Graphical machine learning algorithm platform

Abstract

The present application discloses a graphical machine learning algorithm platform, including: an information input module, used to input image information to be processed so as to pre-process the image information; an information segmentation module, which over-segments the image information to obtain a number of small blocks; an information description module, which extracts the geometric features of all facets for each small block, and selects a geometric feature therefrom as the standard feature of the small block; in the technical solution provided by the present application, compared with the feature of the existing solution that each facet needs to be labeled, in this solution, for each small block, a geometric feature of a facet is used as the standard feature of the small block, so in comparison, the data that needs to be labeled will be drastically reduced, thereby reducing the construction cost of the learning algorithm.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a graphical machine learning algorithm platform. Background Art

[0002] The image is a standard two-dimensional plane information, but in image processing technology, especially image feature recognition, it is necessary to read three-dimensional information from the two-dimensional image information in order to build a three-dimensional visualization information management platform.

[0003] The key to identifying 3D information from 2D information is to segment the 2D image into 3D shapes. Currently, the commonly used 3D shape segmentation methods all use deep learning technology. Deep learning technology trains multi-layer neural networks to learn how to identify high-order features in 2D information to complete the automatic segmentation of 3D shapes.

[0004] When segmenting three-dimensional shapes in images, a trained graphical machine learning algorithm platform is generally used. The algorithm platform has a built-in trained image segmentation model that can automatically segment the input image and then mark the segmented area. However, this type of algorithm platform requires a large amount of labeled data to train the model, and in order to ensure recognition accuracy, the algorithm platform needs to identify a large number of features to find the relationship between the features and the labeled data, so the training of this learning algorithm platform is difficult. Summary of the invention

[0005] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0006] As a first aspect of the present application, in order to solve the technical problems mentioned in the above background technology section, some embodiments of the present application provide a graphical machine learning algorithm platform, including: an information input module, used to input image information to be processed, so as to pre-process the image information;

[0007] The information segmentation module over-segments the image information to obtain several small blocks;

[0008] The information description module extracts the geometric features of all the patches for each small block, and selects a geometric feature from them as the standard feature of the small block;

[0009] The prediction module obtains the standard features of each small block, automatically segments the image and annotates the entities based on the differences between the standard features.

[0010] In the technical solution provided in the present application, compared with the feature of the existing solution that each face needs to be labeled, in this solution, the geometric features of a face are used as the standard features of each small block. Therefore, in comparison, during the iteration process, the information of the graphical machine learning algorithm platform in each cycle will be drastically reduced, thereby reducing the training cost of the learning algorithm.

[0011] Furthermore, the information input module performs smoothing filtering and denoising on the image information to reduce the noise in the image information.

[0012] Using the geometric features of one of the patches as the standard features for each small block will, to a certain extent, result in the final standard features being unable to represent the geometric features of the entire patch. To this end, the present application provides the following technical solutions:

[0013] Furthermore, the information segmentation module divides the image into M small blocks according to the edges based on the edge detection algorithm.

[0014] In this solution, the small blocks are divided based on the edge detection algorithm, so each small block is basically divided by the internal edge contour in the image information, so the information inside each small block is relatively uniform, reducing the situation where the patch cannot represent the entire small block.

[0015] The geometric features of each facet are diverse, and it is difficult to accurately find the intrinsic relationship between the geometric features of each facet. Therefore, when selecting standard features, if it is not possible to accurately find the most representative features of all the geometric features, the final selected standard features will not be able to represent the small blocks. To this end, the present application provides the following technical solutions:

[0016] The information description module includes: an information extraction unit, for the i-th small block, extracts its n geometric features to obtain a feature vector F i ;

[0017] F i ={f i1 、f i2 ,…f ij …f in}, where f ij Represents the jth geometric feature of the i-th small block;

[0018] Dimensionality reduction unit, for the feature vector F i The geometric features of each small block in the dimensionality reduction to the two-dimensional feature space, get n two-dimensional space point set A, A = {P i1 , P i2 ,…P ij …P in}, P ijRepresents the data point of the jth geometric feature of the i-th small block in the two-dimensional feature space;

[0019] Among them, P ij The coordinates of (x ij ,y ij ), x ij =f ij *u1,y ij =f ij *u2, * represents the dot product operation, u1 represents the eigenvector F i The first principal component of u2 represents the eigenvector F i The second principal component of

[0020] The information extraction unit calculates the local density ρ for each data point in the two-dimensional space point set A ij , Where exp represents the exponential function, d c represents the cutoff distance parameter, Indicates P ij With P ik The Euclidean distance, j and k both represent the index of the data point;

[0021] For each data point P ij Calculate the minimum distance δ to a point with higher density than it j ;

[0022] By δ j is the vertical axis, ρ ij As the horizontal axis, draw the distribution diagram of all data points, select the cluster center based on the comprehensive values ​​of the horizontal and vertical axes, and use the geometric features represented by the data points corresponding to the cluster center as the standard features.

