An efficient two-step multi-scale feature extraction method

Abstract

The application discloses a kind of efficient two-step multiscale feature extraction method, including the monitoring image data of the acquisition preset water area, the monitoring image data is preprocessed;The regional positioning and jigsaw patch of the monitoring image data obtain partition image, according to the partition image and moon jelly polyp, obtain detailed area and fuzzy area by feature similarity comparison;In detailed area, the macroscopic identification of the monitoring image is carried out to moon jelly polyp, and the coarse-grained feature is obtained in fuzzy area, and the microscopic identification of the monitoring image is carried out to moon jelly polyp, and the fine-grained feature is obtained;The learnable space reorganization of the fine-grained feature is obtained to improve fine-grained feature, according to the coarse-grained feature and the improved fine-grained feature, and key fusion feature is obtained by selection fusion;According to the fusion feature, construct moon jelly polyp multiscale feature extraction model, according to the data to be extracted output extraction result.

Description

Technical Field

[0001] This invention relates to the field of image feature extraction, and in particular to an efficient two-step multi-scale feature extraction method. Background Technology

[0002] With the continuous development of marine ecological environment monitoring technology, underwater organism identification and monitoring has become an important field of marine scientific research. As a typical marine planktonic organism, the identification and monitoring of the polyps of the moon jellyfish is of great significance for understanding jellyfish population dynamics and ecosystem changes. However, existing technologies still face many challenges in underwater image feature extraction. Traditional underwater image feature extraction methods typically employ single-scale feature analysis, which is insufficient to effectively address the complexity of the underwater environment. Underwater images are often affected by factors such as uneven illumination, scattering noise, and color distortion, making morphological feature extraction of moon jellyfish polyps difficult. Furthermore, the morphological differences exhibited by polyps at different developmental stages are significant, requiring multi-scale feature representation capabilities, from the microscopic distribution of nematocysts to the macroscopic umbrella-like outline.

[0003] In recent years, deep learning-based methods have made some progress in underwater image processing, but they still have limitations when dealing with organisms with complex morphological features, such as the polyps of the moon jellyfish. Existing convolutional neural network methods often struggle to simultaneously consider global contour features and local details, resulting in insufficient feature extraction accuracy in blurred regions. Furthermore, traditional methods lack specific designs for the characteristics of underwater organisms and cannot effectively utilize prior morphological knowledge of the polyps. On the other hand, existing multi-scale feature fusion methods often employ simple feature concatenation or weighted summation, failing to fully consider the semantic relationships between features at different scales, leading to low feature fusion efficiency. Especially when processing blurred regions in underwater images, conventional feature extraction methods often fail to reconstruct complete biological morphological features.

[0004] Therefore, there is an urgent need to develop a two-step multi-scale feature extraction method that can adapt to complex underwater environments, take into account both macroscopic and microscopic features, and has efficient feature fusion capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an efficient two-step multi-scale feature extraction method.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0007] This invention includes the following steps:

[0008] Collect monitoring image data of a preset water area and preprocess the monitoring image data;

[0009] The monitored image data is used to perform regional positioning and mosaic patching to obtain partitioned images. Based on the partitioned images and the polyps of the moon jellyfish, feature similarity comparison is performed to obtain detailed and blurred regions.

[0010] In the detailed area, macroscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain coarse-grained features, and in the blurred area, microscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain fine-grained features.

[0011] The fine-grained features are reorganized into a learnable space to obtain improved fine-grained features, and key fused features are obtained by selecting and fusing the coarse-grained features and the improved fine-grained features.

[0012] Based on the fusion features, a multi-scale feature extraction model for the polyps of the moon jellyfish is constructed. The data to be extracted is input into the multi-scale feature extraction model for the polyps of the moon jellyfish, and the extraction results are output.

[0013] Furthermore, the method for obtaining partitioned images by performing region localization and mosaic patching on the monitored image data includes:

[0014] The monitoring image is uniformly divided into multiple blocks, and a lightweight segmentation network based on convolutional neural network is constructed to output three types of heat maps: umbrella contour heat map, hydroid heat map, and noise mask.

[0015] Based on the heatmap response value, a dual threshold is set, dividing the blocks with heatmap response values ​​greater than 0.751 into high response zones, the blocks with heatmap response values ​​less than 0.498 into low response zones, and the blocks with heatmap response values ​​greater than 0.498 and less than 0.751 into transition zones.

[0016] The high-response zone uses dynamic grid segmentation, which adaptively adjusts the size of the grid blocks according to the hydroid density. The adjusted blocks include complete tentacle or nematocyst structures. The low-response zone uses fixed-size blocks.

[0017] Biological features within the grid blocks are obtained, and the biological features are screened to obtain the polyp features in the high-response and low-response areas. The polyp feature blocks in the high-response area are moved along the biological axis with a fixed step length to highlight the marginal membrane folds and tentacle arrangement patterns. K groups of mixed washing are performed according to the correlation of polyp features.

[0018] A shared transformer layer is used to encode the combination pattern of grid blocks, generating k local feature representations; the local feature representations are composed of both global and local features.

[0019] Based on local feature representation, the blocks are classified into two categories: regions containing moon jellyfish polyps and regions not containing moon jellyfish polyps.

[0020] We employ a combination of crossover loss and weighted regularized triplet loss to supervise the training of local feature representations in a convolutional neural network. Given the total loss function, the expression is:

[0021]

[0022]

[0023]

[0024] in For the total loss function, The crossover loss function, To weighted regularize the triplet loss, For global feature representation, For the i-th grid block in the j-th local feature branch, For the return of the ridge, For the i-th grid block, the true label is in category a. Let be the predicted probability that the i-th grid block belongs to category a. Let be the Euclidean distance between the i-th grid block and the same type of positive sample j. Let be the Euclidean distance between the i-th grid block and the out-of-class negative sample a. Let be the weight of the i-th grid block and the negative sample a. Let i be the weight of the i-th grid block and the positive sample j of the same class. The number of grid blocks, This represents the number of negative samples from the outlier class. The weight parameters for the regularization term, ;

[0025] Continue training until the total loss tends to converge, then output the partitioned images. Output the fine image set of the high response area as the target region image, and output the compressed image set of the low response area as the blurred region image.

