An image processing method and system based on deep learning

Abstract

The application discloses a kind of based on deep learning's image processing method and system, it is related to deep learning image processing technical field.The method comprises the following steps: receiving target image and performing multistage semantic feature extraction to obtain basic semantic feature map;Dynamic receptive field adjustment operation of spatial perception is executed to basic semantic feature map, and adaptive spatial feature map is obtained;According to adaptive spatial feature map, image semantic graph containing global context association is constructed;The node state iteration update operation on multiple graph structures is executed to image semantic graph, and after each update, the current node state matrix is updated according to graph topological structure, until the target image semantic graph is obtained by meeting preset stop condition;Multilevel feature reconstruction is carried out to each node feature in target image semantic graph, and image semantic segmentation graph is generated.The method adaptively adjusts receptive field and dynamically optimizes image structure, and can effectively improve the precision of image semantic segmentation.

Description

Technical Field

[0001] This invention relates to the field of deep learning image processing technology, specifically to an image processing method and system based on deep learning. Background Technology

[0002] Image semantic segmentation, aiming to assign semantic category labels to each pixel in an image, is a fundamental task in computer vision. Existing deep learning-based image semantic segmentation methods typically employ fully convolutional networks or encoder-decoder architectures, extracting image features by stacking convolutional and pooling layers and progressively restoring spatial resolution to generate pixel-by-pixel predictions. These methods generally use convolutional kernels with fixed receptive fields during feature extraction, such as dilated convolutions with a single dilation rate to expand the receptive field. However, target objects in images exhibit significant differences in scale, shape, and spatial distribution. Convolutional operations with fixed receptive fields cannot adapt to these spatial variations, leading to small-scale targets being easily overlooked while large-scale targets lack sufficient contextual information, thus limiting segmentation accuracy. Simultaneously, to capture global contextual information, some methods introduce graph neural networks to model the relationships between image regions, treating pixels or superpixels in the feature map as graph nodes and constructing an initial graph structure based on the similarity of node features. However, in subsequent graph information transmission, the graph topology usually remains unchanged, and node state updates are limited to the initially determined neighborhood, making it difficult to discover new, long-distance semantic associations during iteration. This static graph structure fails to fully utilize the gradually enriched node state information during iteration to adjust connectivity, thus limiting long-distance context modeling capabilities and making it difficult to guarantee semantic consistency between regions. Therefore, how to adaptively adjust the receptive field to match target scale changes at different spatial locations, and how to dynamically optimize the graph topology during node state iteration to more effectively capture global semantic interactions, are the problems that this technical solution aims to solve. Summary of the Invention

[0003] This invention provides an image processing method and system based on deep learning, aiming to overcome the shortcomings of existing image semantic segmentation methods, such as the inability of fixed receptive fields to adapt to changes in spatial scale and the limitation of global context modeling capabilities by static graph structures. By adjusting the spatially aware dynamic receptive field and updating the dynamic graph topology based on the node state matrix, the accuracy and region consistency of semantic segmentation are improved.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides an image processing method and system based on deep learning. The method is applied to a first computing node in an image processing system, which includes multiple computing nodes and an image storage node. The first computing node is any one of the computing nodes. The method includes: receiving a target image sent by an image storage node; performing multi-level semantic feature extraction operations on the target image to obtain a basic semantic feature map; performing a spatially aware dynamic receptive field adjustment operation on the basic semantic feature map to obtain an adaptive spatial feature map; constructing an image semantic map of the target image based on the adaptive spatial feature map, wherein the image semantic map contains the global contextual relationships between various image regions in the target image; performing multiple iterative update operations on the node state on the graph structure of the image semantic map, and updating the graph topology of the image semantic map according to the node state matrix of the current iteration after each node state update operation, until a preset iteration stopping condition is reached to obtain the target image semantic map; performing multi-level feature reconstruction operations on the features of each node in the target image semantic map to generate an image semantic segmentation map corresponding to the target image.