[0023] In the technical solution provided by the present application, when segmenting the three-dimensional model, the image information is segmented into very small blocks, and then each block is divided into dozens of facets. This fine-grained division method can accurately capture the local three-dimensional features of the image information and increase the segmentation accuracy of the image information. And when selecting the description method for each small block, using average level facets to represent the entire small block can reduce the impact of noise and outliers and increase the representativeness of the features. When describing a specific small block, the three-dimensional model is mapped to a two-dimensional feature space, which reduces the data dimension and can intuitively describe the similarity relationship between features, facilitating subsequent segmentation and labeling.

[0024] Furthermore, the prediction module has a built-in neural network model, which learns the correspondence between standard features and labels and automatically labels each small block.

[0025] Furthermore, the prediction module includes:

[0026] The initialization unit initializes each small block and initializes the label of the small block to zero. The initialization here means setting the so-called position in the image information to zero.

[0027] The iteration unit randomly selects a small block i and selects a segmentation label according to the state information of the small block i. The state information includes the standard features of the small block, the standard features of the adjacent small blocks, and the labels of the adjacent small blocks.

[0028] Reward calculation unit, calculates the reward value r of the segmentation label assigned to small block i t ;

[0029] The loop unit guides the iterative unit to loop continuously until the maximum cumulative reward value R is obtained.

[0030] Furthermore, in the iteration unit:

[0031] Status information includes S wi , S ei , S li ;

[0032] For a small block i, its standard characteristic is S wi , the average standard feature of the remaining small blocks adjacent to small block i is S ei ; The label distribution of the remaining small blocks adjacent to small block i is S li ;

[0033] S ei ={E 1i 、E 2i 、E 3i …E ei …}, where E ei is the standard feature of the e-th block among the remaining blocks adjacent to block i;

[0034] S li ={ L 1i 、L 2i 、L 3i …L li ...}, L li is a distribution vector, where l The element represents the number of small blocks adjacent to small block i. l A small piece of label.

[0035] In the technical solution provided in the present application, when updating the label of a small block, the update is performed based on the standard features of the small block and the labels of adjacent small blocks. Therefore, relatively speaking, less annotation data is required, which increases the accuracy of the small block annotation and enables better finding of the connection between information during the iteration process.

[0036] When training the model, some labeled data is needed. To a certain extent, these labeled data are not all correct data. There are also a lot of wrong data. These wrong data are actually difficult to distinguish. When used in the training of the prediction module, the prediction module will learn some wrong information. To this end, the present application provides the following technical solutions:

[0037] Furthermore, after all small blocks are assigned labels, the recurrent unit selects small blocks with different labels from surrounding small blocks from all small blocks, and then inputs them into the iteration unit for re-iteration.

[0038] In the technical solution provided in the present application, based on the distribution of labels between small blocks, small blocks that may have labeling errors are selected, and then these small blocks are guided to perform further reinforcement learning, thereby reducing the impact of erroneous data on prediction accuracy.

[0039] Furthermore, the model is affected by the reward value during the iteration process, and the design of the reward value will affect the convergence rate of the model and increase the training cost of the model. For this purpose, the present application provides the following technical solutions:

[0040] Furthermore, the reward value is r t , , is a Gaussian function, min(d i ) represents the segmentation boundary G from small block i to the true value i The shortest geodesic distance among all points on the network. The geodesic distance refers to the length of the shortest path from one point to another. i This refers to the correct segmentation result of small block i, from small block i to the true segmentation boundary G i The minimum geodesic distance is the label of patch i, α and β are a pair of proportional parameters, t is the number of iterations, a t Represents the action at the tth iteration.

