Image segmentation microstructure identification method based on convolutional neural network model

Abstract

The invention relates to the field of image recognition, in particular to an image segmentation microstructure recognition method based on a convolutional neural network model, and the method comprises the steps: carrying out the sampling of a turbine blade, obtaining a plurality of samples, and carrying out the image collection of each sample after material processing, so as to obtain a plurality of target images; performing noise processing on each target image, determining an image feature state according to a contrast uniformity value and a local difference value of each target image after noise processing, and determining a final binarization threshold corresponding to each target image according to a pixel gray value corresponding to an inter-class variance or a sub-region based on the image feature state; performing threshold segmentation marking on the target image after binarization processing to obtain a label image; the ResUnet model is trained through the training data set; according to the invention, adaptive adjustment is carried out on sample processing and model training according to the actual collection condition of the collected sample, and the model training effect is improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to an image segmentation and micro-tissue recognition method based on a convolutional neural network model. Background Technology

[0002] Convolutional neural network models are widely used in the field of microstructure image recognition. For microstructure images of turbine blades in service, convolutional neural network models are often used to process the images to distinguish between the reinforcement phase and the matrix phase. However, in practical applications, the randomness of sample collection often makes it difficult to control the learning process of a single model to meet the actual training requirements. Therefore, how to improve the training accuracy of convolutional neural network models is a technical problem that needs to be solved by those skilled in the art.

[0003] Chinese patent publication CN120927719A discloses a multi-scale polycrystalline high-temperature alloy. A quantitative characterization method for phases, comprising: sequentially performing metallographic sample preparation, electrolytic polishing, and... In-situ electrolytic etching was performed to acquire microelectronic images, followed by image processing and multi-scale analysis. Phase labeling and data augmentation are used to obtain training samples; the training samples are then used to... The convolutional neural network architecture undergoes iterative training to obtain multi-scale... Phase feature extraction model; to extract the surface features of the polycrystalline superalloy under test. Phase secondary electronic images, input to multi-scale In the phase feature extraction model, binarization is obtained. Phase feature images are statistically distributed to obtain each The statistical distribution data of the phases. Therefore, it is evident that the above technical solution has the following problems: it cannot adaptively adjust sample processing and model training based on the actual sample collection situation, resulting in the model training effect failing to meet training requirements. Summary of the Invention

[0004] To address this issue, the present invention provides an image segmentation and micro-tissue recognition method based on a convolutional neural network model, which overcomes the problem in existing technologies that cannot adaptively adjust image processing and model training according to the actual collection conditions of the collected samples, resulting in the model training effect failing to meet training requirements.

[0005] To achieve the above objectives, this invention provides an image segmentation and micro-tissue recognition method based on a convolutional neural network model, comprising: Samples were taken from each target location corresponding to the turbine blade to obtain several samples, and images were acquired from each sample after material processing to obtain several target images. Noise processing is performed on each target image separately. The image feature state is determined based on the contrast uniformity and local difference value of each noise-processed target image. The final threshold for each target image is determined based on the pixel gray value corresponding to the inter-class variance or sub-region based on the image feature state. Based on the final threshold, the target images after noise processing are binarized, and the binarized target images are used as label images. The dataset is augmented based on labeled images, and the resulting dataset is partitioned. The Res_Unet model is then trained using the training dataset.

[0006] Furthermore, for image feature states where the contrast uniformity value is greater than the preset contrast uniformity value and the local difference value is less than or equal to the preset local difference value, the final threshold is determined based on the inter-class variance.

[0007] Furthermore, determining the final threshold based on inter-class variance includes: Several thresholds to be selected are obtained, and the target image is segmented based on each threshold. The inter-class variance between the foreground and background of the segmented target image is calculated. The threshold that maximizes the inter-class variance is selected as the final threshold.

[0008] Furthermore, for image feature states where the contrast uniformity value is less than or equal to the preset contrast uniformity value or the local difference value is greater than the preset local difference value, the final threshold is determined based on the pixel grayscale value corresponding to the sub-region.

[0009] Furthermore, the final threshold is determined based on the pixel grayscale values ​​corresponding to the sub-regions, including: The target image is divided into several sub-regions by local uniform division. The average value of the pixel grayscale value corresponding to each sub-region is recorded as the final threshold for each sub-region.

[0010] Furthermore, the cross-entropy loss function corresponding to the training process of the Res_Unet model is: , in, For the first The label value corresponding to each pixel. For the first The label probability corresponding to each pixel This represents the total number of pixels.

