61 results about "Background subtraction" patented technology

The invention relates to the technical field of ocean structure monitoring, in particular to a method for monitoring the scouring degree of a pile foundation of an offshore wind power station through bubble tracing, which comprises the following steps of: 1, constructing a bubble feature data acquiring an underwater image, manually segmenting and selecting key bubbles, extracting contour features of the key bubbles and expressing the contour features by vectors, a Zernike moment is used for grouping according to shapes, and a bubble feature database is constructed; and 2, identifying underwater bubbles: obtaining an underwater background image, constructing a dynamic background, obtaining an image containing suspended particles, extracting a foreground main body through a background subtraction method, comparing contour features of the foreground main body with a bubble feature database, and screening out the bubbles. By introducing an innovative method combining image recognition, feature modeling and physical inversion, the key problems of bubble misjudgment, track loss, inaccurate flow velocity and the like in a traditional bubble tracing technology are effectively solved, and high-precision, low-cost and automatic monitoring of the scouring process of the offshore wind power pile foundation is achieved.

The embodiment of the invention provides an actor recognition method, system and device based on ViT gait analysis and a medium, and belongs to the technical field of gait recognition. The method comprises the following steps: acquiring behavior video data of a target object; performing background subtraction processing on the behavior video data to obtain a binary image of the gait of the target object; inputting the binary image into a feature analysis model, wherein feature analysis is constructed and formed based on Transform; performing attitude feature extraction through a double-path parallel multi-head attention module of the feature analysis model, fitting a feature weight of the attitude feature through a feature average fusion module, and outputting an embedded feature vector; and comparing the embedded feature vector with a preset feature vector, and identifying the identity of the target object according to a comparison result. The embodiment of the invention aims to improve the accuracy of performing actor identity recognition through gaits.

An inspection method performed by a computer for inspecting a display panel includes: obtaining an anomalous portion image that includes an anomalous portion of a pixel region of the display panel, the anomalous portion being acquired by performing image processing using a background subtraction method on an inspection image of the pixel region; generating, using a trained generative model, a label image from the anomalous portion image by converting a region indicating the anomalous portion into a region of a color corresponding to a fault mode of the anomalous portion; and determining, based on the color of the region in the label image, whether the fault mode of the anomalous portion has a possibility of being a seepage defect in which light is not emitted due to deterioration of a functional layer in the pixel region. The fault mode includes the seepage defect and a dark-dot defect.

The invention relates to the technical field of environment monitoring, in particular to an internet-of-things-based breeding environment monitoring system for pond turtles, and the system comprises a turtle movement trajectory tracking module which is used for collecting real-time images of an incubator, comparing pixel differences among continuous image frames to carry out background subtraction, and recognizing a moving foreground target. According to the method, the pixel differences between image frames are continuously compared, the geometric center point of the moving target is calculated, continuous recording of the moving tracks of the water tortoise individuals in the cultivation box is achieved, the moving area is gridded, the occurrence probability is calculated by counting the frequency that coordinates fall into all units, an objective data basis is provided for environment optimization, and the method is suitable for large-scale popularization and application. Meanwhile, multiple groups of temperature sensors are arranged at key positions such as a spawning area and a drying platform, multiple paths of environment sensing redundant numerical values are obtained, the state range is predicted according to a water body thermal inertia model, multiple paths of readings are updated and calibrated, deviation data are effectively recognized and eliminated, and the reliability of environment monitoring data and the fault tolerance of the system are improved.

The invention discloses a fan non-stop routing inspection blade tip triggering method, which comprises the following steps: S1, an unmanned aerial vehicle executes a fixed route to reach a blade tip waypoint, a single-point laser radar is adopted to measure and calculate the spatial pose of a fan blade in real time, whether the unmanned aerial vehicle triggers photographing within a preset time is judged, and if not, the unmanned aerial vehicle stops routing inspection; if so, calling a leaf tip perception compensation algorithm to calculate a new leaf tip photographing waypoint; s2, through a blade tip sensing compensation algorithm, realizing stable tracking and picture locking of the pan-tilt camera on the blade tip of the fan, and obtaining an image of the fan at a horizontal position through a background difference method; s3, calculating the proportion of the blade in the picture, and calculating the distance from the current waypoint position to the blade tip in the horizontal direction by combining the internal reference of the pan-tilt camera and the inspection distance; and S4, the unmanned aerial vehicle is controlled to fly to the new blade tip waypoint, and blade image capturing is executed. According to the scheme provided by the invention, a closed-loop mechanism is formed through the single-point radar, the compensation algorithm, the background difference and the waypoint offset calculation process, and accurate acquisition of the blade tip area image is completed for accurately calculating the target point again.