[0041] To sum up: in the technical solution provided by the present application, compared with the characteristic of the existing solution that each face needs to be labeled, in this solution, the geometric features of a face are used as the standard features of each small block. Therefore, in comparison, during the iteration process, the information of the graphical machine learning algorithm platform in each cycle will be drastically reduced, thereby reducing the training cost of the learning algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The drawings and descriptions of the exemplary embodiments of this application are used to explain this application and do not constitute an improper limitation on this application.

[0043] In addition, throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the components and elements are not necessarily drawn to scale.

[0044] In the attached picture:

[0045] Figure 1 This is a schematic diagram of the structure of the graphical machine learning algorithm platform.

[0046] Figure 2 It is a structural diagram of the feature extraction unit. DETAILED DESCRIPTION

[0047] Embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms, and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0048] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0049] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0050] Image segmentation in this solution is mainly to identify the corresponding objects in the image. For example, in a surveillance video of an intersection, there are multiple pedestrians, bicycles, buses and other features. In this picture, which parts are pedestrians and which parts are bicycles? The image needs to be segmented and then marked. This process is the image segmentation process. The graphical machine learning algorithm platform uses training samples for learning. After learning, it can automatically identify objects in the picture. For example, if 1,000 sample data are provided, each sample data will include the label of each object. After the 1,000 sample data are input into the graphical machine algorithm platform, the graphical machine algorithm platform will automatically learn the image segmentation rules in the 1,000 sample data, and then automatically iterate and update. After the update, when the unlabeled image information is input into the graphical machine algorithm platform, the image information will be automatically segmented, and then each segmented area will be marked.

[0051] Reference Figure 1 The graphical machine learning algorithm platform includes an information input module, an information segmentation module, an information description module, and a prediction module. Among them, the information input module performs smoothing filtering and denoising on the image information to reduce the noise in the image information. The image information is the picture that needs to be marked, which is in the common RBG image format. Before processing, smoothing filtering and denoising are required. In practice, the picture can also be grayscale processed. Whether grayscale processing is performed mainly depends on whether color information is introduced during segmentation. If color information is not introduced, the picture information needs to be grayscale processed.

[0052] The information segmentation module over-segments the image information to obtain several small blocks.

[0053] Over-segmentation is to divide the image information into many small areas, which are small blocks. If each small block is identified and then the small blocks of the same type are spliced ​​together, all objects in the image information will be constructed. Over-segmentation is to try to segment more areas when segmenting image information to make the range of each small block smaller. Specifically:

[0054] Specifically, the information segmentation module divides the image into M small blocks according to the edges based on the edge detection algorithm.

[0055] In this solution, when segmenting image information, an edge detection algorithm is used. The edge detection algorithm can distinguish the corresponding edge features from the image information, and use the edge features as the boundaries of the small blocks to segment the small blocks. The specific segmentation process is a prior art and will not be repeated here. In this solution, the image is segmented based on the edge detection algorithm, that is, the part of the image information with edge information is used as the segmentation basis.

[0056] For example: the Sobel operator is used to calculate the gradient of image information to detect edges, increase the sensitivity of edge detection, and thus generate more edges. The acquired edges are further refined through morphological operations (such as thinning and skeletonization), the edge contours are identified and tracked, and the area inside the closed contour is used as a segmentation block. The image information is represented as a graph, where nodes represent pixels or superpixels, and edges represent the similarity between pixels (or edge strength). Then, the image is divided into multiple small blocks through a graph cutting algorithm (such as minimum spanning tree, graph cut algorithm, etc.). Starting from the edge point or seed point, the area is gradually grown until another edge is encountered. After over-segmentation, some small blocks that do not meet the requirements (such as too small, irregular shape, etc.) may be generated. The segmentation results can be optimized through post-processing steps such as merging adjacent small blocks and filtering out too small areas.

[0057] In other ways, it can also be based on features such as image texture. In this solution, segmentation is performed based on an edge detection algorithm, mainly because objects in the field of image processing are more sensitive to edge information.

[0058] The information description module extracts the geometric features of all the patches for each small block, and selects a geometric feature from them as the standard feature of the small block.

[0059] After the image information is divided into multiple small blocks, the label corresponding to the small block is determined based on the relationship between the geometric features in the small block and the label. However, the range of the small block is relatively large, and when describing the geometric features of the small block, there will be more redundant information. To this end, this application provides the following solution:

[0060] The information description module includes: an information extraction unit and a dimension reduction unit.

[0061] Information extraction unit, for the i-th small block, extracts its n geometric features to obtain the feature vector F i ;

[0062] F i ={f i1 、f i2 ,…f ij …f in}, where f ij Represents the jth geometric feature of the i-th patch.