[0011] Furthermore, regarding the first For each pixel, the corresponding enhancement phase probability is output according to the Sigmoid activation function, which is: , in, For the first The enhancement phase probability corresponding to each pixel. For the first The pixel value corresponding to each pixel.

[0012] Furthermore, during the training of the Res_Unet model, the learning rate is set according to the number of training epochs. The formula for calculating the learning rate is: , in, For learning rate, As the baseline number of rounds, For the number of learning rounds.

[0013] Furthermore, if the image feature diversity is less than the preset image feature diversity, the number of reference rounds will be reduced. The reduction in the baseline round number is negatively correlated with the image feature diversity.

[0014] Furthermore, the material processing includes mounting the sample with cold mounting resin, and then grinding, polishing, etching, rinsing with alcohol, and drying the mounted sample with cold air.

[0015] Compared with the prior art, the beneficial effect of the present invention is that the technical solution of the present invention reflects the distribution difference of gray values ​​in the target image by using the contrast uniformity value and the local difference value, and selects different methods to determine the final threshold accordingly, so as to improve the accuracy of the final threshold determination, thereby improving the accuracy of subsequent image processing, and thus improving the recognition effect of the model of the present invention on the micro-structure of turbine blades.

[0016] Furthermore, the technical solution of this invention uses the cross-entropy loss function in conjunction with the Sigmoid activation function to provide a stable and well-gradient optimization objective for binary classification and segmentation tasks. Moreover, the learning rate can be dynamically adjusted, which is beneficial for fine-tuning the model parameters and converging to a better local optimum.

[0017] Furthermore, in the technical solution of the present invention, the baseline number of learning rounds is adaptively adjusted according to the diversity of image features, so that the setting of the learning rate is more in line with the actual image features, thereby saving computing resources and time costs. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the image segmentation and micro-tissue recognition method based on a convolutional neural network model according to the present invention; Figure 2 This is a flowchart illustrating the process of determining the final threshold for each target image based on the pixel grayscale value corresponding to the inter-class variance or sub-region according to the image feature state of the present invention. Figure 3 This is a flowchart illustrating the process of determining whether to reduce the baseline round number based on the diversity of image features according to the present invention. Detailed Implementation

[0019] To make the objectives and advantages of this invention clearer, the invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0021] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0022] Please see Figures 1 to 3 As shown, this invention provides an image segmentation and micro-tissue recognition method based on a convolutional neural network model, comprising: Samples were taken from each target location corresponding to the turbine blade to obtain several samples, and images were acquired from each sample after material processing to obtain several target images. Noise processing is performed on each target image separately. The image feature state is determined based on the contrast uniformity and local difference value of each noise-processed target image. The final threshold for each target image is determined based on the pixel gray value corresponding to the inter-class variance or sub-region based on the image feature state. Based on the final threshold, the target images after each noise processing step are binarized to obtain the label images; The dataset is augmented based on labeled images, and the resulting dataset is partitioned. The Res_Unet model is then trained using the training dataset.

[0023] The specific application of this invention is to use a convolutional neural network model to distinguish between the reinforcing phase and the matrix phase in the microstructure of a turbine blade in service. The network architecture of the Res_Unet model includes a left encoder, a central connecting bridge, and a right decoder. For the left encoder, it consists of four cascaded residual encoder blocks (REBs). The internal process of each REB is 3×3 convolution → BN → ReLU → 3×3 convolution → BN. The feature map is added pixel-by-pixel to the input identity mapping, and finally downsampled by a 3×3 convolution with stride=2, which halves the feature map size and doubles the number of channels.

[0024] The connection bridge uses two standard residual blocks (RBs) stacked together, with the number of channels maintained at 512 and the feature map size maintained at 32×32, to further encode global context information and smoothly transition to the decoding stage.

[0025] The right-side decoder consists of four residual decoder blocks (RDBs). Each RDB first doubles the feature map size using a 3×3 transposed convolution (stride=2); then, it concatenates the high-resolution features from the corresponding REBs of the encoder along the channel dimension to form a hybrid "semantic + detail" feature; finally, it is refined by two 3×3 convolution-BN-ReLU residual units to gradually restore the spatial resolution. After four upsampling operations, the feature map is restored to 512×512, and the number of channels is reduced to 64.

[0026] In addition, the model uses 1×1 convolution + Softmax to generate binary probability maps, corresponding to pixel-level confidence of the reinforcement phase and the matrix phase. The total number of network parameters is 31.7 M, and the cross-entropy loss function is used for supervision during the training phase.