Image sensor, which includes the following: a pixel array including multiple pixels configured and functional to detect light from a scene, wherein the scene includes an object to be detected and a background image, wherein each pixel of the multiple pixels is configured to generate a new voltage signal in response to the intensity of light received by the pixel; a memory for storing digital values ​​corresponding to a predetermined voltage bandwidth around a prior voltage signal R of at least some of the multiple pixels corresponding to the light detected from the background image, wherein the voltage bandwidth defines a positive delta voltage +Δ and a negative delta voltage -Δ relative to the background voltage signal R; a sample-and-hold or track-and-hold module capable of sampling the new voltage signal corresponding to one of the several pixels selected; a digital-to-analog converter that generates a reference voltage R+Δ or R-Δ from the digital values ​​stored in the memory; a comparator that receives the new voltage signal sample value from the hold module and receives one of the reference voltages R+Δ or R-Δ for the selected pixel to compare it with the voltage signal sample value, wherein the comparator is configured to produce a high-voltage output signal if the voltage signal sample value is greater than the reference voltages R+Δ or R-Δ, and to produce a low- or zero-voltage output if the voltage signal sample value is less than the reference voltages R+Δ or R-Δ; a two-bit register containing a most significant bit (MSB) and a least significant bit (LSB), wherein the two-bit register provides a two-bit output signal; and a logic module that receives the voltage output from the comparator, wherein the logic module is configured and functional as follows: Selecting the reference voltage R+Δ to be supplied to the comparator, and setting the MSB to a logic "0" in response to a low-voltage output signal from the comparator or to a logic "1" in response to a high-voltage output signal from the comparator; and then Selecting the reference voltage R-Δ to be supplied to the comparator and setting the LSB to a logic "0" in response to a low-voltage output signal from the comparator or to a logic "1" in response to a high-voltage output signal from the comparator; a background subtraction module configured and functional to receive the two-bit output signal from the two-bit register for the selected pixel and, based on the two-bit output signal, to identify the selected pixel as detecting light from a scene that is different from the light detected by the same pixel in the background image; and a processing module for generating an output image that includes the selected pixel identified by the background subtraction module.

The invention provides a method and system for measuring the side deformation degree and the girdling depth of a capacitor, and relates to the technical field of intelligent quality inspection, and the method comprises the steps: carrying out the background subtraction and image binarization of a side image of the capacitor, so as to obtain a binary image of the capacitor; carrying out segmented detection on the capacitance binary image by adopting a slope and curvature combined mode to obtain wave crest information and wave trough information of a side surface deformation area and a girdling area; and performing side deformation degree measurement according to the wave crest information and the wave trough information of the side deformation area, and performing girdling depth measurement according to the wave crest information and the wave trough information of the girdling area so as to obtain a side deformation degree measurement result and a girdling depth measurement result. The side deformation degree and the girdling depth of the capacitor are measured based on the machine vision technology, and the problems that the side deformation degree and the girdling depth of an existing capacitor are high in measurement cost and poor in measurement precision are solved.

The invention discloses a fire fighting access occupation detection method based on multi-target tracking. According to the method, firstly, foreground targets are extracted through background subtraction and morphological processing, and then an independent life cycle management mechanism is established for each target by adopting a multi-target tracking algorithm based on intersection-to-parallel ratio IOU matching. By calculating the continuous existence time of the target and combining the spatial relationship between the target position and the preset no-parking area, the intelligent judgment that the alarm is triggered only when the continuous parking of the target in the area exceeds the set threshold value is realized. According to the method, the problem that the false alarm rate of a traditional area intrusion detection system is high is effectively solved, instant passing and illegal occupation behaviors can be accurately distinguished, and the accuracy and practicability of fire fighting access occupation detection are remarkably improved.

The application discloses an infrared gas imaging detection method based on a filter array, and relates to the technical field of gas analysis. The method comprises the following steps: setting a filter unit array on a single infrared camera focal plane to determine the corresponding relationship between each filter channel and a pixel; determining a characteristic channel with strong absorption and weak absorption to other gases for each target gas, and constructing a channel-gas response matrix; dividing infrared radiation into multiple bands by using the array, collecting original multi-channel images on the same imaging surface, and deconstructing each channel image; calculating the normalized absorption of each channel after background subtraction and normalization, and combining the normalized absorption into an observation vector; performing multi-gas unmixing by using a non-negative least square method according to the response matrix, solving the component images of at least two target gases, and outputting the types, positions and distribution ranges of the target gases. The application can simultaneously separate the spatial distribution of at least two target gases, eliminate parallax and registration errors, and improve the accuracy and stability of multi-gas detection.