[0063] Specifically, the information extraction unit divides each small block i into n facets. For each small block, the number of facets is more than 10. Facets are the basic building blocks in a three-dimensional network model, consisting of vertices and edges, and are automatically divided according to geometric features.

[0064] After the image information is divided into small blocks, each small block will be automatically divided into multiple facets. Then the geometric features of each facet are calculated, so we can get n geometric features. Therefore, in this solution, the jth geometric feature of the ith small block is actually the geometric feature of the jth facet. The facet division method is generally automatic division using modeling tools, and the specific division method will not be described here.

[0065] The geometric features in this solution can be one feature or multiple features. It is mainly selected based on the model accuracy. In this implementation scheme, the geometric features include geometric center, SDF (shape diameter function), AGD (average geometric distance), SI (spin image), WKS (wavelet kernel signature), SIHKS scale-invariant heat kernel features. The extraction method of the above features is the existing technology, and the extraction method will not be repeated here.

[0066] In the information extraction unit, for each small block, the geometric features of each facet are extracted, so the feature dimensions of such small blocks are too many. It is necessary to reduce the dimension of the geometric features of each small block. To this end, the present application provides the following technical solutions:

[0067] Dimensionality reduction unit, for the feature vector F i The geometric features of each small block in the dimensionality reduction to the two-dimensional feature space, get n two-dimensional space point set A, A = {P i1 , P i2 ,…P ij …P in}, P ij Represents the data point of the jth geometric feature of the ith block in the two-dimensional feature space.

[0068] Among them, P ij The coordinates of (x ij ,y ij ), x ij =f ij *u1,y ij =f ij *u2, * represents the dot product operation, u1 represents the eigenvector F i The first principal component of u2 represents the eigenvector F i The first and second principal components here need to be obtained by principal component analysis, that is, the eigenvector F i As a set, after principal component analysis, the first principal component and the second principal component are obtained. Principal component analysis is a prior art, and the specific acquisition method is not further described here.

[0069] The information extraction unit calculates the local density ρ for each data point in the two-dimensional space point set A ij , ; where exp represents the exponential function, d c represents the cutoff distance parameter, Indicates P ij With P ik The Euclidean distance, j and k both represent the index of the data point;

[0070] For each data point P ij Calculate the minimum distance δ to a point with higher density than it j ;

[0071] By δ j is the vertical axis, ρ ij As the horizontal axis, draw the distribution map of all data points, select the cluster center based on the comprehensive value of the horizontal axis and the vertical axis, and use the geometric features represented by the data points corresponding to the cluster center as the standard feature. The comprehensive value here mainly considers the influence of the vertical axis and the horizontal axis. In this scheme, the cluster center in the distribution map of all data points is selected. The specific method for obtaining the cluster center is a prior art and will not be described here.

[0072] The data dimensionality reduction in this solution is actually a cluster analysis, which maps the geometric features of all faces into a two-dimensional space to obtain a scatter plot in the two-dimensional space. The scatter plot is analyzed to find the central point. After this data point is converted from the two-dimensional space to the three-dimensional space, the corresponding geometric features can be obtained.

[0073] Because the geometric feature may be one or more, when the geometric feature is one, it is processed according to the normal process. When there are multiple geometric features, there will be multiple corresponding feature vectors F i .

[0074] Specifically, if there are three geometric features, feature 1, feature 2, and feature 3, the feature vector will also include three feature vectors 1, 2, and 3, and the corresponding standard features will also include three. For feature 1, the feature 1 of all the facets in the small block is clustered for cluster analysis (dimensionality reduction) to obtain a standard feature. Similarly, the three geometric features can be reduced in dimension to obtain three standard features.

[0075] The standard features are the features provided in the aforementioned information description module. In this solution, the prediction module also includes a feature extraction unit, because geometric features are relatively low-level features, and it is difficult to find the connection between features and labels during the cycle. For this reason, in this solution, a feature extraction unit is set up, which converts the input standard features into high-order features. Specifically:

[0076] The feature extraction units include: fully connected network (FN), convolutional neural network (CNN, CN), and long short-term memory network (LSTM, LN).

[0077] In practice, it also includes geometric center, SDF (shape diameter function), AGD (average geometric distance), SI (spin image), WKS (wavelet kernel signature), and SIHKS scale-invariant heat kernel features.

