Search

CN-121982037-A - Ocean engineering material microstructure defect detection method and system based on image detection

CN121982037ACN 121982037 ACN121982037 ACN 121982037ACN-121982037-A

Abstract

The application discloses a method and a system for detecting microstructure defects of ocean engineering materials based on image detection, belonging to the field of image processing, wherein the method comprises the steps of acquiring an original scanning electron microscope image, extracting global first structure evolution characteristics and initializing denoising parameters; denoising an original scanning electron microscope image to obtain a current iteration denoising image, extracting local second structure evolution features and calculating feature deviation degree with global first structure evolution features, generating a current space constraint probability mask, inputting the current iteration denoising image and the current space constraint probability mask to generate an countermeasure network for reconstruction, outputting a current iteration restoration image, calculating reconstruction errors, decoupling the reconstruction errors, updating denoising parameters, performing closed loop iteration until convergence, and performing differential calculation on an optimal denoising image and the optimal restoration image to obtain a defect topography map.

Inventors

  • LI YATING
  • SHI DANDA

Assignees

  • 上海海事大学

Dates

Publication Date
20260505
Application Date
20260409

Claims (10)

  1. 1. The detection method for the microstructure defects of the ocean engineering materials based on image detection is characterized by comprising the following steps of: Acquiring an original scanning electron microscope image, extracting global first structural evolution characteristics and initializing denoising parameters; denoising the original scanning electron microscope image by using the denoising parameters to obtain a current iteration denoising image; Extracting local second structure evolution features and calculating feature deviation degree with the global first structure evolution features; dynamically distributing self-adaptive weights for the feature deviation degree to generate a current space constraint probability mask; Inputting the current iteration denoising image and the current space constraint probability mask into a generated countermeasure network for reconstruction, and outputting a current iteration repair image; Calculating the reconstruction errors of the current iteration repair image and the current iteration denoising image in the background area outside the mask, decoupling the reconstruction errors, updating the denoising parameters, performing closed-loop iteration until convergence, and outputting an optimal denoising image and an optimal repair image; and carrying out differential calculation on the optimal denoising image and the optimal repairing image to obtain a defect morphology image.
  2. 2. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 1, wherein the method for extracting the global first structural evolution features comprises the steps of extracting global space continuation reference features of an original scanning electron microscope image Global structural consistency benchmark features Global fluctuation reference feature Performing dimension splicing on the extracted result to obtain a global first structure evolution characteristic As a reference frame for evaluating the degree of abnormality, wherein, Is the spatial coordinates of the image pixels.
  3. 3. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 2, wherein the steps of extracting the local second structural evolution feature and calculating the feature deviation from the global first structural evolution feature comprise the following steps: Extracting a current iteration denoising image In the first place Local spatial continuity features of turns Local structural consistency feature Local statistics ; Performing mathematical difference on the local second structural evolution features and the global first structural evolution features of corresponding dimensions to generate feature deviation degree of the current turn: calculating spatial continuity deviation ; Calculating structural consistency deviation ; Calculating the deviation degree of the statistical characteristics ; Wherein, the The Fr Luo Beini Usnea norm of the matrix is represented.
  4. 4. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 3, wherein the method for dynamically distributing the self-adaptive weight for the characteristic deviation comprises the following steps: Quantifying the physical texture state of each pixel neighborhood in the current field of view, and calculating the local variance complexity : Introducing background reconstruction error term of previous round of feedback Constructing a confidence level of decay : ; Wherein, the For the preset system error tolerance scale parameter, in the first iteration At the time, the setting is initialized 。
  5. 5. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 4, wherein the pixel-level adaptive weights are constructed and distributed by combining the local variance complexity and the confidence attenuation degree: setting basic weight of space continuity deviation degree ; Setting basic weight of structure consistency deviation degree ; Setting basic weight of deviation degree of statistical characteristics ; Normalization processing, calculating dynamic self-adaptive weight set : ; Wherein, the , 、 、 Respectively marking a deviation feature set of space continuity, structural consistency and statistical features; ; as a denominator of the total weight, Representation and marking Corresponding basis weights.
  6. 6. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 5, wherein the generating the current space constraint probability mask specifically comprises the following steps: Mapping feature deviation degree of each dimension into local abnormal probability through truncation function Using dynamically adaptive weight sets Fusion is carried out to obtain a defect probability response diagram : ; After the spatial cohesiveness constraint is applied to the defect probability response diagram, the defect probability response diagram is input into a soft threshold mapping function to generate output values which are continuously distributed in Probability mask for current space constraint between : ; Wherein, the In order to obtain a response graph after spatial cohesiveness constraint, In order to map the slope parameter(s), Is the adaptive response center threshold for the current round.
  7. 7. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 6, wherein the current iteration denoising image and the current space constraint probability mask are input to generate an countermeasure network to reconstruct, and the current iteration restoration image is output The reconstruction formula of (2) is: ; Wherein, the Generating a deduction residual error output by a deduction sub-network for the defect prior in the reactance network, The microcosmic compensation residual error output by the self-coding sub-network for background fidelity, Is a preset weight factor.
  8. 8. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 7, wherein the method for calculating the reconstruction errors of the current iteration repair image and the current iteration denoising image in the background area outside the mask and decoupling the reconstruction errors specifically comprises the following steps: Constructing a pixel-level reconstruction disparity map specific to an out-of-mask background region : ; Separating kernels using two-dimensional gaussian frequencies Decoupling the reconstructed difference map into a low frequency reconstructed difference component And high frequency reconstruction differential : ; ; Wherein, the The convolution is represented by a representation of the convolution, Is a preset standard deviation of frequency separation.
  9. 9. The method for detecting the microstructure defects of the ocean engineering material based on the image detection according to claim 8, wherein updating the denoising parameters for closed loop iteration specifically comprises: global weighted absolute integration is respectively carried out on the low-frequency reconstruction difference component and the high-frequency reconstruction difference component, and a low-frequency penalty term is calculated And a high frequency penalty term ; Based on a finite difference approximation cutting method, homomorphic filtering Gaussian kernel scale parameters and density clustering neighborhood radius threshold values in the next round of denoising algorithm are adaptively updated: ; ; Wherein, the And (3) with Iteration step parameters which are respectively and independently set for low-frequency dimensions and high-frequency dimensions; updating background reconstruction total error item of current round feedback Returning to iteration with updated parameters until convergence conditions are met, wherein And The weights are penalized for a pre-set frequency, And (3) with Respectively the current first Round and last one Homomorphism filtering Gaussian kernel scale parameters of rounds; and (3) with Respectively the current first Round and last one A round density clustering neighborhood radius threshold; and (3) with Respectively the current first Round and last one Low frequency penalty term of round; and (3) with Respectively the current first Round and last one The high frequency penalty term for the turn, And (3) with Respectively the current first Round and last one Low frequency penalty term for round.
  10. 10. An image detection-based marine engineering material microstructure defect detection system, characterized in that the system comprises: The denoising module is used for denoising the original scanning electron microscope image by utilizing the denoising parameters to obtain a current iteration denoising image; the mask generation module is used for extracting local second structure evolution features and calculating feature deviation degree with the global first structure evolution features, dynamically distributing self-adaptive weights for the feature deviation degree and generating a current space constraint probability mask; the reconstruction module is used for inputting the current iteration denoising image and the current space constraint probability mask to generate an countermeasure network for reconstruction and outputting a current iteration restoration image; calculating the reconstruction errors of the current iteration repair image and the current iteration denoising image in the background area outside the mask, decoupling the reconstruction errors, updating the denoising parameters, performing closed-loop iteration until convergence, and outputting an optimal denoising image and an optimal repair image; and the detection module is used for carrying out differential calculation on the optimal denoising image and the optimal repairing image to obtain a defect morphology image.

