CN-121999006-A - Image segmentation micro-organization identification method based on convolutional neural network model

CN121999006ACN 121999006 ACN121999006 ACN 121999006ACN-121999006-A

Abstract

The invention relates to the field of image recognition, in particular to an image segmentation micro-tissue recognition method based on a convolutional neural network model, which comprises the steps of sampling a turbine blade to obtain a plurality of samples, respectively carrying out image acquisition on the samples subjected to material processing to obtain a plurality of target images, respectively carrying out noise processing on the target images, determining an image characteristic state according to a contrast uniformity value and a local difference value of the target images subjected to the noise processing, determining a final binarization threshold value corresponding to each target image based on a pixel gray value corresponding to an inter-class variance or a sub-region according to the image characteristic state, carrying out threshold segmentation marking on the target images subjected to the binarization processing to obtain a label image, training a Res_ Unet model through a training data set, and carrying out adaptive adjustment on sample processing and model training according to the acquisition condition of actual acquired samples, thereby improving the model training effect.

Inventors

SUN YANTAO
FU ZHIZHONG
FAN TIANYU
ZHOU HONGBIN
JI PENGFEI
ZHANG SHENGLIANG
ZHENG XIAOMEI
XIA AIGUO
DONG LIWEI

Assignees

中国人民解放军93208部队

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. The image segmentation micro-organization identification method based on the convolutional neural network model is characterized by comprising the following steps of: Sampling each target position corresponding to the turbine blade to obtain a plurality of samples, and respectively carrying out image acquisition on each sample after material treatment to obtain a plurality of target images; Respectively carrying out noise processing on each target image, determining an image characteristic state according to the contrast uniformity value and the local difference value of each target image subjected to the noise processing, and determining a final threshold value corresponding to each target image based on the pixel gray value corresponding to the inter-class variance or the sub-region according to the image characteristic state; Performing binarization processing on the target image subjected to noise processing based on a final threshold value, and taking the image subjected to binarization processing as a label image; And carrying out data set expansion based on the label image, carrying out data division on the final data set obtained by expansion, and training the Res_ Unet model through a training data set.
2. The convolutional neural network model-based image segmentation micro-organization recognition method of claim 1, wherein the final threshold is determined based on the inter-class variance for image feature states having a contrast uniformity value greater than a preset contrast uniformity value and a local variance value less than or equal to a preset local variance value.
3. The convolutional neural network model-based image segmentation micro-organization recognition method of claim 2, the determining a final threshold based on an inter-class variance comprising: And obtaining a plurality of thresholds to be selected, dividing the target image based on the thresholds to be selected respectively, calculating the inter-class variance between the foreground and the background of the divided target image, and selecting the threshold to be selected with the maximum inter-class variance as a final threshold.
4. The convolutional neural network model-based image segmentation micro-organization recognition method of claim 1, wherein for image feature states where a contrast uniformity value is less than or equal to a preset contrast uniformity value or a local difference value is greater than a preset local difference value, a final threshold is determined based on pixel gray values corresponding to the sub-regions.
5. The method for identifying the image segmentation micro-organization based on the convolutional neural network model according to claim 4, wherein determining the final threshold based on the pixel gray value corresponding to the sub-region comprises: and carrying out local uniform division on the target image to obtain a plurality of subareas, and detecting an average value of pixel gray values corresponding to each subarea as a final threshold value corresponding to each subarea.
6. The convolutional neural network model-based image segmentation micro-organization recognition method of claim 1, wherein the corresponding cross entropy loss function in training the res_ Unet model is: , Wherein, the Is the first The label value corresponding to each pixel point, Is the first The label probabilities corresponding to the individual pixels, Is the total number of pixels.
7. The method for identifying image segmentation micro-organization based on convolutional neural network model as set forth in claim 6, wherein for the first step And outputting the corresponding reinforced phase probability of each pixel point according to a Sigmoid activation function, wherein the Sigmoid activation function is as follows: , Wherein, the Is the first The probability of the strengthening phase corresponding to each pixel point, Is the first And pixel values corresponding to the pixel points.
8. The method for identifying the image segmentation microstructure based on the convolutional neural network model according to claim 7, wherein in the training process of the res_ Unet model, the learning rate is set according to the number of learning rounds, and the calculation formula of the learning rate is as follows: , Wherein, the In order for the rate of learning to be high, As the number of reference wheels to be used, Is the number of learning rounds.
9. The convolutional neural network model-based image segmentation micro-organization recognition method of claim 8, wherein if the image feature diversity is less than a preset image feature diversity, performing reduction adjustment for a reference number of rounds; The reduction amount of the reference wheel number and the image feature diversity are in a negative correlation relationship.
10. The method for identifying the image segmentation microstructure based on the convolutional neural network model according to claim 1, wherein the material treatment comprises embedding the sample by using cold mosaic resin, and grinding, polishing, corroding, washing with alcohol and drying with cold air the embedded sample.