[0026] Furthermore, the method for comparing the feature similarity between the partitioned image and the polyp of the moon jellyfish includes:

[0027] Based on the morphological characteristics of the polyps of the moon jellyfish, core contrast features were selected and extracted from the fine image set and the compressed image set, respectively.

[0028] In the high-response region, the core contrast features are numerically quantified and weighted according to feature stability. A weighted cosine similarity formula is then used to fuse the distances of multiple features, as expressed below:

[0029]

[0030] in The number of core contrast features, The weight of the z-th core contrast feature. Let be the Euclidean distance between the z-th hydroid feature and the core contrast feature in the region to be compared. Cosine similarity;

[0031] When the cosine similarity is greater than the first similarity threshold, it is confirmed that the region to be compared contains the polyp of the moon jellyfish, and the feature matching details are output; when the cosine similarity is lower than the second similarity threshold, it is confirmed that the region to be compared does not contain the polyp of the moon jellyfish; when the cosine similarity is greater than the second similarity threshold and less than the first similarity threshold, it is determined that the region to be compared is suspected to contain the polyp of the moon jellyfish.

[0032] For low-response regions and high-response regions suspected of containing jellyfish polyps, robust features are obtained by preserving highly robust morphological features. Binarization matching is used to obtain the number of matching terms between the robust features and the polyp features in the region to be matched. When the number of matching terms is greater than or equal to 2, the region to be matched contains the suspected target; when the number of matching terms is less than 2, the region to be matched does not contain the target.

[0033] The regions to be compared containing the polyp of the moon jellyfish are output as detailed regions, while the regions to be compared that are suspected to contain the polyp of the moon jellyfish and the regions to be matched that contain the suspected target are output as fuzzy regions.

[0034] Furthermore, a method for obtaining coarse-grained features by performing macroscopic identification of jellyfish polyps in the monitoring image within a detailed region includes:

[0035] Based on the life history characteristics of the moon jellyfish, core coarse-grained features are extracted from three dimensions: morphology, ecology, and texture. Contrast-adaptive histogram enhancement is applied to fine-grained regions. A segmentation network is used to generate an umbrella mask, and the perimeter-to-area ratio of the outline is calculated.

[0036]

[0037] in The area of ​​the umbrella. The circumference of the umbrella body. The ratio of the perimeter of the outline to the area;

[0038] If the ratio of the perimeter of the outline to the area is greater than the upper limit threshold, the umbrella is considered a complete umbrella; if the ratio of the perimeter of the outline to the area is less than the lower limit threshold, it indicates damage or juvenile status.

[0039] The detailed region is divided into grid blocks. An improved real-time target detection model is used to detect polyps. A confidence threshold is set, and non-maximum suppression is applied to overlapping detection boxes to retain individual polyps. The polyp density is then calculated.

[0040]

[0041] in For the density of hydroids in a detailed area, For the number of polyps in a detailed area, For detailed area measurements;

[0042] Nonlocal mean filtering is used to suppress underwater scattering noise. The reflection component is decomposed by a classic image enhancement algorithm in the field of computer vision to enhance the contrast of myofibril texture and output the core coarse-grained features in the detailed area as coarse-grained features.

[0043] Furthermore, a method for obtaining fine-grained features by performing microscopic identification of the polyps of the moon jellyfish in the monitoring image within a blurred region includes:

[0044] Based on the morphological and biological characteristics of the polyps of the moon jellyfish, a non-learning fine-grained feature extraction system was constructed, which specifically includes three levels of features: the first level features are the area of ​​the basal disc and the number of tentacles, the second level features are the density of nematocysts and the bifurcation angle of tentacles, and the third level features are the distance between muscle rings and the diameter of the oral disc.

[0045] Potential hydroid regions were located using dynamic threshold segmentation: the fluorescence channel used Otsu threshold segmentation to segment the nematocyst region, and the visible light channel used Cannibal edge detection to extract the basal disk contour to obtain the basal disk area; the region of interest was obtained by intersection operation, and an area filtering threshold was set.

[0046] The basement and tentacles are separated by distance transformation combined with watershed algorithm. The number of nematocysts per unit area is counted by neighboring connected components. The tentacles are then skeletonized. Linear fitting is performed on the skeletonized tentacles to calculate the angle between adjacent tentacle branches.

[0047] If the basal disk area is greater than or equal to the area threshold, the number of tentacles is greater than or equal to the number threshold, and the nematocyst density is greater than or equal to the density threshold, then the corresponding region of interest is determined to contain a moon jellyfish polyp, and the basal disk area, number of tentacles, and nematocyst density are output as fine-grained features.

[0048] Furthermore, a method for reorganizing the fine-grained features into a learnable space to obtain improved fine-grained features includes:

[0049] Fluorescence channels were extracted, and the contrast between nematocyst particles and the background was enhanced by Gaussian filtering. The blurred region of the attachment point of the basal disc of the hydroid was iteratively optimized based on the nonlocal mean filtering algorithm.

[0050] A weight matrix is ​​generated using a lightweight convolutional network, dynamically dividing the blurred region feature map into multiple sub-maps. Pixel-level feature recombination is then performed using Hadamard product, expressed as follows:

[0051]

[0052] in For the h-th subgraph based on the original feature map Learnable weighted matrix The output feature of the h-th sub-image after pixel-level feature recombination. For the original feature map in spatial location Eigenvalues ​​at;

[0053] The complexity of fuzzy regions is quantified using information entropy, where the information entropy of high-entropy regions exceeds the entropy threshold. The number of basis vectors is adjusted based on feature entropy; vectors are added to capture details of antennae forking in complex regions, while computation is reduced in simple regions. The expression is as follows:

[0054]

[0055] in The number of basis vectors. The maximum entropy value of the training set. Let be the base value of the basis vector. The feature entropy of the original feature map;

[0056] Using temperature coefficient The logic function focuses on the peak region of fluorescence intensity in the nematocyst, reconstructs the continuity of muscle texture in the transparent hydra by iteratively adjusting the distribution of basis vectors, and outputs the improved fine-grained features as improved fine-grained features.

