Pinctada martensii internal pearl recognition image processing method
Through the combination of dynamic distance data partitioning module, multi-dimensional Gaussian probability density modeling and region weighting enhancement module, the problem of difficult identification of pearls inside the Martha pearl beads is solved, and higher image processing accuracy and recognition robustness are achieved.
Patent Information
- Application Number
- CN202510725916.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-03
- Publication Date
- 2025-07-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The traditional internal pearl recognition image processing method of Mahayana's mother-of-pearl is difficult to adapt to the complex texture and morphological characteristics of the mother-of-pearl, resulting in low accuracy of the recognition results, especially in the case of poor image quality or high background noise, the recognition effect is poor.
A method of internal pearl recognition image processing for Mahayana mother beads is proposed. By introducing a dynamic distance data partition module and a regional center reconstruction strategy, the image area is accurately divided and weighted according to the regional characteristics. At the same time, a multi-dimensional Gaussian probability density modeling and semantic weight calculation module is designed to capture the high-order feature and structural information of the image, and non-linear weighting is performed through the region weighting enhancement module.
It significantly improves the accuracy and effect of image processing, improves the semantic consistency and boundary clarity of partitioned images, and can restore more image details in low resolution or complex backgrounds, improving the accuracy and robustness of recognition.
Smart Images
Figure CN120235802A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image enhancement, and particularly relates to an image processing method for identifying pearls inside Pinctada martensii. Background Art
[0002] Traditional image processing methods for identifying pearls inside Pinctada martensii mostly rely on classical image processing techniques such as threshold segmentation and edge detection. These methods usually rely on simple image features and are difficult to adapt to the complex texture and morphological features inside the pearl oyster. They lack effective region division and feature enhancement mechanisms, and often cannot accurately distinguish pearls from the background, resulting in low accuracy of the recognition results. Especially in the case of poor image quality or more background noise, the recognition effect is greatly reduced.
[0003] Although some current image enhancement and object recognition methods based on deep learning have made certain progress, there are still some areas for improvement. Existing methods usually only focus on pixel-level classification or segmentation, ignoring the semantic associations and structural information between regions. As a result, in the case of complex internal structures and large texture variations in Pinctada martensii, it is difficult to provide accurate region segmentation. Many methods lack effective modeling of multi-scale and non-linear features when processing images, resulting in limited image enhancement effects. Especially when processing low-resolution or blurred images, they often cannot recover enough details.
[0004] The present invention proposes an image processing method for identifying pearls inside Pinctada martensii, significantly improving the accuracy and effect of image processing. By introducing a dynamic distance data partitioning module and a regional center reconstruction strategy, the image regions can be accurately divided and weighted according to regional characteristics, improving the semantic consistency and boundary clarity of the partitioned image. The designed multi-dimensional Gaussian probability density modeling and semantic weight calculation module can better capture the high-order features and structural information of the image, making the enhanced image results more detailed. Especially in the case of low resolution or complex background, more image details can be restored, improving the accuracy and robustness of recognition. Summary of the Invention
[0005] The present invention provides an image processing method for identifying pearls inside Pinctada martensii, aiming to improve the accuracy and robustness of image recognition through region partitioning, semantic weight calculation, and non-linear weighting strategies.
[0006] The present invention aims to propose an image enhancement model for identifying pearls inside Pinctada martensii and provides an image processing method for identifying pearls inside Pinctada martensii, including the following steps: S1. Collect images related to Pinctada martensii pearls to generate an original dataset for identifying pearls inside Pinctada martensii; S2. Define a symmetric perturbation distance metric function, introduce a regional center reconstruction strategy, and construct a dynamic distance data partitioning module. The internal pearl recognition image of Pinctada martensii is partitioned by the dynamic distance data partitioning module to obtain a partitioned internal pearl recognition image of Pinctada martensii. S3. Design a multi-dimensional Gaussian probability density modeling, construct a semantic weight calculation module, and the partitioned internal pearl recognition image of Pinctada martensii obtains regional semantic weights through the semantic weight calculation module. S4. Introduce a regional semantics and non-linear weighting strategy, construct a regional weighted enhancement module, and the internal pearl recognition image of Pinctada martensii obtains an enhanced internal pearl image of Pinctada martensii through the regional weighted enhancement module. S5. Integrate the dynamic distance data partitioning module, the semantic weight calculation module, and the regional weighted enhancement module to construct an internal pearl recognition image enhancement model of Pinctada martensii. S6. Input the internal pearl recognition image of Pinctada martensii into the internal pearl recognition image enhancement model of Pinctada martensii for image enhancement, and output a high-resolution internal pearl recognition image of Pinctada martensii.