Description

Ocean engineering material microstructure defect detection method and system based on image detection Technical Field The application relates to the field of image processing, in particular to a method and a system for detecting microstructure defects of ocean engineering materials based on image detection. Background The macroscopic mechanical failure of the ocean engineering material often originates from defect evolution of a microstructure layer of the material, such as holes, microcracks, abnormal grain boundaries or discontinuous local tissues and the like, when the ocean engineering material is in service in complex environments such as high humidity, high salt, high pressure, alternating load and the like for a long time. Therefore, microstructure image analysis based on a scanning electron microscope has become an important technical means in marine engineering material quality assessment, life prediction and failure early warning. With the development of deep learning, part of the prior art performs microstructure reconstruction by introducing a generated countermeasure network, compares the reconstructed SEM image with an original SEM image, and determines detailed defect morphology, thereby realizing defect detection. However, in the process of generating SEM images, the SEM images are easily affected by factors such as sample preparation process differences, magnification changes, electron beam parameter fluctuations, and the like, so that a large amount of random disturbance noise exists in the images, and the random disturbance noise can be optimized as structural defects by an countermeasure network, so that misjudgment of subsequent defect identification is caused. The existing method is mostly processed by adopting filtering, smoothing or deep denoising network, and does not distinguish real microstructure evolution characteristics from false microstructure changes introduced by imaging, so that structural details are easy to delete or false defects are easy to introduce when denoising is caused. In addition, existing generation countermeasure networks mostly target pixel reconstruction errors, and perform pixel-level reconstruction on the full map. When processing complex micro textures of materials, the model often fine-tunes a normal background area in order to pursue global visual similarity, so that extra noise is introduced into the normal background area, and the noise is easily mistakenly identified as a real defect in a subsequent defect detection process, so that the accuracy and the reliability of defect detection are reduced. Therefore, how to eliminate false features introduced by the imaging environment and reduce the false generation probability of the generation countermeasure network, so as to reduce the misjudgment and missed judgment probability of the final defect is a technical problem to be solved at present. Disclosure of Invention The invention provides a detection method of microstructure defects of ocean engineering materials based on image detection, which comprises the following steps: Acquiring an original scanning electron microscope image, extracting global first structural evolution characteristics and initializing denoising parameters; denoising the original scanning electron microscope image by using the denoising parameters to obtain a current iteration denoising image; Extracting local second structure evolution features and calculating feature deviation degree with the global first structure evolution features; dynamically distributing self-adaptive weights for the feature deviation degree to generate a current space constraint probability mask; Inputting the current iteration denoising image and the current space constraint probability mask into a generated countermeasure network for reconstruction, and outputting a current iteration repair image; Calculating the reconstruction errors of the current iteration repair image and the current iteration denoising image in the background area outside the mask, decoupling the reconstruction errors, updating the denoising parameters, performing closed-loop iteration until convergence, and outputting an optimal denoising image and an optimal repair image; and carrying out differential calculation on the optimal denoising image and the optimal repairing image to obtain a defect morphology image. Extracting global first structural evolution characteristics, which concretely comprises extracting global space continuation reference characteristics of an original scanning electron microscope imageGlobal structural consistency benchmark featuresGlobal fluctuation reference featurePerforming dimension splicing on the extracted result to obtain a global first structure evolution characteristicAs a reference frame for evaluating the degree of abnormality, wherein,Is the spatial coordinates of the image pixels. Extracting local second structure evolution features and calculating feature deviation degree with global first