[0057] Further, the method for selecting and fusing based on the coarse-grained features and the improved fine-grained features includes:

[0058] Multi-scale convolutional parallel attention is employed to capture the outline of the jellyfish bell and the folds of the marginal membrane through multi-scale parallel convolutional kernels. The enhanced parallel attention module separates global semantics from local positional information.

[0059] Based on the receptive field residual dense attention network, three residual dense blocks are stacked to extract the attachment point of the basal disc of the hydroid, the forking angle of the antennae, and the density of the nematocysts.

[0060] A coarse-grained feature map is constructed based on the coarse-grained features. Semantic segmentation is performed on the coarse-grained feature map to generate a polyp region mask. Only the features of the mask-covered region are retained. The channel contribution rate of the feature map is calculated based on the class feature variance. The channel contribution rates are sorted from largest to smallest, and the first one-third of the channels are retained.

[0061] Calculate the similarity matrix between coarse-grained features and improved fine-grained features. Multiply the association weights element-wise with the feature maps to semantically align the fine-grained features of the hydroid basal disc region with the coarse-grained features of the jellyfish attachment region. Perform biofeedback weighted fusion using dynamic weights and attention enhancement. The expression is as follows:

[0062]

[0063] in For balance coefficient, To improve fine-grained feature weights, For coarse-grained feature weights, As a feature of fusion, To improve fine-grained features, It is characterized by coarse-grainedness;

[0064] The dimensionality of the fusion features is reduced, and the fusion features with a cumulative contribution greater than the cumulative threshold are output as key fusion features.

[0065] Furthermore, the method for constructing a multi-scale feature extraction model of the moon jellyfish polyp based on the fused features includes:

[0066] A multi-scale feature extraction model for the polyp of the moon jellyfish is constructed based on hierarchical feature extraction, dynamic weighted fusion, and biological feature enhancement. The multi-scale feature extraction model for the polyp of the moon jellyfish adopts a dual-branch feature flow and performs collaborative learning of micro and macro features through cross-layer feature association and attention mechanism.

[0067] The multi-scale feature extraction model for the polyp of the moon jellyfish includes an input layer and preprocessing module, a multi-scale feature extraction module, a feature fusion module, and a classification and discrimination module.

[0068] Input layer and preprocessing module: The input image resolution is based on the monitoring image data, the coordinates of the target area are output based on the area positioning module, the region of interest is automatically cropped, and the underwater image is enhanced by low light enhancement, noise suppression and color correction.

[0069] Multi-scale feature extraction module: Constructs multiple feature extraction stages, outputs feature maps of 5 resolutions, and extracts biological features at different scales in a hierarchical manner;

[0070] Feature fusion module: Upsamples five feature maps of different resolutions to a uniform size using deconvolution; generates a biofeature heatmap through image segmentation, assigning high weights to the hydroid basal disk and jellyfish stoma, and applying channel attention weighting to multiple channels of the fused feature map; obtains cross-branch feature associations by performing feature interaction between macroscopic and microscopic features through mosaic patches, and obtains fused features by applying spatial attention weighting to the biofeature heatmap;

[0071] Classification and discrimination module: The dimensionality of the fused features is reduced through a fully connected layer, and a classifier is used to output three types of labels. The three types of labels include regions containing moon jellyfish polyps, regions containing moon jellyfish adults, and regions not containing the target. The total loss function is calculated based on cross-entropy classification loss, weighted regularized triplet loss, and ridge regression regularization term, and it is iterated until the total loss function tends to converge.

[0072] The beneficial effects of this invention are:

[0073] This invention is a highly efficient two-step multi-scale feature extraction method. Compared with existing technologies, this invention has the following technical advantages:

[0074] This invention employs a preprocessing, image partitioning, feature similarity comparison, macroscopic recognition, microscopic recognition, learnable space recombination, selective fusion, and model building steps. It achieves precise partitioning through region localization and mosaic patching techniques, combined with a dual-path feature extraction strategy. In detailed regions, macroscopic recognition is used to acquire coarse-grained features such as the umbrella-shaped outline, while in ambiguous regions, microscopic recognition captures fine-grained features such as the distribution of nematocysts. It innovatively introduces a learnable space recombination mechanism to effectively reconstruct the continuity of the muscle texture of the transparent polyp. A dynamic weighted fusion method is used to significantly improve the recognition accuracy and robustness of the moon jellyfish polyp in complex underwater environments. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating the steps of an efficient two-step multi-scale feature extraction method of the present invention. Detailed Implementation

[0076] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0077] The present invention provides an efficient two-step multi-scale feature extraction method comprising the following steps:

[0078] like Figure 1 As shown, this embodiment includes the following steps:

[0079] Collect monitoring image data of a preset water area and preprocess the monitoring image data;

[0080] In actual evaluation, preprocessing includes low-light enhancement, noise suppression (non-local mean filtering), and color correction. Data format differences are eliminated by unifying the image resolution (e.g., adjusting to 1024×1024 pixels), providing standardized input for subsequent region localization.

[0081] The monitored image data is used to perform regional positioning and mosaic patching to obtain partitioned images. Based on the partitioned images and the polyps of the moon jellyfish, feature similarity comparison is performed to obtain detailed and blurred regions.

[0082] In the detailed area, macroscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain coarse-grained features, and in the blurred area, microscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain fine-grained features.

[0083] The fine-grained features are reorganized into a learnable space to obtain improved fine-grained features, and key fused features are obtained by selecting and fusing the coarse-grained features and the improved fine-grained features.

[0084] Based on the fusion features, a multi-scale feature extraction model for the polyps of the moon jellyfish is constructed. The data to be extracted is input into the multi-scale feature extraction model for the polyps of the moon jellyfish, and the extraction results are output.

