolution layer, and is characterized in that the residual network architecture is used for processing a one-dimensional original audio waveform; the residual error network architecture is provided with at least one convolution middle layer, the convolution middle layer is formed by a plurality of convolution layers in parallel, and at least one soft threshold value operation unit is arranged between the convolution middle layer and the tail end of the residual error network architecture and is used for carrying out soft threshold value calculation and filtering. According to the invention, the underwater sound one-dimensional signal waveform can be effectively sensed.