Description

Image segmentation micro-organization identification method based on convolutional neural network model Technical Field The invention relates to the field of image recognition, in particular to an image segmentation micro-organization recognition method based on a convolutional neural network model. Background The convolutional neural network model is widely applied in the field of micro-organization image recognition, and the micro-organization images aiming at the service turbine blades are usually processed by adopting the convolutional neural network model to distinguish the areas of the strengthening phase and the matrix phase, but in practical application, the control of a single model learning process is difficult to meet the practical training due to the randomness of the sample acquisition condition, so that how to improve the training precision of the convolutional neural network model is a technical problem to be solved urgently by a person in the field. Chinese patent publication No. CN120927719A discloses a multi-scale polycrystalline superalloyThe quantitative characterization method of the phase comprises the steps of sequentially carrying out metallographic phase sample preparation, electrolytic polishing and phase analysis on a standard superalloy sampleIn-situ electrolysis etching, microscopic electron image is collected, image processing and multi-scale processing are carried outPhase labeling, data augmentation, obtaining training sample, utilizing training sample pairPerforming iterative training on the convolutional neural network architecture to obtain a multi-scale structureExtracting phase characteristics from the surface of the polycrystalline superalloy to be measuredPhase secondary electron image input to multiscaleIn the phase feature extraction model, binarization is obtainedCarrying out statistical distribution characterization on the phase characteristic images to obtain eachStatistical distribution data of phases. Therefore, the technical scheme has the following problems that sample processing and model training cannot be adaptively adjusted according to the actual sample acquisition condition, so that the model training effect is difficult to meet the training requirement. Disclosure of Invention Therefore, the invention provides an image segmentation micro-organization identification method based on a convolutional neural network model, which is used for solving the problem that in the prior art, the model training effect is difficult to meet the training requirement because the image processing and the model training cannot be adaptively adjusted according to the acquisition condition of an actual acquired sample. In order to achieve the above object, the present invention provides an image segmentation micro-organization identification method based on a convolutional neural network model, comprising: Sampling each target position corresponding to the turbine blade to obtain a plurality of samples, and respectively carrying out image acquisition on each sample after material treatment to obtain a plurality of target images; Respectively carrying out noise processing on each target image, determining an image characteristic state according to the contrast uniformity value and the local difference value of each target image subjected to the noise processing, and determining a final threshold value corresponding to each target image based on the pixel gray value corresponding to the inter-class variance or the sub-region according to the image characteristic state; performing binarization processing on each noise processed target image based on the final threshold value, and taking the target image subjected to the binarization processing as a label image; And carrying out data set expansion based on the label image, carrying out data division on the final data set obtained by expansion, and training the Res_ Unet model through a training data set. Further, for image feature states where the contrast uniformity value is greater than a preset contrast uniformity value and the local variance value is less than or equal to a preset local variance value, a final threshold is determined based on the inter-class variance. Further, the determining the final threshold based on the inter-class variance includes: And obtaining a plurality of thresholds to be selected, dividing the target image based on the thresholds to be selected respectively, calculating the inter-class variance between the foreground and the background of the divided target image, and selecting the threshold to be selected with the maximum inter-class variance as a final threshold. Further, for image feature states where the contrast uniformity value is less than or equal to a preset contrast uniformity value or where the local variance value is greater than a preset local variance value, a final threshold is determined based on the pixel gray values corresponding to the sub-regions. Further, determining a final