[0027] Specifically, for image feature states where the contrast uniformity value is greater than the preset contrast uniformity value and the local difference value is less than or equal to the preset local difference value, the final threshold is determined based on the inter-class variance.

[0028] For a single target image, the formula for calculating its contrast uniformity is: , in, ,in, For the first The average gray value of all pixels corresponding to an initial local region. The average of the average grayscale values ​​of all pixels corresponding to the entire initial local region. The initial local region is the total number of regions. The initial local region is the target image that is uniformly divided into several regions (e.g., 16x16 or 32x32 pixels). The divided regions are recorded as the initial local regions. The number of initial local regions is set by the user. The higher the user's requirement for the accuracy of the contrast uniformity value, the larger the number of initial local regions. Preferably, the target image is uniformly divided into several 16x16 initial local regions. The local difference value is the absolute value of the difference between the maximum and minimum values ​​of the average gray values ​​corresponding to the initial local regions.

[0029] Regarding the preset uniform contrast value and preset local difference value, it is known that the larger the values ​​of the uniform contrast value and the local difference value, the greater the difference in the distribution of gray values ​​in the target image, and the greater the impact on the accuracy of determining the final threshold based on inter-class variance. Therefore, the smaller the user's tolerance for the difference in the distribution of gray values ​​and the higher the accuracy requirement for the final threshold setting, the larger the value of the preset uniform contrast value and the smaller the value of the preset local difference value. One approach is to extract the uniform contrast value and the local difference value from the parameter statistics, and record the average values ​​of the uniform contrast value and the local difference value as the preset uniform contrast value and the preset local difference value, respectively.

[0030] Specifically, determining the final threshold based on inter-class variance includes: Several thresholds to be selected are obtained, and the target image is segmented based on each threshold. The inter-class variance between the foreground and background of the segmented target image is calculated. The threshold that maximizes the inter-class variance is selected as the final threshold.

[0031] The inter-class variance between foreground and background = foreground pixel ratio × background pixel ratio × (mean foreground pixel - mean background pixel)².

[0032] Foreground pixel ratio = area of ​​foreground region / total area of ​​target image; background pixel ratio = 1 - foreground pixel ratio.

[0033] It is understood that the present invention sets the selectable range of grayscale values ​​to [0, 255], and the threshold values ​​to be selected are grayscale values ​​within [0, 255]. The threshold values ​​to be selected are sorted from small to large to obtain a threshold sequence. The absolute value of the difference between any two adjacent threshold values ​​to be selected within the threshold sequence is the same. The number of threshold values ​​to be selected is set by the user. The higher the user's requirement for the accuracy of the final threshold determination, the larger the number of threshold values ​​to be selected. One set of values ​​is provided, and the threshold values ​​to be selected include 20, 40, 80, 120, 160, 200, and 240. Obtaining the threshold values ​​to be selected through historical experience and multiple experiments is also a subject of knowledge for those skilled in the art, and will not be elaborated here.

[0034] Specifically, after segmenting the target image based on a single threshold to be selected, the foreground of the segmented target image is the region in the segmented target image where the pixel value is greater than the threshold to be selected, and the background of the segmented target image is the region in the segmented target image where the pixel value is less than or equal to the threshold to be selected.

[0035] Specifically, for image feature states where the contrast uniformity value is less than or equal to the preset contrast uniformity value or the local difference value is greater than the preset local difference value, the final threshold is determined based on the pixel grayscale value corresponding to the sub-region.

[0036] Specifically, the final threshold is determined based on the pixel grayscale values ​​corresponding to the sub-region, including: The target image is divided into several sub-regions by local uniform division. The average value of the pixel grayscale value corresponding to each sub-region is recorded as the final threshold for each sub-region.

[0037] If the final threshold is determined based on the inter-class variance, then global binary processing is performed on the corresponding target image based on the final threshold. If the final threshold is determined based on the pixel grayscale value corresponding to the sub-region, then the target image is binarized based on the corresponding final threshold for each sub-region and the images are stitched together to obtain the binarized target image.

[0038] Specifically, the cross-entropy loss function used during the training of the Res_Unet model is: , in, For the first The label value corresponding to each pixel. For the first The label probability corresponding to each pixel The total number of pixels is denoted as . For a single pixel, the label value is determined as follows: if the pixel actually belongs to the strengthening phase, the label value is 1; if the pixel actually belongs to the matrix phase, the label value is 0.

[0039] Specifically, regarding the first For each pixel, the corresponding enhancement phase probability is output according to the Sigmoid activation function, which is: , in, For the first The enhancement phase probability corresponding to each pixel. For the first The pixel value corresponding to each pixel.