This invention discloses an adaptive adjustment method and system for a virtual environment, relating to the field of virtual reality technology. The method includes: performing background subtraction processing on a target image sequence when a user is experiencing a virtual environment; extracting detailed features from the target image sequence, and identifying key points using target anchor boxes based on the detailed feature information; determining the coordinates of three-dimensional key points based on the target key point information for user action behavior analysis; extracting a user eye image sequence from the target image sequence to determine eyeball image information, and analyzing the gaze direction based on the eyeball image information; analyzing the user's attention focus using image brightness distribution based on the gaze direction information; and adjusting elements of the virtual environment based on the user's action behavior information and user attention focus information. This invention enables more fine-grained virtual environment adjustment, improving the quality and naturalness of the user's immersive experience in the virtual environment.

The invention discloses a fixed scene real-time target counting system based on example driving, and the system comprises a cross-modal interaction module which receives a visual frame selection example and a natural language instruction to generate a unified example feature vector, and a lightweight feature extraction module which extracts a fixed scene image multi-scale feature map based on a post-distillation lightweight DINO backbone network. The space-time correlation processing module filters static interference and focuses on a dynamic target through an optical flow method and a background subtraction algorithm by utilizing the background constant characteristic of a fixed scene, and the real-time density map generation module normalizes a strategy for a priori optimization density map; and the edge computing optimization module performs edge device optimization deployment to reduce the computing power consumption and the storage demand of the embedded device. The method has the beneficial effects that multi-modal input processing, lightweight efficient feature extraction, accurate space-time target tracking, real-time density map generation counting and low-power optimization of edge equipment are realized, and the real-time performance, the accuracy and the deployment flexibility of the system are improved.

The invention discloses an inland waterway traffic complexity evaluation method based on a radar sequence image, belongs to the technical field of intelligent monitoring of inland waterways, and aims to solve the problems of poor traffic situation perception real-time performance, low ship detection precision and insufficient dynamic interaction quantification in the prior art. Radar noise is removed through background subtraction and a gray region threshold value, a target and a background are separated, and accurate ship detection is achieved; kalman filtering and a Hungary algorithm are adopted to construct a continuous motion trail, an Euclidean distance cost function is utilized to distribute identification numbers, and the problem of number jump in shielding and dense scenes is solved; integrating relative distance, encounter angle and trend dynamic information, and quantifying ship interaction strength through a nonlinear function; and constructing an undirected weighted graph model, and realizing traffic complexity visual evaluation based on the symmetry of the adjacent matrix. According to the method, high-risk ship pairs can be marked in real time, accurate data support is provided for collision avoidance decision making and channel scheduling, and channel traffic management efficiency and safety are remarkably improved.

The present application belongs to the field of computer vision, and particularly relates to a video anomaly detection method, system and device based on scene classification. The steps of the present application method are as follows: step S1, performing gray processing on original video frames, using background subtraction method and frame difference method to extract foreground targets, and extracting pixel features through Double-Canny algorithm; step S2, using a pre-trained model to detect the number of people in the video frames and generate a heat map of the video frames; step S3, using K-means clustering algorithm to classify scenes of the video frames, and dividing the video frames into two categories of dense scenes and sparse scenes; and step S4, using an anomaly detection module to respectively perform anomaly detection on the dense scenes and the sparse scenes. The present application can select different network structures for anomaly detection according to scenes with different crowd densities, is better applied to complex real scenes, and achieves better anomaly behavior detection effect.

This invention discloses a visualization measurement method for solid-liquid phase change of electrically controlled solid propellants based on image recognition. The method is implemented using an electrode clamping device and includes synchronous control of image acquisition and laser illumination, background subtraction combined with adaptive threshold binarization, application of regional constraints and morphological image processing, extraction and tracking of phase change interface features, and statistical analysis of target parameters. This method enables continuous and intuitive optical observation of the solid-liquid phase change process, achieving high-fidelity visualization of the phase change process before combustion. Through the synergistic effect of laser directional illumination and image recognition algorithms, it effectively distinguishes the solid region, liquid molten region, and suspended particle region of the propellant. Candidate motion regions are segmented using background subtraction combined with adaptive threshold binarization, reducing light interference caused by laser illumination. This achieves effective identification and interface extraction of the main solid-liquid phase change region and weak phase change region, improving the accuracy of phase change morphology characterization.