[0007] Preferably, in S2, constructing the dynamic distance data partitioning module specifically includes the following steps: Step S21. Define a symmetric perturbation distance metric function , calculate the similarity distance between the pixel and the regional center . The mathematical model is: ; Among them, is the metric function, is the pixel value of the image at the position , is the central position of the th region, initialized by random sampling, is the perturbation sensitive gradient of the pixel value to the regional center, is for partial derivative calculation, is the angle between the pixel vector and the regional center, is the pixel vector dimension, is the contribution ratio of the direction difference term, controlling the influence intensity of the perturbation gradient term. Designing the symmetric perturbation distance metric function can accurately calculate the similarity between the pixel and the regional center, which helps to improve the partitioning accuracy.
[0008] Step S22. Introduce a regional center reconstruction strategy, and update the center of each region based on the symmetric perturbation distance metric function. The mathematical model is: ; Among them, is the temperature parameter, which controls the smoothness of the weight. The smaller the value, the more attention is paid to the low-distance pixels. is the th region, is the number of pixels in the th region, is the current number of regions, is the updated region center of the th region. By introducing the region center reconstruction strategy, the focusing ability of the center on low-disturbance distance pixels is strengthened.
[0009] Step S23: Based on each updated region center, assign each pixel of the image to the corresponding region to obtain the region division result of the image. The mathematical model is: ; where, represents the region to which the pixel is assigned, , represents that the pixel is assigned to the region with the minimum similarity value.
[0010] Step S24: Iteratively execute Steps S21 to S23, continuously optimize the region center and pixel assignment until the change of all region centers is less than the set threshold to achieve the final region division of the image. The mathematical model is: ; where represents the preset convergence threshold, represents the number of iterations.
[0011] Preferably, through the dynamic distance data partitioning module constructed by the above S2, in S21, by defining the symmetric perturbation distance metric function, the similarity between the pixel and the region center can be accurately calculated, avoiding the error accumulation of the traditional method and improving the segmentation accuracy; in S22, by introducing the region center reconstruction strategy, the region center can be adaptively updated according to the pixel distribution, improving the robustness and accuracy of the region division; in S23, by assigning the pixel to the most similar region, the consistency and accuracy of the region division are ensured; the iterative optimization process in S24 can further refine the region segmentation until convergence, ensuring the optimality of the partitioning result; overall, this module significantly improves the accuracy and robustness of the image partitioning through the dynamic adjustment of the region center and the fine pixel assignment strategy.
[0012] Preferably, in S3, constructing the semantic weight calculation module specifically includes the following steps: Step S31: Design a multi-dimensional Gaussian probability density model to calculate the likelihood value of the pixels in each region. The mathematical model is: ; Among them, is the covariance matrix of the th region, which controls the ellipsoidal direction and expansion degree of the Gaussian function, represents the generation likelihood value of pixel under all regions, and is an important indicator of the generative discriminative ability.
[0013] Step S32: Calculate the semantic weight of each region based on the generation likelihood value of the pixels in the region. The mathematical model is: ; Among them, represents the number of pixels in the th region, represents the proportion of the number of region pixels, represents the calculation of taking the exponential of the negative entropy, represents the region under the semantic weight.
[0014] Preferably, through the semantic weight calculation module constructed by the above S3, in S31, the multi-dimensional Gaussian probability density modeling can accurately characterize the distribution characteristics of the pixels in each region, improving the modeling ability of the image structure information; in S32, the semantic weight of the region is calculated based on the generation likelihood value of the pixels, making the weight allocation more semantically consistent and discriminative; overall, this module effectively enhances the semantic credibility and discrimination ability of the region expression through the modeling of the pixel distribution in the region and the entropy-aware weighting strategy.
[0015] Preferably, in S4, constructing the region weighted enhancement module specifically includes the following steps: Step S41: Introduce the sine function curve change and the distance metric from the pixel to the region center, and calculate the distribution perception weight of the pixels in the region in combination with the region semantic weight. The mathematical model is: ; Among them, is the distribution perception weight of pixel , is the th pixel in the region.