[0085] In this embodiment, the method for obtaining partitioned images by performing region localization and mosaic patching on the monitored image data includes:

[0086] The monitoring image is uniformly divided into multiple blocks, and a lightweight segmentation network based on convolutional neural network is constructed to output three types of heat maps: umbrella contour heat map, hydroid heat map, and noise mask.

[0087] Based on the heatmap response value, a dual threshold is set, dividing the blocks with heatmap response values ​​greater than 0.751 into high response zones, the blocks with heatmap response values ​​less than 0.498 into low response zones, and the blocks with heatmap response values ​​greater than 0.498 and less than 0.751 into transition zones.

[0088] The high-response zone uses dynamic grid segmentation, which adaptively adjusts the size of the grid blocks according to the hydroid density. The adjusted blocks include complete tentacle or nematocyst structures. The low-response zone uses fixed-size blocks.

[0089] Biological features within the grid blocks are obtained, and the biological features are screened to obtain the polyp features in the high-response and low-response areas. The polyp feature blocks in the high-response area are moved along the biological axis with a fixed step length to highlight the marginal membrane folds and tentacle arrangement patterns. K groups of mixed washing are performed according to the correlation of polyp features.

[0090] A shared transformer layer is used to encode the combination pattern of grid blocks, generating k local feature representations; the local feature representations are composed of both global and local features.

[0091] Based on local feature representation, the blocks are classified into two categories: regions containing moon jellyfish polyps and regions not containing moon jellyfish polyps.

[0092] We employ a combination of crossover loss and weighted regularized triplet loss to supervise the training of local feature representations in a convolutional neural network. Given the total loss function, the expression is:

[0093]

[0094]

[0095]

[0096] in For the total loss function, The crossover loss function, To weighted regularize the triplet loss, For global feature representation, For the i-th grid block in the j-th local feature branch, For the return of the ridge, Let be the true label of sample a in category for the i-th grid block. Let be the predicted probability that the i-th grid block belongs to category a. Let be the Euclidean distance between the i-th grid block and the same type of positive sample j. Let be the Euclidean distance between the i-th grid block and the out-of-class negative sample a. Let be the weight of the i-th grid block and the negative sample a. Let i be the weight of the i-th grid block and the positive sample j of the same class. The number of grid blocks, Let denoted as the number of outlier negative samples, and 'a' be the number of local feature branches. The weight parameters for the regularization term, ;

[0097] Continue training until the total loss tends to converge, output the partitioned images, output the fine image set of the high response area as the target region image, and output the compressed image set of the low response area as the blurred region image.

[0098] In actual evaluation, the fine image set retains the original resolution, while the compressed image set retains only global background information;

[0099] k-group washing: The value of k is dynamically adjusted according to the density of hydra in the high response zone. When the density is ≥5 hydra / cm², k=5; when the density is <5 hydra / cm², k=3, to ensure that each group of washing contains complete hydra feature blocks.

[0100] The baseline value of the basis vector is set to 8, which is determined through cross-validation of the training set; the number of local feature branches is set according to the resolution of the monitored image, and a=4 when the image resolution is 1024×1024 pixels.

[0101] In this embodiment, the method for comparing the feature similarity between the partitioned image and the polyp of the moon jellyfish includes:

[0102] Based on the morphological characteristics of the polyps of the moon jellyfish, core contrast features were selected and extracted from the fine image set and the compressed image set, respectively.

[0103] In the high-response region, the core contrast features are numerically quantified and weighted according to feature stability. A weighted cosine similarity formula is then used to fuse the distances of multiple features, as expressed below:

[0104]

[0105] in The number of core contrast features, The weight of the z-th core contrast feature. Let be the Euclidean distance between the z-th hydroid feature and the core contrast feature in the region to be compared. Cosine similarity;

[0106] When the cosine similarity is greater than the first similarity threshold, it is confirmed that the region to be compared contains the polyp of the moon jellyfish, and the feature matching details are output; when the cosine similarity is lower than the second similarity threshold, it is confirmed that the region to be compared does not contain the polyp of the moon jellyfish; when the cosine similarity is greater than the second similarity threshold and less than the first similarity threshold, it is determined that the region to be compared is suspected to contain the polyp of the moon jellyfish.

[0107] For low-response regions and high-response regions suspected of containing jellyfish polyps, robust features are obtained by preserving highly robust morphological features. Binarization matching is used to obtain the number of matching terms between the robust features and the polyp features in the region to be matched. When the number of matching terms is greater than or equal to 2, the region to be matched contains the suspected target; when the number of matching terms is less than 2, the region to be matched does not contain the target.

[0108] The region to be compared containing the polyp of the moon jellyfish is output as a detailed region, and the region to be compared that is suspected to contain the polyp of the moon jellyfish and the region to be matched that contains the suspected target are output as a fuzzy region.

[0109] In actual evaluation, the core contrast features of the high response area are nematocyst density, number of tentacle bifurcations, basal disc diameter, basal disc diameter ratio, muscle ring texture direction, and fluorescence intensity. The core contrast features of the low response area are nematocyst presence, tentacle outline integrity, basal disc outline presence, body height range (≥2 mm), texture continuity, and average gray value (≥128).

[0110] The weights for nematocyst density are 0.35, basal disk diameter is 0.25, fluorescence intensity is 0.20, and the total weight for other features is 0.20. The fluorescence intensity weight is dynamically adjusted with underwater light intensity. The first similarity threshold is 0.759, and the second similarity threshold is 0.504.

[0111] Robust features include the presence of nematocysts, the presence of basal disc contours, and improved range. In the suspected target-containing matching area, the fluorescence signal detection threshold is reduced, the contour detection algorithm is simplified, the matching speed is improved, and the number of matching items is optimized.

[0112] The coordinates of target region 1 are (128, 156, 342, 410). The matching features are nematocyst density of 0.75 / 100μm² (match), basal disk diameter of 0.9mm (match), and fluorescence intensity of 520nm (match). The similarity score is 0.82, confirming the presence of the moon jellyfish polyp.