The invention discloses a non-contact vital sign recognition and detection system, which relates to the technical field of vital sign monitoring, and comprises a time-sharing echo acquisition module for acquiring echo signals and constructing an original data matrix; the phase processing module is used for executing background subtraction and phase extraction processing on the original data matrix and resolving a phase signal; the vital sign separation module is used for processing the phase signal and generating a breathing waveform; the breathing event judgment module is used for converting the breathing waveform into a physical displacement waveform, and judging and generating an apnea early warning signal; the radar dynamic scheduling module is used for generating a radar switching instruction according to the calculated target position; and the power compensation module obtains the signal quality parameter and generates a power adjustment instruction. According to the invention, through cooperation of double radars and dynamic scheduling, global blind-area-free continuous monitoring is realized, obstacles can be penetrated and breath and heartbeat can be accurately separated by using an advanced signal processing method, and low power consumption and high environmental adaptability are realized while high precision is ensured.

According to the automobile hub classification method and system based on machine vision provided by the invention, the working rhythm of a structural light source is designed, an industrial camera collects a plurality of images, a background subtraction method is utilized to obtain common features, namely a spoke window feature region, of a hub, the hub images are subjected to normalization processing, and a hub region is detected and positioned by utilizing a Hough circle; the area of the circular region of the hub is converted into a rectangle through polar coordinate transformation, and then multi-dimensional feature extraction and classification matching are carried out on a non-spoke window region. Multi-dimensional features including color, diameter, spoke number, spoke width, center hole number, gray level co-occurrence matrix and the like are extracted, and classification decision is carried out by adopting a hierarchical matching strategy. Hub geometric measurement, classification identification and template library self-updating can be realized. And requirements of high reliability, simple integration and the like of industrial vision are met.

The invention discloses a machine vision-based sundry road occupation detection method. The method comprises the following steps of: acquiring and preprocessing data; dual-mode differential detection: running a template differential channel and a background subtraction channel in parallel, calculating an absolute difference value between a current frame and a reference template by the template differential channel, and performing binarization processing and morphological closed operation to obtain a newly added static target area; the background subtraction channel dynamically models a background based on an MOG2 algorithm and separates a moving target; removing all pixels overlapped with the dynamic target mask in the newly added static target area through matrix subtraction operation, and outputting a static abnormal target image; target analysis; target tracking; and decision output. According to the embodiment, high accuracy can be achieved in various environments, the method can effectively distinguish temporary articles from illegal lane occupation, real-time alarm is supported, an efficient solution is provided for intelligent security and protection, and the method is suitable for intelligent supervision requirements of factories, fire fighting accesses and other scenes.

In an FMCW radar system, the chirp pair modes of the master and slave radar units are selected such that when one chirp in the master radar's chirp pair (A,B) coincides in time with one chirp in the slave radar's chirp pair (C,D), the other chirs in these pairs do not coincide in time. This allows for background subtraction and enables the slave radar to achieve self-synchronization with the master radar by detecting the in-band tone in the difference between the spectral responses obtained when the slave radar chirps and the master radar signals are mixed, and driving it to a specific setpoint after detecting the tone.

The embodiment of the invention discloses a laser point cloud background differencing method and system, and the method comprises the steps: obtaining a current frame point cloud; preprocessing the point cloud of the current frame according to the scene prior information to obtain a processed point cloud; performing grid division on the processing point cloud according to a preset grid side length, and calculating a grid coordinate of a grid corresponding to each point in the processing point cloud; performing Morton index calculation according to the grid coordinates and a preset rule to obtain a first index value corresponding to each point; sorting all points in the processing point cloud according to the first index value to obtain a first sorted point cloud; and traversing each point in the first sorted point cloud, searching adjacent points in the second sorted point cloud according to a preset search radius, and determining background points according to the number of the adjacent points. And the background difference efficiency is improved.