[0016] Step S42: Use the distribution perception weight to weight pixel to obtain the enhanced recognition image of the pearls inside the Pinctada martensii, and the mathematical model is: ; Among them, is the enhanced recognition image of the pearls inside the Pinctada martensii.
[0017] Preferably, in the regional weighted enhancement module constructed by S4, in S41, by introducing the combined calculation of the sine function curve change and the regional semantic weight to calculate the distribution perception weight, the response of pixels to the regional features is made smoother and more discriminative; in S42, the distribution perception weight is used to perform weighted enhancement on the image, so that the semantic key regions are further highlighted on the basis of maintaining the original structure of the image; overall, this module realizes local enhancement by fusing spatial distribution and semantic weight, effectively improving the clarity of the image and the saliency of target recognition.
[0018] Preferably, in S5, constructing an internal pearl recognition image enhancement model for Pinctada martensii specifically includes the following steps: Step S51, the original internal pearl recognition image of Pinctada martensii passes through the dynamic distance data partitioning module to obtain the partitioned data of the internal pearl recognition image of Pinctada martensii. The mathematical model is: ; Among them, is the original internal pearl recognition image of Pinctada martensii, represents the partitioned data of the internal pearl recognition image of Pinctada martensii, represents the dynamic distance data partitioning module.
[0019] Step S52, the original internal pearl recognition image of Pinctada martensii and the partitioned data of the internal pearl recognition image of Pinctada martensii pass through the semantic weight calculation module to obtain the regional semantic weight distribution of the internal pearl recognition image of Pinctada martensii. The mathematical model is: ; Among them, is the regional semantic weight distribution of the internal pearl recognition image of Pinctada martensii, is the semantic weight calculation module.
[0020] Step S53, the original internal pearl recognition image of Pinctada martensii and the regional semantic weight distribution of the internal pearl recognition image of Pinctada martensii pass through the regional weighted enhancement module to obtain the enhanced internal pearl recognition image of Pinctada martensii. The mathematical model is: ; Among them, is the enhanced internal pearl recognition image of Pinctada martensii, is the regional weighted enhancement module.
[0021] In summary, due to the adoption of this technical solution, compared with the prior art, the beneficial effects of the present invention are as follows: The use of the dynamic distance data partitioning module to partition the original image helps to extract the structural information and regional features of the image; by the semantic weight calculation module, the original image and the partition data are fused to obtain a more discriminative semantic weight distribution, improving the semantic separation degree between regions; through the region weighted enhancement module, the image is enhanced under the guidance of weights, making the key regions more prominent and the boundaries clear; overall, through the integrated design of partitioning, semantic modeling, and weighted enhancement, the model realizes the precise enhancement and semantic strengthening of the internal pearl recognition image of Pinctada martensii. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 It is a flowchart of the method for processing the internal pearl recognition image of Pinctada martensii.
[0023] Figure 2 It is a structural diagram of the dynamic distance data partitioning module.
[0024] Figure 3 It is a structural diagram of the semantic weight calculation module.
[0025] Figure 4 It is a structural diagram of the region weighted enhancement module.
[0026] Figure 5 It is a general structural diagram of the internal pearl recognition image enhancement model of Pinctada martensii.
[0027] Figure 6 It is a comparison diagram of the internal pearl recognition images of Pinctada martensii before and after enhancement. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0028] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.
[0029] Please refer to the attached Figure 1 - attached Figure 6 , the present invention provides a method for processing the internal pearl recognition image of Pinctada martensii.
[0030] As shown in the attached Figure 1 flowchart, the present invention proposes a method for processing the internal pearl recognition image of Pinctada martensii. The specific implementation method includes the following steps: S1. Collect images related to Pinctada martensii pearls to generate an original dataset of internal pearl recognition images of Pinctada martensii.
[0031] Further, as described in S1 in the appendix Figure 1 Collect the internal pearl recognition image data of Pinctada martensii, specifically, keep the shell position stable, control the light source, and use a high-resolution industrial camera device to non-destructively photograph the complete shell to generate an internal pearl recognition image dataset of Pinctada martensii, with a total of 652 image data. The dataset format is PNG, and each image is uniformly preprocessed to a size of 224×224 pixels.
[0032] S2. Define a symmetric perturbation distance metric function, introduce a regional center reconstruction strategy, and construct a dynamic distance data partitioning module. The internal pearl recognition image of Pinctada martensii is partitioned by the dynamic distance data partitioning module to obtain the partitioned internal pearl recognition image of Pinctada martensii.