[0113] The coordinates of target region 2 are (512,204,630,328). The matching features are nematocyst density of 0.4 per 100μm² (low) and number of tentacle branches of 3 (standard 4). The similarity score is 0.65, which suggests that it may contain the polyps of the moon jellyfish.

[0114] The coordinates of the blurred region 1 are (800, 450, 880, 520). The matching features are nematocysts (1) and basal disc outlines (1). The number of matching items is 2, and it is suspected to contain the polyp of the moon jellyfish.

[0115] The coordinates of the blurred region 2 are (210, 600, 280, 660). The matching features are: nematocysts do not exist (0), the basal disk outline is incomplete (0), the number of matching items is 0, and it does not contain the polyp of the moon jellyfish.

[0116] In this embodiment, the method for obtaining coarse-grained features by performing macroscopic identification of the polyps of the moon jellyfish in the monitoring image within a detailed region includes:

[0117] Based on the life history characteristics of the moon jellyfish, core coarse-grained features are extracted from three dimensions: morphology, ecology, and texture. Contrast-adaptive histogram enhancement is applied to fine-grained regions. A segmentation network is used to generate an umbrella mask, and the perimeter-to-area ratio of the outline is calculated.

[0118]

[0119] in The area of ​​the umbrella. The circumference of the umbrella body. The ratio of the perimeter of the outline to the area;

[0120] If the ratio of the perimeter of the outline to the area is greater than the upper limit threshold, the umbrella is considered a complete umbrella; if the ratio of the perimeter of the outline to the area is less than the lower limit threshold, it indicates damage or juvenile status.

[0121] The detailed region is divided into grid blocks. An improved real-time target detection model is used to detect polyps. A confidence threshold is set, and non-maximum suppression is applied to overlapping detection boxes to retain individual polyps. The polyp density is then calculated.

[0122]

[0123] in For the density of hydroids in a detailed area, For the number of polyps in a detailed area, For detailed area measurements;

[0124] Nonlocal mean filtering is used to suppress underwater scattering noise. The reflection component is decomposed by a classic image enhancement algorithm in the field of computer vision to enhance the contrast of myofibril texture and output the core coarse-grained features in the detailed area as coarse-grained features.

[0125] In actual assessment, core coarse-grained features include the complete shape of the umbrella body outline, the attachment status of the hydroid basal disc, the number and arrangement of tentacles, cluster density, spatial relationship with the adult, the average gray value of the gonadal region, the orientation of myofibrils on the umbrella body surface, and the number of steatoid sacs.

[0126] The upper limit threshold is 0.851, the lower limit threshold is 0.697, and the confidence threshold is 0.783.

[0127] Detailed region 1: The macroscopic feature is the complete outline of the umbrella, with an extraction result of 0.92 (complete), indicating that the adult is in good health and has a confidence level of 0.95;

[0128] Detailed region 2: The macroscopic feature is the density of hydroid clusters, with an extraction result of 8 individuals / cm², indicating an active reproductive period, with a confidence level of 0.88;

[0129] Detailed region 3: The macroscopic feature is the gray mean of the gonad region, with an extraction result of 115, for adult individuals, and a confidence level of 0.91.

[0130] In this embodiment, the method for obtaining fine-grained features by performing microscopic identification of the polyps of the moon jellyfish in the monitoring image within a blurred region includes:

[0131] Based on the morphological and biological characteristics of the polyps of the moon jellyfish, a non-learning fine-grained feature extraction system was constructed, which specifically includes three levels of features: the first level features are the area of ​​the basal disc and the number of tentacles, the second level features are the density of nematocysts and the bifurcation angle of tentacles, and the third level features are the distance between muscle rings and the diameter of the oral disc.

[0132] Potential hydroid regions were located using dynamic threshold segmentation: the fluorescence channel used Otsu threshold segmentation to segment the nematocyst region, and the visible light channel used Cannibal edge detection to extract the basal disk contour to obtain the basal disk area; the region of interest was obtained by intersection operation, and an area filtering threshold was set.

[0133] The basement and tentacles are separated by distance transformation combined with watershed algorithm. The number of nematocysts per unit area is counted by neighboring connected components. The tentacles are then skeletonized. Linear fitting is performed on the skeletonized tentacles to calculate the angle between adjacent tentacle branches.

[0134] If the basal disk area is greater than or equal to the area threshold, the number of tentacles is greater than or equal to the number threshold, and the nematocyst density is greater than or equal to the density threshold, then the corresponding region of interest is determined to contain a moon jellyfish polyp, and the basal disk area, number of tentacles, and nematocyst density are output as fine-grained features.

[0135] In the actual assessment, the area threshold was set at 300 μm² (based on the statistical value of the basal disc area of ​​the larval polyps of the moon jellyfish), the number threshold was set at 6 (excluding deformed or damaged individuals), and the density threshold was set at 10 individuals / 100 μm² (the minimum effective density determined through comparative experiments).

[0136] Sample 1: Basement area 527 μm², number of tentacles 11, nematocyst density 15 / 100 μm², average bifurcation angle 42°, developmental stage determined to be glandular, confidence level 0.92; Sample 2: Basement area 389 μm², number of tentacles 7, nematocyst density 13 / 100 μm², developmental stage determined to be suspected larva, confidence level 0.68.

[0137] In this embodiment, the method for obtaining improved fine-grained features by reorganizing the fine-grained features into a learnable space includes:

[0138] Fluorescence channels were extracted, and the contrast between nematocyst particles and the background was enhanced by Gaussian filtering. The blurred region of the attachment point of the basal disc of the hydroid was iteratively optimized based on the nonlocal mean filtering algorithm.