The invention belongs to the technical field of quality detection, particularly relates to an online quality detection method and system for casting powder, and aims to solve the technical problems of poor accuracy and reliability of a detection method in the prior art. The detection method comprises the following steps: S1, obtaining a spectral signal from which a background is deducted, and calculating the global signal-to-noise ratio of the spectral signal; s2, calculating a second derivative spectrum of the spectral signal; selecting an effective spectrum peak; s3, classifying a plurality of effective spectrum peaks smaller than a preset clustering coefficient into a spectrum peak cluster according to the ratio of the distance between the center positions of the adjacent effective spectrum peaks to the sum of the respective estimated full width at half maximum; performing synchronous fitting on each spectrum peak cluster by adopting linear superposition of a plurality of asymmetric Lorentz functions; and S4, based on the peak area of the characteristic spectral line of each element obtained by fitting, establishing a working curve between the peak area and the content of the corresponding element in the casting powder, and determining the component content of the casting powder. And the accuracy of quantitative analysis results of the element content of the casting powder is ensured.

The invention discloses a flame image recognition and feature extraction method and system, and belongs to the field of engine testing. The method comprises the following steps: performing background noise reduction and boundary enhancement preprocessing on an engine combustion flame image; a preliminary mask is generated by adopting an adaptive threshold method, and fine segmentation is realized in combination with a U-Net network; tracking flame movement by using a background subtraction method and an optical flow method, and analyzing the propagation speed and direction; three-dimensional reconstruction is realized based on MonoDepth2 monocular depth estimation and binocular vision parallax calculation, and a flame space position and an external cube are determined; constructing a convolutional neural network model by taking ResNet-50 as a backbone network, and respectively predicting continuity parameters and discrete parameters through regression branches and classification branches; and fusing the flame area, the speed, the color histogram and the edge features, and finishing final recognition through an SVM (Support Vector Machine) classifier. According to the invention, automatic flame identification, three-dimensional positioning and multi-feature parameter extraction are realized, and manual intervention is reduced.

This invention provides a mobile object detection device and a mobile object detection program that efficiently detect moving objects using the background subtraction method. [Solution] The moving object detection device 200 comprises an object detection unit 250 and a background update unit 290. The object detection unit 250 detects an object estimation region represented in the input data 610 based on difference data 620 representing the difference between background data 600 and input data 610. The background update unit 290 updates the background data 600 based on fixed object data representing the object when background conditions are met, including the fact that the velocity of the object represented in the object estimation region is below a threshold. The difference data 620 is used to detect the moving object 20 represented in the input data 610.

A method for detecting spilled objects by integrating differential video sequences and convolutional neural networks is proposed. First, a static background model of the scene is constructed. Then, a differential video sequence is built using a Gaussian-weighted background subtraction method to extract the foreground. Next, stationary targets within the foreground are extracted and combined with a convolutional neural network model to analyze and exclude non-sprayed objects from the stationary targets in the foreground, thus detecting the spilled object in the current video frame. This invention fully utilizes existing tunnel monitoring facilities, minimizing hardware investment. It can automatically and in real-time monitor the occurrence of spilled object events in tunnels, achieving intelligent road detection of spilled object events in tunnel traffic.

The invention relates to the technical field of range-gated imaging, and provides a range-gated imaging noisy point elimination method for adaptive background suppression, which comprises the following steps: acquiring each pixel point of a range-gated image, and inputting the pixel point into a position encoder to generate a high-dimensional feature vector; constructing a background noise model, executing adaptive background subtraction on the range gating image, and outputting an intermediate range gating image; splicing the intermediate range gating image and the high-dimensional feature vector in a channel dimension to output a fusion feature; and inputting the fusion feature to the dynamic network structure, and outputting a reconstructed range gating image. According to the invention, through adaptive background modeling, position coding enhancement and a dynamic network structure, balance between noise suppression and signal fidelity is realized.

The application provides a Mach-Zehnder interference structure free-state molecule interaction analysis method and system, in a free system, a light intensity imaging module is used to perform molecule interaction monitoring, through a Mach-Zehnder interference optical path, a refractive index change caused by free system biomolecule combination is converted into a phase difference between two different polarization components formed by a detection light at a detection arm and a reference arm; subsequently, a weak value amplification effect is introduced by near-orthogonal post-selection, so that the small phase difference is converted into a light intensity change, and system noise interference is reduced through light intensity image processing, baseline correction and background subtraction; the application can solve the problems of spatial steric hindrance, limited mass transfer, complex detection process, difficulty in truly reflecting the molecule interaction process in the free system, insufficient sensitivity to weak signals, and easy environmental noise interference in the existing fixed system label-free detection technology, and can realize high-sensitivity, real-time and high-throughput detection of biomolecule interaction in a free solution system.