[0033] Further, as described in S2 in the appendix Figure 1 Construct the dynamic distance data partitioning module. The specific construction steps of the module are as shown in the appendix Figure 2 and are as follows: Further, step S21. Define a symmetric perturbation distance metric function and calculate the similarity distance between a pixel and the regional center. The mathematical model is: ; where is the pixel value at position of the image, is the center position of the th region, initialized by random sampling, is the perturbation sensitivity gradient of the pixel value to the regional center, is for partial derivative calculation, is the angle between the pixel vector and the regional center, with an initial value of 0, is the pixel vector dimension, with a value of 3, is the contribution ratio of the direction difference term, set in the range of 0.1 - 1, with an initial value of 0.1, and changes with the change of the pixel center point, is the symmetric perturbation distance metric function, and the implementation code is: def symmetric_perturbation_distance(I, mu_k, grad_I, theta, lam=0.1): """ Calculate the symmetric perturbation distance metric function Args: I: (B, C, H, W) Input image mu_k: (K, C) Regional center grad_I: Pixel gradient of (B, C, H, W) theta: Angle between pixel and region center of (B, K, H, W) lam: Weight of λ direction difference Returns: D: Perturbation distance of (B, K, H, W) """ B, C, H, W = I.shape K = mu_k.shape[0] I_exp = I.unsqueeze(1) # (B, 1, C, H, W) mu_k_exp = mu_k.view(1, K, C, 1, 1) # (1, K, C, 1, 1) grad_exp = grad_I.unsqueeze(1) # (B, 1, C, H, W) diff = torch.abs(I_exp - mu_k_exp) * torch.abs(grad_exp) log_term = torch.log(1 + diff) log_sum = log_term.sum(dim=2) direction_term = lam * torch.sin(theta) ** 2 D = log_sum + direction_term return D
[0034] Furthermore, in step S22, a region center reconstruction strategy is introduced to update the center of each region based on the symmetric perturbation distance metric function. The mathematical model is as follows: ; Wherein, is the temperature parameter, which controls the smoothness of the weight. The initial value is set to 0.1 and gradually decreases according to the convergence situation. is the th region. is the th number of pixels in the region. is the current number of regions, set to 3. is the th standard deviation of the region. It can be set to 1 initially, representing the uniformity of the region pixel values. For the updated region center of the th region, the implementation code is: def update_region_center(I, D, mu_k, sigma_k, tau): """ Update the region center mu_k' Args: I: (B, C, H, W) Input image D: (B, K, H, W) Distance from pixels to each region center mu_k: (K, C) Current region center sigma_k: (K, C) Region standard deviation tau: Temperature parameter (float) Returns: mu_k_new: (K, C) Updated region center """ B, C, H, W = I.shape K = mu_k.shape[0] I_flat = I.view(B, C, -1).permute(0, 2, 1) # (B, H*W, C) D_flat = D.view(B, K, -1).permute(0, 2, 1) # (B, H*W, K) # softmax weight calculation weight = torch.softmax(-D_flat / tau, dim=2) # (B, H*W, K) mu_k_exp = mu_k.view(1, 1, K, C) # (1, 1, K, C) sigma_k_exp = sigma_k.view(1, 1, K, C) # (1, 1, K, C) I_exp = I_flat.unsqueeze(2) # (B, H*W, 1, C) response = torch.tanh((I_exp - mu_k_exp) / (sigma_k_exp + 1e-6)) #(B, H*W, K, C) weighted_response = weight.unsqueeze(-1) * response # (B, H*W, K,C) sum_response = weighted_response.sum(dim=1) # (B, K, C) count = weight.sum(dim=1).unsqueeze(-1) + 1e-6 # (B, K, 1) mu_k_new = (sum_response / count).mean(dim=0) # (K, C) return mu_k_new。
[0035] Furthermore, in step S23, based on each updated regional center, each pixel of the image is assigned to the corresponding region to obtain the regional division result of the image. The mathematical model is as follows: ; wherein, represents the region to which the pixel (i, j) is assigned , means that the pixel is assigned to the region with the minimum similarity value.
[0036] Furthermore, in step S24, steps S21 to S23 are iteratively executed to continuously optimize the regional center and pixel assignment until the change of all regional centers is less than the set threshold, so as to achieve the final regional division of the image. The mathematical model is as follows: ; where represents the preset convergence threshold, which is set to 0.1, represents the number of iterations, and the initial value is 1.