[0078] Among them, the geometric center, SDF, and AGD are input into the fully connected network (FN) to obtain the first high-level features;

[0079] SI is input into the convolutional neural network to obtain the second high-level features;

[0080] WKS and SIHKS are input into the long short-term memory network to obtain the third high-level features;

[0081] The first high-level feature, the second high-level feature, and the third high-level feature are fused through weights or self-attention weights. Generally speaking, fully connected networks (FN), convolutional neural networks (CNN, CN), and long short-term memory networks (LSTM, LN) are separately trained network models, which are then used for feature extraction after being trained separately.

[0082] The training process of the fully connected network (FN), convolutional neural network (CNN, CN), and long short-term memory network (LSTM, LN) in the feature extraction unit is:

[0083] Data preparation: Collect several sample data. Each sample data includes the label and geometric features of a small block.

[0084] The sample data are used to train the fully connected network (FN), convolutional neural network (CNN, CN), and long short-term memory network (LSTM, LN) respectively. During the training process, the labels in the sample data are used as supervisory information.

[0085] Through the back-propagation algorithm and optimizer (such as SGD, Adam, etc.), the network parameters are updated to minimize the loss between the predicted label and the true label. In this way, the fully connected network (FN), convolutional neural network (CNN, CN), and long short-term memory network (LSTM, LN) can complete the corresponding training.

[0086] Feature fusion: After the three network parts are trained, their outputs are concatenated to form high-level geometric features.

[0087] The connection process here can be done by weight connection or self-attention weight connection.

[0088] After training is complete:

[0089] For a new small block, input the standard features of the small block;

[0090] The geometric center, SDF, and AGD are input into the fully connected network (FN) to obtain the first high-level features;

[0091] SI is input into the convolutional neural network to obtain the second high-level features;

[0092] WKS and SIHKS are input into the long short-term memory network to obtain the third high-level features;

[0093] The first high-level feature, the second high-level feature, and the third high-level feature are connected to form a final high-level geometric feature.

[0094] Therefore, in this scheme, a feature extraction unit will be used to extract the standard features of each small block to obtain high-level geometric features.

[0095] The reason why features are extracted in this way in this scheme is mainly to adapt to the characteristics of different geometric features. Different geometric features have different characteristics and structures. The geometric center, SDF, and AGD use fully connected networks (FN), which can better discover the change information between numerical values. SI uses convolutional networks, which takes advantage of the greater advantage of convolutional networks in processing two-dimensional information. For WKS and SIHKS, which have information that depends on the sequence before and after, a long short-term memory network is used, so that the changes in sequence information can be captured. In this way, the feature extraction unit provided in this scheme has a higher generalization ability.

[0096] The prediction module in this solution is essentially a neural network model. In the iterative and cyclic process, it can learn the intrinsic relationship between standard features (high-level geometric features) and labels, and then automatically label the image information. Specifically:

[0097] The initialization unit initializes each small block and initializes the label of the small block to zero. This step is an initialization link.

[0098] The iteration unit randomly selects a small block i and selects a segmentation label according to the state information of the small block i. The state information includes the standard features of the small block, the standard features of the adjacent small blocks, the standard features of the adjacent small blocks, and the labels of the adjacent small blocks.

[0099] In this scheme, the reason why the standard features of the surrounding small blocks are considered is to facilitate finding the distribution pattern of the labeled objects in the image information, so the training data provided can be less, or each training data does not need to be fully labeled.

[0100] The iteration unit is actually the embodiment of learning and improvement. The process of setting labels for small blocks in the iteration unit is generally controlled by the step size. The longer the step size, the larger the range of label selection during iteration.

[0101] The reward calculation unit calculates the reward value r of the segmentation label assigned to the small block i;

[0102] The reward value is essentially an adaptability function. The larger the reward value r is, the higher the operation of the iterative unit is, and vice versa.

[0103] The loop unit guides the iterative unit to loop continuously until the maximum cumulative reward value R is obtained.

[0104] In this solution, the kernel of the iteration unit is the hidden layer of the neural network model to be trained, which is the RLSegNet network. The loop unit controls the number of iterations during training, and the reward calculation unit is used to determine the effectiveness of this iteration operation.