The present disclosure provides a particle size statistical method of granular minerals of shale, comprising: collecting shale samples; cutting the collected samples into the size of 1 cm×1 cm×1 cm; performing argon ion polishing; magnifying the samples by the field emission scanning electron microscopy, collecting 500-1000 images; using Adobe Photoshop to seamlessly splice the collected images; using ImageJ in combination with EDS energy spectrum to determine types of mineral particles, performing background subtraction, and setting a reasonable grayscale threshold to classify the particle minerals; performing binary processing on the images of selected mineral particle types; identifying edges of granular minerals, and using a Analyze Particles command to statistically analyze and measure particle sizes, and drawing a particle size distribution histogram. The method can perform precise measurement and statistics on the particle sizes of shale particle minerals, and can perform full-scale statistical analysis on the distribution of different mineral particle sizes in the shale.

The invention belongs to the technical field of radiation detection, and particularly relates to a synthetic aperture radiometer target detection method and device under a high dynamic condition based on continuous frame image processing. Utilizing a synthetic aperture radiometer to continuously obtain millimeter wave signals of a target scene, and obtaining visibility functions of a plurality of continuous frames according to preset integral time; performing background subtraction, Gaussian low-pass filtering and Fourier transform on the visibility function sequence to obtain a plurality of frames of brightness temperature maps; solving a rotation matrix of the current frame image according to the attitude angle at the current moment, and performing registration fusion on the brightness temperature image according to the rotation matrix to obtain a preliminary result image; a convolution kernel is set according to antenna array arrangement and platform motion imaging properties, the preliminary result image is enhanced, then the image is further enhanced by adopting multi-scale white top-hat transform, and a final result image is obtained; and setting a target detection window, screening candidate target points according to pixels in the window, and finally setting a threshold value according to a full-image pixel mean value and variance to determine the positions of the target points.

The application discloses a flexible temperature and pressure sensor based on lensless imaging and a measuring method thereof, and belongs to the technical field of multi-modal flexible sensing. The sensor stack structure comprises an image sensor, a coding mask, a beam splitter, a transparent elastic pressure-sensitive layer, a thermochromic temperature-sensitive film and a top reflective layer in sequence. A side light source forms a normal illumination light path along the stack structure after being folded by the beam splitter. The illumination light is reflected by the reflective layer after penetrating the pressure-sensitive layer and the temperature-sensitive layer, and the light field carrying temperature and pressure modulation information is captured by the image sensor after being spatially encoded by the mask. The application adopts 0D perovskite (C16H20N2) SbBr5 as a temperature-sensitive material, effectively suppresses optical crosstalk noise by combining a differential background subtraction algorithm, extracts a spatial geometric deformation field and an RGB normalized color ratio of a coding pattern synchronously by double-branch preprocessing, and realizes high-precision orthogonal decoupling of temperature and pressure signals by combining global gray feedback closed-loop control. The application has the advantages of ultra-thin structure, flexible sensing layer and strong anti-interference capability, and is suitable for intelligent wearing and industrial curved surface monitoring.

Disclosed are systems and methods that leverage machine learning techniques to provide information about items that are on display during a live event. A 3D volumetric capture system is used to construct a 3D representation of the event in real time. For example, depth and color data from multiple perspectives may be used to construct a 3D mesh of colored polygons that represent the scene. Users may view the live event with a 3D enabled virtual reality or augmented reality headset. Background subtraction may be used to isolate items of interest. In some configurations a machine learning model may process the 3D mesh to generate a textual description of items in the scene. The textual description may then be used to search for more information about a particular item. Additionally, or alternatively, the machine learning model may directly predict one or more items in the scene.

The application provides a kind of based on deep learning's radiographic image weld defect detection method, utilizes Faster RCNN network to carry out target detection to radiographic image, on the basis of traditional Faster RCNN, the application increases background subtraction network layer, obtains the Faster RCNN network structure of five modules, the background subtraction network layer has with the parameter of updating, can constantly optimize and reduce the effect of background along with network training, highlights the feature of defect, simultaneously, utilizes three branch area recommendation network layer instead of traditional area recommendation network, it has two regression branches, so that the task of predicting defect position information is subdivided, one branch is responsible for predicting the horizontal coordinate and width of the center point of defect, one branch is responsible for predicting the vertical coordinate and height of the center point of defect, enhances the accuracy of defect position prediction.