[0037] S3. Design a multi-dimensional Gaussian probability density model and construct a semantic weight calculation module. After partitioning, the internal pearl recognition image of Pinctada martensii passes through the semantic weight calculation module to obtain the regional semantic weight.
[0038] Furthermore, as described in S3 of the appendix Figure 3 The specific construction steps of the constructed semantic weight calculation module are as shown in the appendix Figure 3 The specific implementation of the module includes the following steps: Furthermore, in step S31, design a multi-dimensional Gaussian probability density model and calculate the likelihood value of the pixels in each region. The mathematical model is as follows: ; wherein, For the covariance matrix of the th region, the initial value is set to the identity matrix. Denote the likelihood value generated by the pixel under all regions. The implementation code is as follows: def gaussian_likelihood(I, mu_k, sigma_k): """ Calculate the likelihood value generated by each pixel under each Gaussian region Args: I: (B, C, H, W) Input image mu_k: (K, C) Mean vector of each region (mu_k') sigma_k: (K, C, C) Covariance matrix of each region (can be initialized as the identity matrix) Returns: likelihood: (B, K, H, W) Probability density value of each pixel under each region """ B, C, H, W = I.shape K = mu_k.shape[0] I_flat = I.view(B, C, -1).permute(0, 2, 1).unsqueeze(2) # (B, H*W,1, C) mu_k_exp = mu_k.view(1, 1, K, C) # (1, 1, K, C) diff = I_flat - mu_k_exp # (B, H*W, K, C) sigma_inv = torch.inverse(sigma_k + 1e-6 * torch.eye(C).to(sigma_k.device)) # (K, C, C) sigma_det = torch.det(sigma_k + 1e-6 * torch.eye(C).to(sigma_k.device)) # (K,) # Mahalanobis distance diff_T = diff.unsqueeze(-2) # (B, H*W, K, 1, C) diff_B = diff.unsqueeze(-1) # (B, H*W, K, C, 1) sigma_inv_exp = sigma_inv.view(1, 1, K, C, C) # (1, 1, K, C, C) mahal = torch.matmul(torch.matmul(diff_T, sigma_inv_exp), diff_B).squeeze(-1).squeeze(-1) # (B, H*W, K) norm_const = torch.sqrt((2 * torch.pi) ** C * sigma_det).view(1, 1,K) likelihood = torch.exp(-0.5 * mahal) / (norm_const + 1e-8) return likelihood.permute(0, 2, 1).view(B, K, H, W) # (B, K, H, W).
[0039] Furthermore, in step S32, based on the likelihood values generated for pixel-based regions, calculate the semantic weights for each region. The mathematical model is as follows: ; where represents the number of pixels in the k-th region, represents the proportion of the number of regional pixels, represents the exponential calculation of negative entropy, represents the semantic weight under region k. The implementation code is as follows: def compute_semantic_weights(p_likelihood, region_mask, max_entropy=None): """ Calculate the semantic weight W_k for each region Args: p_likelihood: (B, K, H, W) The likelihood value generated for each pixel under each region region_mask: (B, K, H, W) Region assignment mask, where pixels belonging to region k are 1 and the rest are 0 max_entropy: float, the maximum entropy value in all regions (used for normalization) Returns: W: (B, K) regional semantic weight """ eps = 1e-8 B, K, H, W = p_likelihood.shape # Region size |C_k| region_size = region_mask.sum(dim=(2, 3)) + eps # (B, K) # Pixel negative entropy term: p * log(p) p = p_likelihood * region_mask # Only keep pixels belonging to the region neg_entropy = (p * torch.log(p + eps)).sum(dim=(2, 3)) # (B, K) # Maximum entropy value normalization if max_entropy is None: max_entropy = neg_entropy.max().item() + eps norm_entropy = neg_entropy / max_entropy # (B, K) # Region ratio term region_ratio = region_size / (region_size.sum(dim=1, keepdim=True)+ eps) # (B, K) # Final semantic weight W = region_ratio * torch.exp(-norm_entropy) # (B, K) return W.
[0040] S4. Introduce regional semantics and non-linear weighting strategy to construct a regional weighted reinforcement module, and the internal pearl recognition image of Pinctada martensii is enhanced by the regional weighted reinforcement module to obtain an enhanced internal pearl image of Pinctada martensii.