[0139] A weight matrix is ​​generated using a lightweight convolutional network, dynamically dividing the blurred region feature map into multiple sub-maps. Pixel-level feature recombination is then performed using Hadamard product, expressed as follows:

[0140]

[0141] in For the h-th subgraph based on the original feature map Learnable weighted matrix The output feature of the h-th sub-image after pixel-level feature recombination. For the original feature map in spatial location Eigenvalues ​​at;

[0142] The complexity of fuzzy regions is quantified using information entropy, where the information entropy of high-entropy regions exceeds the entropy threshold. The number of basis vectors is adjusted based on feature entropy; vectors are added to capture details of antennae forking in complex regions, while computation is reduced in simple regions. The expression is as follows:

[0143]

[0144] in The number of basis vectors. The maximum entropy value of the training set. Let be the base value of the basis vector. The feature entropy of the original feature map;

[0145] Using temperature coefficient The logic function focuses on the peak region of fluorescence intensity in the nematocyst, reconstructs the continuity of muscle texture in the transparent hydra by iteratively adjusting the distribution of basis vectors, and outputs the improved fine-grained features as improved fine-grained features.

[0146] In this embodiment, the method for selecting and fusing based on the coarse-grained features and the improved fine-grained features includes:

[0147] Multi-scale convolutional parallel attention is employed to capture the outline of the jellyfish bell and the folds of the marginal membrane through multi-scale parallel convolutional kernels. The enhanced parallel attention module separates global semantics from local positional information.

[0148] Based on the receptive field residual dense attention network, three residual dense blocks are stacked to extract the attachment point of the basal disc of the hydroid, the forking angle of the antennae, and the density of the nematocysts.

[0149] A coarse-grained feature map is constructed based on the coarse-grained features. Semantic segmentation is performed on the coarse-grained feature map to generate a polyp region mask. Only the features of the mask-covered region are retained. The channel contribution rate of the feature map is calculated based on the class feature variance. The channel contribution rates are sorted from largest to smallest, and the first one-third of the channels are retained.

[0150] Calculate the similarity matrix between coarse-grained features and improved fine-grained features. Multiply the association weights element-wise with the feature maps to semantically align the fine-grained features of the hydroid basal disc region with the coarse-grained features of the jellyfish attachment region. Perform biofeedback weighted fusion using dynamic weights and attention enhancement. The expression is as follows:

[0151]

[0152] in For balance coefficient, To improve fine-grained feature weights, For coarse-grained feature weights, As a feature of fusion, To improve fine-grained features, It is characterized by coarse-grainedness;

[0153] Dimensionality reduction is performed on the fusion features, and the fusion features with a cumulative contribution greater than the cumulative threshold are output as key fusion features;

[0154] In actual assessments, key fusion features include: umbrella aspect ratio, hydroid basal disc diameter, nematocyst density, muscle ring texture direction, relative position of jellyfish and hydroid, and cluster distribution density.

[0155] In this embodiment, the method for constructing a multi-scale feature extraction model of the moon jellyfish polyp based on the fused features includes:

[0156] A multi-scale feature extraction model for the polyp of the moon jellyfish is constructed based on hierarchical feature extraction, dynamic weighted fusion, and biological feature enhancement. The multi-scale feature extraction model for the polyp of the moon jellyfish adopts a dual-branch feature flow and performs collaborative learning of micro and macro features through cross-layer feature association and attention mechanism.

[0157] The multi-scale feature extraction model for the polyp of the moon jellyfish includes an input layer and preprocessing module, a multi-scale feature extraction module, a feature fusion module, and a classification and discrimination module.

[0158] Input layer and preprocessing module: The input image resolution is based on the monitoring image data, the coordinates of the target area are output based on the area positioning module, the region of interest is automatically cropped, and the underwater image is enhanced by low light enhancement, noise suppression and color correction.

[0159] Multi-scale feature extraction module: Constructs multiple feature extraction stages, outputs feature maps of 5 resolutions, and extracts biological features at different scales in a hierarchical manner;

[0160] Feature fusion module: Upsamples five feature maps of different resolutions to a uniform size using deconvolution; generates a biofeature heatmap through image segmentation, assigning high weights to the hydroid basal disk and jellyfish stoma, and applying channel attention weighting to multiple channels of the fused feature map; obtains cross-branch feature associations by performing feature interaction between macroscopic and microscopic features through mosaic patches, and obtains fused features by applying spatial attention weighting to the biofeature heatmap;

[0161] The classification and discrimination module reduces the dimensionality of the fused features through a fully connected layer and outputs three types of labels using a classifier. These three types of labels include regions containing the polyp of the moon jellyfish, regions containing the adult moon jellyfish, and regions not containing the target. The total loss function is calculated based on cross-entropy classification loss, weighted regularized triplet loss, and ridge regression regularization term, and iterated until the total loss function tends to converge.

[0162] In the actual evaluation, the layers included shallow, intermediate, and deep layers; shallow layer: receptive field size was 8×8~16×16, the extraction targets were nematocyst distribution and tentacle bifurcation angle, the core parameter was 3×3 convolution (stride 1, padding 1), and the number of channels was 64→128; intermediate layer: receptive field size was 32×32~64×64, the extraction targets were morphology of the hydrocephalus basal disc and umbrella folds, the core parameter was 3×3 grouped convolution (16 groups), and the number of channels was 256→512; deep layer: receptive field size was 128×128~256×256, the extraction targets were jellyfish umbrella diameter and gonad location (macroscopic structure), and the core parameter was 1×1 convolution dimensionality reduction (number of channels 1024→512).

[0163] Upsampling: Shallow features: retain details such as tentacle bifurcation and nematocyst edges, upsampling is performed through bilinear interpolation; Deep features: retain global structures such as umbrella outline and gonad distribution, upsampling is performed through transposed convolution (kernel=4, stride=2);

[0164] The mosaic patch in the region localization stage is used for image partitioning, dividing only high / low response regions; the mosaic patch in the feature fusion stage is used for feature interaction, aligning macroscopic features (such as umbrella outline) and microscopic features (such as nematocyst distribution) along biological axes to avoid functional overlap and improve model running efficiency.