[0040] Specifically, during the training of the Res_Unet model, the learning rate is set according to the number of training epochs. The formula for calculating the learning rate is: , in, For learning rate, As the baseline number of rounds, The number of learning rounds is the number of times the entire training set has been fully traversed in the current iteration. The number of baseline rounds is a preset value, with an initial preset value of 100.

[0041] Specifically, if the image feature diversity is less than the preset image feature diversity, the number of reference rounds will be reduced. The reduction in the baseline round number is negatively correlated with the image feature diversity.

[0042] The method for confirming image feature diversity is to detect the average pixel value corresponding to each target image, and record the absolute value of the difference between the maximum and minimum values ​​of the average pixel value as the image feature diversity.

[0043] The preset image feature diversity value can be understood as follows: the greater the image feature diversity, the richer the sample, and the lower the probability of the model converging quickly. Therefore, the lower the user's tolerance for rapid model convergence, the higher the preset image feature diversity. A method is provided to extract the image feature diversity from the parameter statistics and record the average value of the image feature diversity as the preset image feature diversity.

[0044] When adjusting the base number of rounds, the adjusted base number of rounds = 100 - t, where t = preset image feature diversity / image feature diversity, and the calculated result of t is an integer rounded up.

[0045] Specifically, the material processing includes mounting the sample with cold mounting resin, and then grinding, polishing, etching, rinsing with alcohol, and drying the mounted sample with cold air.

[0046] The target locations include the blade regions corresponding to the blade root, blade middle, and blade tip of the turbine blade. The blade regions include the leading edge, trailing edge, blade base, and blade back. Image acquisition is achieved using a scanning electron microscope (SEM). The number of images acquired is set by the user. The greater the user's requirement for model training accuracy, the greater the number of target images acquired.

[0047] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for image segmentation and micro-tissue recognition based on a convolutional neural network model, characterized in that, include: Samples were taken from each target location corresponding to the turbine blade to obtain several samples, and images were acquired from each sample after material processing to obtain several target images. Noise processing is performed on each target image separately. The image feature state is determined based on the contrast uniformity and local difference values ​​of each noise-processed target image. The final threshold for each target image is determined based on the pixel gray value corresponding to the inter-class variance or sub-region based on the image feature state. Based on the final threshold, the target images after noise processing are binarized, and the binarized images are used as label images. The dataset is augmented based on labeled images, and the resulting dataset is partitioned. The Res_Unet model is then trained using the training dataset.

2. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 1, characterized in that, For image feature states where the contrast uniformity is greater than the preset contrast uniformity and the local difference is less than or equal to the preset local difference, the final threshold is determined based on the inter-class variance.

3. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 2, wherein determining the final threshold based on inter-class variance includes: Several thresholds to be selected are obtained, and the target image is segmented based on each threshold. The inter-class variance between the foreground and background of the segmented target image is calculated. The threshold that maximizes the inter-class variance is selected as the final threshold.

4. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 1, characterized in that, For image feature states where the contrast uniformity value is less than or equal to the preset contrast uniformity value or the local difference value is greater than the preset local difference value, the final threshold is determined based on the pixel gray value corresponding to the sub-region.

5. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 4, characterized in that, The final threshold is determined based on the pixel grayscale values ​​corresponding to the sub-regions, including: The target image is divided into several sub-regions by local uniform division. The average value of the pixel grayscale value corresponding to each sub-region is recorded as the final threshold for each sub-region.

6. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 1, characterized in that, The cross-entropy loss function used during the training of the Res_Unet model is: ， in, For the first The label value corresponding to each pixel. For the first The label probability corresponding to each pixel This represents the total number of pixels.

7. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 6, characterized in that, Regarding the first For each pixel, the corresponding enhancement phase probability is output according to the Sigmoid activation function, which is: ， in, For the first The enhancement phase probability corresponding to each pixel. For the first The pixel value corresponding to each pixel.

8. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 7, characterized in that, During the training of the Res_Unet model, the learning rate is set according to the number of training epochs. The formula for calculating the learning rate is: ， in, For learning rate, As the baseline number of rounds, To learn the number of rounds.

9. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 8, characterized in that, If the image feature diversity is less than the preset image feature diversity, the number of reference rounds will be reduced. The reduction in the baseline round number is negatively correlated with the image feature diversity.

10. The image segmentation and micro-tissue recognition method based on a convolutional neural network model according to claim 1, characterized in that, The material processing includes mounting the sample with cold mounting resin, and then grinding, polishing, etching, rinsing with alcohol, and drying the mounted sample with cold air.

Images