[0041] Furthermore, as described in S4 of the appendix Figure 4 The construction steps of the module for constructing the regional weighted reinforcement module are as described in the appendixFigure 4 As shown in the figure, the specific implementation of the module includes the following steps: Further, in step S41, a sine function curve change and the distance metric from the pixel to the center of the region are introduced, and the distribution perception weight of the pixels within the region is calculated by combining the regional semantic weight. The mathematical model is: ; where, is the distribution perception weight of pixel , which is dynamically calculated by the model, is the th pixel within the region.
[0042] Further, in step S42, the distribution perception weight is used to weight pixel to obtain the enhanced recognition image of the pearls inside the Pinctada martensii. The mathematical model is: ; where, is the enhanced recognition image of the pearls inside the Pinctada martensii.
[0043] S5. The dynamic distance data partitioning module, semantic weight calculation module, and regional weighting reinforcement module are integrated to construct an enhanced model for recognizing the pearls inside the Pinctada martensii.
[0044] Further, for the construction of the enhanced model for recognizing the pearls inside the Pinctada martensii described in S5 in the appendix Figure 5 , the specific construction steps of the model are as shown in appendix Figure 5 As shown in the figure, the specific implementation of the module includes the following steps: Further, in step S51, the original recognition image of the pearls inside the Pinctada martensii passes through the dynamic distance data partitioning module to obtain the partition data of the recognition image of the pearls inside the Pinctada martensii. The mathematical model is: ; where, is the original recognition image of the pearls inside the Pinctada martensii, represents the partition data of the recognition image of the pearls inside the Pinctada martensii, represents the dynamic distance data partitioning module.
[0045] Further, in step S52, the original recognition image of the pearls inside the Pinctada martensii and the partition data of the recognition image of the pearls inside the Pinctada martensii pass through the semantic weight calculation module to obtain the regional semantic weight distribution of the recognition image of the pearls inside the Pinctada martensii. The mathematical model is: ; where, For the semantic weight distribution of the internal pearl recognition image area of Pinctada martensii It is a semantic weight calculation module
[0046] Furthermore, in step S53, the original internal pearl recognition image of Pinctada martensii and the semantic weight distribution of the internal pearl recognition image area of Pinctada martensii pass through the region weighted enhancement module to obtain the enhanced internal pearl recognition image of Pinctada martensii. The mathematical model is: ; Wherein, is the enhanced internal pearl recognition image of Pinctada martensii, is the region weighted enhancement module
[0047] S6. Input the internal pearl recognition image of Pinctada martensii into the internal pearl recognition image enhancement model of Pinctada martensii for image enhancement, and output the high-resolution internal pearl recognition image of Pinctada martensii
[0048] Furthermore, as described in S6 of the attached Figure 1 The internal pearl recognition image enhancement model of Pinctada martensii inputs the internal pearl recognition image of Pinctada martensii and outputs the high-quality internal pearl recognition image of Pinctada martensii. As shown in the attached Figure 6 The specific implementation includes the following steps: Furthermore, the operating system platform used by the model in step S6 is the Linux system, the language is Python 3.9.1, the processor used is Jetson Xavier, the image processing library uses OpenCV and PIL, the physical memory of the server is 64G, and the dataset includes community care wound images under interference background, low resolution, and poor lighting conditions, totaling 652 images
[0049] The above is only the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention
Claims
1. An image processing method for identifying internal pearls of Pinctada martensii, characterized in that, It includes the following steps: S1. Collect images related to pearls of Pinctada martensii, and generate an original dataset of internal pearl recognition images of Pinctada martensii; S2. Define a symmetric perturbation distance metric function, introduce a regional center reconstruction strategy, and construct a dynamic distance data partitioning module. The internal pearl recognition image of Pinctada martensii passes through the dynamic distance data partitioning module to obtain the partitioned internal pearl recognition image of Pinctada martensii; S3. Design a multi-dimensional Gaussian probability density modeling, construct a semantic weight calculation module, and the partitioned internal pearl recognition image of Pinctada martensii passes through the semantic weight calculation module to obtain regional semantic weights; S4. Introduce a regional semantics and non-linear weighting strategy, construct a regional weighted enhancement module, and the internal pearl recognition image of Pinctada martensii passes through the regional weighted enhancement module to obtain an enhanced internal pearl image of Pinctada martensii; S5. Integrate the dynamic distance data partitioning module, the semantic weight calculation module, and the regional weighted enhancement module to construct an enhancement model for internal pearl recognition images of Pinctada martensii; S6. Input the internal pearl recognition image of Pinctada martensii into the enhancement model for internal pearl recognition images of Pinctada martensii for image enhancement, and output a high-resolution internal pearl recognition image of Pinctada martensii.