[0165] Channel attention weighting: High-weighted channels correspond to nematocyst fluorescence features (520nm channel) and muscle texture features (regions with drastic gradient changes); Low-weighted channels correspond to background noise (suspended particles, light spots), which are dynamically suppressed by the Sigmoid function.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An efficient two-step multi-scale feature extraction method, characterized in that, Includes the following steps: Collect monitoring image data of a preset water area and preprocess the monitoring image data; The monitored image data is used to perform regional positioning and mosaic patching to obtain partitioned images. Based on the partitioned images and the polyps of the moon jellyfish, feature similarity comparison is performed to obtain detailed and blurred regions. In the detailed area, macroscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain coarse-grained features, and in the blurred area, microscopic identification of the polyps of the moon jellyfish is performed on the monitoring image to obtain fine-grained features. The fine-grained features are reorganized into a learnable space to obtain improved fine-grained features; key fused features are obtained by selecting and fusing the coarse-grained features and the improved fine-grained features; including: Fluorescence channels were extracted, and the contrast between nematocyst particles and the background was enhanced by Gaussian filtering. The blurred region of the attachment point of the basal disc of the hydroid was iteratively optimized based on the nonlocal mean filtering algorithm. A weight matrix is ​​generated using a lightweight convolutional network, dynamically dividing the blurred region feature map into multiple sub-maps. Pixel-level feature recombination is then performed using Hadamard product, expressed as follows: in For the h-th subgraph based on the original feature map Learnable weighted matrix The output feature of the h-th sub-image after pixel-level feature recombination. For the original feature map in spatial location Eigenvalues ​​at; The complexity of fuzzy regions is quantified using information entropy, where the information entropy of high-entropy regions exceeds the entropy threshold. The number of basis vectors is adjusted based on feature entropy; vectors are added to capture details of antennae forking in complex regions, while computation is reduced in simple regions. The expression is as follows: in The number of basis vectors. The maximum entropy value of the training set. Let be the base value of the basis vector. The feature entropy of the original feature map; Using temperature coefficient The logic function focuses on the peak region of fluorescence intensity in the nematocyst, and reconstructs the continuity of muscle texture of transparent hydra by iteratively adjusting the distribution of basis vectors, and outputs the improved fine-grained features as improved fine-grained features. Based on the fusion features, a multi-scale feature extraction model for the polyps of the moon jellyfish is constructed. The data to be extracted is input into the multi-scale feature extraction model for the polyps of the moon jellyfish, and the extraction results are output.

2. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, A method for obtaining partitioned images by performing region localization and mosaic patching on the monitored image data includes: The monitoring image is uniformly divided into multiple blocks, and a lightweight segmentation network based on convolutional neural network is constructed to output three types of heat maps: umbrella contour heat map, hydroid heat map, and noise mask. Based on the heatmap response value, a dual threshold is set, dividing the blocks with heatmap response values ​​greater than 0.751 into high response zones, the blocks with heatmap response values ​​less than 0.498 into low response zones, and the blocks with heatmap response values ​​greater than 0.498 and less than 0.751 into transition zones. The high-response zone uses dynamic grid segmentation, which adaptively adjusts the size of the grid blocks according to the hydroid density. The adjusted blocks include complete tentacle or nematocyst structures. The low-response zone uses fixed-size blocks. Biological features within the grid blocks are obtained, and the biological features are screened to obtain the polyp features in the high-response and low-response areas. The polyp feature blocks in the high-response area are moved along the biological axis with a fixed step length to highlight the marginal membrane folds and tentacle arrangement patterns. K groups of mixed washing are performed according to the correlation of polyp features. A shared transformer layer is used to encode the combination pattern of grid blocks, generating k local feature representations; the local feature representations are composed of both global and local features. Based on local feature representation, the blocks are classified into two categories: regions containing moon jellyfish polyps and regions not containing moon jellyfish polyps. We employ a combination of crossover loss and weighted regularized triplet loss to supervise the training of local feature representations in a convolutional neural network. Given the total loss function, the expression is: in For the total loss function, The crossover loss function, To weighted regularize the triplet loss, For global feature representation, For the i-th grid block in the j-th local feature branch, For the return of the ridge, For the i-th grid block, the true label is in category a. Let be the predicted probability that the i-th grid block belongs to category a. Let be the Euclidean distance between the i-th grid block and the same type of positive sample j. Let be the Euclidean distance between the i-th grid block and the out-of-class negative sample a. Let be the weight of the i-th grid block and the negative sample a. Let i be the weight of the i-th grid block and the positive sample j of the same class. The number of grid blocks, This represents the number of negative samples from the outlier class. The weight parameters for the regularization term, ; Continue training until the total loss tends to converge, then output the partitioned images. Output the fine image set of the high response area as the target region image, and output the compressed image set of the low response area as the blurred region image.

3. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, The method for comparing feature similarity between the partitioned image and the polyp body of the moon jellyfish includes: Based on the morphological characteristics of the polyps of the moon jellyfish, core contrast features were selected and extracted from the fine image set and the compressed image set, respectively. In the high-response region, the core contrast features are numerically quantified and weighted according to feature stability. A weighted cosine similarity formula is then used to fuse the distances of multiple features, as expressed below: in The number of core contrast features, The weight of the z-th core contrast feature. Let be the Euclidean distance between the z-th hydroid feature and the core contrast feature in the region to be compared. Cosine similarity; When the cosine similarity is greater than the first similarity threshold, it is confirmed that the region to be compared contains the polyp of the moon jellyfish, and the feature matching details are output; when the cosine similarity is lower than the second similarity threshold, it is confirmed that the region to be compared does not contain the polyp of the moon jellyfish; when the cosine similarity is greater than the second similarity threshold and less than the first similarity threshold, it is determined that the region to be compared is suspected to contain the polyp of the moon jellyfish. For low-response regions and high-response regions suspected of containing jellyfish polyps, robust features are obtained by preserving highly robust morphological features. Binarization matching is used to obtain the number of matching terms between the robust features and the polyp features in the region to be matched. When the number of matching terms is greater than or equal to 2, the region to be matched contains the suspected target; when the number of matching terms is less than 2, the region to be matched does not contain the target. The regions to be compared containing the polyp of the moon jellyfish are output as detailed regions, while the regions to be compared that are suspected to contain the polyp of the moon jellyfish and the regions to be matched that contain the suspected target are output as fuzzy regions.

4. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, A method for obtaining coarse-grained features by performing macroscopic identification of the polyps of the moon jellyfish in the monitored image within a detailed region includes: Based on the life history characteristics of the moon jellyfish, core coarse-grained features are extracted from three dimensions: morphology, ecology, and texture. Contrast-adaptive histogram enhancement is applied to fine-grained regions. A segmentation network is used to generate an umbrella mask, and the perimeter-to-area ratio of the outline is calculated. in The area of ​​the umbrella. The circumference of the umbrella body. The ratio of the perimeter of the outline to the area; If the ratio of the perimeter of the outline to the area is greater than the upper limit threshold, the umbrella is considered a complete umbrella; if the ratio of the perimeter of the outline to the area is less than the lower limit threshold, it indicates damage or juvenile status. The detailed region is divided into grid blocks. An improved real-time target detection model is used to detect polyps. A confidence threshold is set, and non-maximum suppression is applied to overlapping detection boxes to retain individual polyps. The polyp density is then calculated. in For the density of hydroids in a detailed area, For the number of polyps in a detailed area, For detailed area measurements; Nonlocal mean filtering is used to suppress underwater scattering noise. The reflection component is decomposed by a classic image enhancement algorithm in the field of computer vision to enhance the contrast of myofibril texture and output the core coarse-grained features in the detailed area as coarse-grained features.

5. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, A method for obtaining fine-grained features by performing microscopic identification of the polyps of the moon jellyfish in the monitored image within a blurred region includes: Based on the morphological and biological characteristics of the polyps of the moon jellyfish, a non-learning fine-grained feature extraction system was constructed, which specifically includes three levels of features: the first level features are the area of ​​the basal disc and the number of tentacles, the second level features are the density of nematocysts and the bifurcation angle of tentacles, and the third level features are the distance between muscle rings and the diameter of the oral disc. Potential hydroid regions were located using dynamic threshold segmentation: the fluorescence channel used Otsu threshold segmentation to segment the nematocyst region, and the visible light channel used Cannibal edge detection to extract the basal disk contour to obtain the basal disk area; the region of interest was obtained by intersection operation, and an area filtering threshold was set. The basement and tentacles are separated by distance transformation combined with watershed algorithm. The number of nematocysts per unit area is counted by neighboring connected components. The tentacles are then skeletonized. Linear fitting is performed on the skeletonized tentacles to calculate the angle between adjacent tentacle branches. If the basal disk area is greater than or equal to the area threshold, the number of tentacles is greater than or equal to the number threshold, and the nematocyst density is greater than or equal to the density threshold, then the corresponding region of interest is determined to contain a moon jellyfish polyp, and the basal disk area, number of tentacles, and nematocyst density are output as fine-grained features.

6. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, The method for selecting and fusing features based on the coarse-grained features and the improved fine-grained features includes: Multi-scale convolutional parallel attention is employed to capture the outline of the jellyfish bell and the folds of the marginal membrane through multi-scale parallel convolutional kernels. The enhanced parallel attention module separates global semantics from local positional information. Based on the receptive field residual dense attention network, three residual dense blocks are stacked to extract the attachment point of the basal disc of the hydroid, the forking angle of the antennae, and the density of the nematocysts. A coarse-grained feature map is constructed based on the coarse-grained features. Semantic segmentation is performed on the coarse-grained feature map to generate a polyp region mask. Only the features of the mask-covered region are retained. The channel contribution rate of the feature map is calculated based on the class feature variance. The channel contribution rates are sorted from largest to smallest, and the first one-third of the channels are retained. Calculate the similarity matrix between coarse-grained features and improved fine-grained features. Multiply the association weights element-wise with the feature maps to semantically align the fine-grained features of the hydroid basal disc region with the coarse-grained features of the jellyfish attachment region. Perform biofeedback weighted fusion using dynamic weights and attention enhancement. The expression is as follows: in For balance coefficient, To improve fine-grained feature weights, For coarse-grained feature weights, As a feature of fusion, To improve fine-grained features, It is characterized by coarse-grainedness; The dimensionality of the fusion features is reduced, and the fusion features with a cumulative contribution greater than the cumulative threshold are output as key fusion features.

7. The efficient two-step multi-scale feature extraction method according to claim 1, characterized in that, A method for constructing a multi-scale feature extraction model of the polyp body of the moon jellyfish based on the fused features includes: A multi-scale feature extraction model for the polyp of the moon jellyfish is constructed based on hierarchical feature extraction, dynamic weighted fusion, and biological feature enhancement. The multi-scale feature extraction model for the polyp of the moon jellyfish adopts a dual-branch feature flow and performs collaborative learning of micro and macro features through cross-layer feature association and attention mechanism. The multi-scale feature extraction model for the polyp of the moon jellyfish includes an input layer and preprocessing module, a multi-scale feature extraction module, a feature fusion module, and a classification and discrimination module. Input layer and preprocessing module: The input image resolution is based on the monitoring image data, the coordinates of the target area are output based on the area positioning module, the region of interest is automatically cropped, and the underwater image is enhanced by low light enhancement, noise suppression and color correction. Multi-scale feature extraction module: Constructs multiple feature extraction stages, outputs feature maps of 5 resolutions, and extracts biological features at different scales in a hierarchical manner; Feature fusion module: Upsamples five feature maps of different resolutions to a uniform size using deconvolution; generates a biofeature heatmap through image segmentation, assigning high weights to the hydroid basal disk and jellyfish stoma, and applying channel attention weighting to multiple channels of the fused feature map; obtains cross-branch feature associations by performing feature interaction between macroscopic and microscopic features through mosaic patches, and obtains fused features by applying spatial attention weighting to the biofeature heatmap; Classification and discrimination module: The dimensionality of the fused features is reduced through a fully connected layer, and a classifier is used to output three types of labels. The three types of labels include regions containing moon jellyfish polyps, regions containing moon jellyfish adults, and regions not containing the target. The total loss function is calculated based on cross-entropy classification loss, weighted regularized triplet loss, and ridge regression regularization term, and it is iterated until the total loss function tends to converge.

Images