[0105] Specific:

[0106] Furthermore, in the iteration unit:

[0107] Status information includes S wi , S ei , S li ;

[0108] For a small block i, its standard characteristic is S wi , the average standard feature of the remaining small blocks adjacent to small block i is S ei ; The label distribution of the remaining small blocks adjacent to small block i is S li ;

[0109] S ei ={E 1i 、E 2i 、E 3i …E ei …}, where E ei is the standard feature of the e-th block among the remaining blocks adjacent to block i;

[0110] S li ={ L 1i 、L 2i 、L 3i …L li ...}, L li is a distribution vector, where l The element represents the number of small blocks adjacent to small block i. l The label of a small block; the adjacent small block generally refers to the small block adjacent to the small block. In this scheme, it mainly refers to the small block within the first ring, that is, the small block directly adjacent to the small block. In practice, it can also be the small block within the second ring, that is, the small block indirectly adjacent.

[0111] Furthermore, after all small blocks are assigned labels, the recurrent unit selects small blocks with different labels from surrounding small blocks from all small blocks, and then inputs them into the iteration unit for re-iteration.

[0112] Furthermore, the model is affected by the reward value during the iteration process, and the design of the reward value will affect the convergence rate of the model and increase the training cost of the model. For this purpose, the present application provides the following technical solutions:

[0113] Furthermore, the reward value is r t , , is a Gaussian function, min(di) represents the distance from small block i to the true segmentation boundary G i The shortest geodesic distance among all points on the network. The geodesic distance refers to the length of the shortest path from one point to another. i This refers to the correct segmentation result of small patch i. The minimum geodesic distance from small patch i to the true segmentation boundary Gi is the label of small patch i. α and β are a pair of proportional parameters, r t represents the index of the tth iteration, t represents the index of the number of iterations, a t represents the action at the tth iteration, so a t =G i In this scheme, α=1 and β=1.

[0114] Specifically, the training process is as follows:

[0115] S1: Initialize the parameters of the neural network model RLSegNet, such as setting the number of iterations, step size, α, and β.

[0116] S2: For each training image, split it into small blocks and initialize the label of each small block to be empty or random.

[0117] S3: Start iteration:

[0118] For each small block i, calculate its standard feature S wi .

[0119] Calculate the average standard feature S of the adjacent small blocks of small block i ei .

[0120] Calculate the label distribution S of the adjacent small blocks of small block i li .

[0121] S wi、 S ei , S li As state information, it is input into the RLSegNet network, and the network outputs the predicted label of small patch i.

[0122] S4: For each small block i, segment the boundary G according to its predicted label and true value i , calculate the minimum geodesic distance min(d i ).

[0123] Calculate the reward value r according to the Gaussian function t , where α=1, β=1, and the cumulative reward value is R.

[0124] Cyclic unit operation:

[0125] Check whether all small blocks have been assigned labels, filter out small blocks with different labels from surrounding small blocks, and re-input these small blocks into the iteration unit for re-iteration. Repeat the iteration process until the maximum number of iterations is reached or the cumulative reward value R no longer increases significantly.

[0126] S4: Based on the cumulative reward value R and the result of the reward calculation unit, the back propagation algorithm is used to update the parameters of the RLSegNet network. Until the RLSegNet network converges or reaches the preset number of training times: save the trained RLSegNet model parameters.

[0127] To increase the convergence speed of the model:

[0128] In this scheme, α and β are dynamically changing, and α and β are related to the number of iterations. Let the current number of iterations be t, and the total number of iterations be T; Among them, U and k are pre-set control parameters, U=3, k=1, and exp is an exponential function. In this way, in this scheme, at the beginning of the iteration, the growth rate of the reward function is low, that is, the reward and penalty are small, so the iterative unit has more choice space during iteration, which can prevent the model from falling into the local optimal solution. In the second half of the iteration, the growth rate of the reward function will increase. At this time, if the choice is wrong, the reward or penalty will be greater. Therefore, when iterating the parameters, more correct operations will be selected as much as possible to increase the convergence rate of the model.

[0129] When predicting:

[0130] Step 1: Load the trained RLSegNet model parameters.

[0131] Step 2: Divide the image to be predicted into small blocks and calculate the standard feature S of each small block wi .

[0132] Step 3: For each small block i, initialize its label to empty or random label, and calculate the average standard feature S of the small blocks adjacent to the small block i ei and label distribution S li , S wi , S ei , S liAs state information, it is input into the RLSegNet network, and the network outputs the predicted label of small block i. The iterative process is repeated until all small blocks are assigned labels or the preset number of iterations is reached. The prediction results are post-processed, such as smoothing boundaries and removing isolated points, to improve the accuracy of the segmentation results.