2. The internal pearl recognition image processing method of Pinctada fucata martensii according to claim 1, wherein In S2, constructing the dynamic distance data partitioning module includes the following steps: S21. Initialize a regional center, define a symmetric perturbation distance metric function , calculate the pixel and the regional center The similarity distance between them is given by the mathematical model: ； Among them, is the symmetric perturbation distance metric function, is the pixel value of the image at position ; is the -th region center position, initialized by random sampling, is the perturbation-sensitive gradient of the pixel value with respect to the region center, is for partial derivative calculation, is the angle between the pixel vector and the region center, is the pixel vector dimension, is the contribution ratio of the direction difference term; S22. Introduce a regional center reconstruction strategy, and update the center of each region based on the symmetric perturbation distance metric function. The mathematical model is: ； Among them, is the temperature parameter, is the th region, is the number of pixels in the th region, is the current number of regions, is the standard deviation of the th region, is the updated region center of the th region; S23. Based on the updated center of each region, assign each pixel of the image to the corresponding region to obtain the regional division result of the image. The mathematical model is: ； Among them, represents the area to which a pixel is assigned , indicating that the pixel is assigned to the area with the minimum similarity value ; S24. Iteratively execute steps S21 to S23, continuously optimize the regional center and pixel assignment until the change of all regional centers is less than the set threshold to achieve the final regional division of the image. The mathematical model is: ； wherein represents a preset convergence threshold value, is represented as the number of iterations.
3. The internal pearl recognition image processing method of Pinctada fucata martensii according to claim 2, characterized in that In S3, constructing the semantic weight calculation module includes the following steps: S31. Design a multi-dimensional Gaussian probability density modeling, and calculate the likelihood value of the pixels within each region. The mathematical model is: ； Among them, is the covariance matrix of the th region, represents the generation likelihood value of pixel under all regions; S32. Based on the likelihood value generated by the pixels of the region, calculate the semantic weight of each region. The mathematical model is: ； Among them, represents the number of pixels in the th region, is expressed as the proportion of the number of pixels in the region, represents the exponential calculation of the negative entropy, represents the semantic weight under the region.
4. A method for processing internal pearl recognition images of Pinctada fucata martensii according to claim 3, characterized in that In S4, constructing the regional weighted enhancement module includes the following steps: S41. Introduce the change of the sine function curve and the distance metric from the pixel to the regional center, and combine the regional semantic weight to calculate the distribution perception weight of the pixels within the region. The mathematical model is: ； Among them, is the distribution-aware weight of pixels , and is the pixel in the S42. Use distribution-aware weights Weight the pixels to obtain an enhanced internal pearl recognition image of Pinctada martensii. The mathematical model is as follows: ； Among them, is the enhanced internal pearl recognition image of Pinctada fucata martensii.
5. A method for identifying internal pearl image processing of Pinctada martensii according to claim 4, characterized in that, In S5, constructing the enhancement model for internal pearl recognition images of Pinctada martensii includes the following steps: S51. The original internal pearl recognition image of Pinctada martensii passes through the dynamic distance data partitioning module to obtain the partitioned data of the internal pearl recognition image of Pinctada martensii. The mathematical model is: ； Among them, is the internal pearl recognition image of the original Pinctada martensii, represents the partition data of the internal pearl recognition image of Pinctada martensii, represents the dynamic distance data partition module; S52. The original internal pearl recognition image of Pinctada martensii and the partitioned data of the internal pearl recognition image of Pinctada martensii pass through the semantic weight calculation module to obtain the regional semantic weight distribution of the internal pearl recognition image of Pinctada martensii. The mathematical model is: ； Among them, is the semantic weight distribution of the internal pearl recognition image area of Pinctada martensii, is the semantic weight calculation module; S53. The original internal pearl recognition image of Pinctada martensii and the regional semantic weight distribution of the internal pearl recognition image of Pinctada martensii pass through the regional weighted enhancement module to obtain the enhanced internal pearl recognition image of Pinctada martensii. The mathematical model is: ; Among them, is the enhanced internal pearl recognition image of Pinctada fucata martensii, is the region weighted enhancement module.