[0133] Output segmentation results: Integrate the prediction results into a complete image segmentation map and output it.

[0134] The above description is only some preferred embodiments of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present application is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above invention concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the embodiments of the present application to form a technical solution.

Claims

1. A graphical machine learning algorithm platform, characterized by: include: An information input module is used to input image information to be processed so as to pre-process the image information; The information segmentation module over-segments the image information to obtain several small blocks; The information description module extracts the geometric features of all the patches for each small block, and selects a geometric feature from them as the standard feature of the small block; The prediction module obtains the standard features of each small block, automatically segments the image and annotates the entities based on the differences between the standard features; The prediction module includes: An initialization unit initializes each small block and initializes the label of the small block to zero; The iteration unit randomly selects a small block i and selects a segmentation label according to the state information of the small block i. The state information includes the standard features of the small block, the average standard features of the adjacent small blocks, and the label distribution of the remaining adjacent small blocks; Reward calculation unit, calculates the reward value r of the segmentation label assigned to small block i t ; The loop unit guides the iteration unit to loop continuously until the maximum cumulative reward value R is obtained; In the iteration unit: Status information includes S wi , S ei , S li ; The standard characteristic of small block i is S wi , the average standard feature of adjacent small blocks is S ei , the label distribution of the remaining adjacent small blocks is S li ; S ei ={E 1i 、E 2i 、E 3i …E ei …}, where E ei is the standard feature of the e-th block among the remaining blocks adjacent to block i; S li ={ L 1i 、L 2i 、L 3i …L li ...}, L li is a distribution vector, where l The element represents the number of small blocks adjacent to small block i. l A small piece of label; Each small block is divided into multiple patches, which are the basic building blocks in the three-dimensional network model; The information description module includes: Information extraction unit, for the i-th small block, extracts its n geometric features to obtain the feature vector F i ; F i ={f i1 、f i2 ,…f ij …f in }, where f ij Represents the jth geometric feature of the i-th small block; Dimensionality reduction unit, for the feature vector F i The geometric features f of each small block in ij Reduce the dimension to two-dimensional feature space and obtain n two-dimensional space point sets A; A={P i1 , P i2 ,…P ij …P in }, P ij Represents the data point of the jth geometric feature of the i-th small block in the two-dimensional feature space; Among them, P ij The coordinates of (x ij ,y ij ), x ij =f ij *u1,y ij =f ij *u2, * represents the dot product operation, u1 represents the eigenvector F i The first principal component of F, u2 represents the eigenvector F i The second principal component of The information extraction unit calculates the local density ρ for each data point in the two-dimensional space point set A ij , ; where exp represents the exponential function, d c represents the cutoff distance parameter, Indicates P ij With P ik The Euclidean distance, j and k both represent the index of the data point; For each data point P ij Calculate the minimum distance δ to a point with higher density than it j ; By δ j is the vertical axis, ρ ij As the horizontal axis, draw the distribution diagram of all data points, select the cluster center based on the comprehensive values ​​of the horizontal and vertical axes, and use the geometric features represented by the data points corresponding to the cluster center as the standard features.

2. The graphical machine learning algorithm platform according to claim 1, characterized in that: The information input module performs smoothing filtering and denoising on the image information to reduce the noise in the image information.

3. The graphical machine learning algorithm platform according to claim 1, characterized in that: The information segmentation module divides the image into M small blocks according to the edges based on the edge detection algorithm.

4. The graphical machine learning algorithm platform according to claim 1, characterized in that: The prediction module has a built-in neural network model, which learns the correspondence between standard features and labels and automatically labels each small block.

5. The graphical machine learning algorithm platform according to claim 1, characterized in that: The reward value is r t , ; is a Gaussian function, min(d i ) represents the segmentation boundary G from small block i to the true value i The shortest geodesic distance among all points on the G, the geodesic distance refers to the length of the shortest path from one point to another, the distance from patch i to the true segmentation boundary G i The minimum geodesic distance is the label of patch i, α and β are a pair of proportional parameters, t is the number of iterations, a t Represents the action at the tth iteration.

6. The graphical machine learning algorithm platform according to claim 5, characterized in that: ; t is the current iteration number, T is the total number of iterations; where U and k are pre-set control parameters, U=3, k=1, and exp is an exponential function.

Images