[0005] Through multi-level semantic feature extraction, visual information of target images can be captured from different levels of abstraction, such as shallow texture, mid-level structure, and high-level semantics. Cross-level features are then superimposed and fused, resulting in a basic semantic feature map that possesses both detailed expressive power and global semantic understanding capabilities. After obtaining the basic semantic feature map, a spatially aware dynamic receptive field adjustment is performed. This adaptively determines the receptive field range of each local region based on the spatial relationship between local regions of the image, thereby applying differentiated feature extraction intensities to image content at different spatial locations. This enhances the adaptability to scale changes and spatial structural diversity, resulting in a more discriminative adaptive spatial feature map.

[0006] When constructing an image semantic graph based on adaptive spatial feature maps, each pixel position is treated as a graph node, and the pixel feature vector is used as the initial node feature. Edge connections between nodes are established based on feature similarity, explicitly associating semantically similar but spatially disjoint regions in the image within the graph structure, effectively representing the global contextual relationships in the target image. Multiple rounds of node state iterative updates are performed on the image semantic graph. In each round, an attention mechanism is used to calculate the interaction weights between the target node and its neighboring nodes, and neighborhood information is selectively aggregated. Based on this, a gating fusion mechanism adaptively integrates the original node state and the aggregated neighborhood state, ensuring that the node state retains its own characteristics while absorbing effective contextual information during information propagation.

[0007] As a technical solution of this invention, after each iteration of updating the node state, the graph topology is rebuilt based on the node state matrix of the current round. Global state correlation is obtained by calculating the covariance matrix of the node state matrix. For each node, a subset of nodes with the strongest correlation are selected to establish new edge connections, while existing edge connections are broken. This allows the graph topology to dynamically adjust as node features evolve, correcting semantically biased connections that may exist in the initial image semantic graph and strengthening information pathways between truly semantically consistent nodes. Through this dynamic graph topology update mechanism, the graph structure can gradually tend towards a more reasonable semantic association state, thereby improving the semantic consistency and boundary accuracy of the final generated image semantic segmentation graph.

[0008] Preferably, after the preset iteration stopping condition is met and the semantic map of the target image is obtained, the updated node state vectors of each graph node are mapped back to the original pixel positions in the adaptive spatial feature map to form a reconstructed feature map. This reconstructed feature map is then progressively upsampled and edge-corrected to restore it to the same spatial resolution as the target image. Class probability mapping is performed on the restored resolution feature map to determine the corresponding semantic class label for each pixel, ultimately generating an image semantic segmentation map. Through multi-level feature reconstruction and progressive upsampling, the abstract semantic information of the graph structure can be accurately restored to pixel-level prediction results, ensuring the fidelity of the segmentation results at the spatial detail level.

[0009] As a technical solution of the present invention, when performing multi-level semantic feature extraction operations on the target image, it is implemented through a pre-configured deep convolutional network in the first computing node. The first encoder module with the smallest downsampling stride is used to perform convolution, batch normalization, and activation function mapping on the target image to obtain an initial shallow texture feature map with a spatial resolution between the target image and subsequent intermediate layer feature maps. This provides the texture foundation required for subsequent layers while preserving more spatial details. Then, the intermediate convolutional module and the high-level convolutional module sequentially perform semantic abstraction and global semantic encoding operations to obtain intermediate layer structural feature maps and high-level semantic feature maps, respectively. These three maps are then fused across different layers to fully extract semantic information from the image at different receptive field scales.

[0010] In the spatially aware dynamic receptive field adjustment operation, the basic semantic feature map is divided into a spatially dimensional grid to obtain multiple local spatial feature blocks. For each local spatial feature block, the receptive field scale adjustment coefficient is obtained by calculating the spatial Euclidean distance between its center position and the centers of other blocks and performing a normalized mapping. This coefficient represents the effective neighborhood range corresponding to the local spatial feature block in the subsequent feature extraction process. Based on this, dilated convolution operations with different dilation rates are applied to different local spatial feature blocks, where the dilation rate is positively correlated with the receptive field scale adjustment coefficient. Then, the updated local spatial feature blocks are concatenated according to their original spatial positions to obtain an adaptive spatial feature map. Through the above operations, the receptive field of feature extraction can be adaptively adjusted according to the spatial context content. A smaller receptive field is maintained at object edges or areas with dense details to preserve detail information, while a larger receptive field is used in flat or semantically consistent areas to obtain more sufficient contextual information.

[0011] In the construction of the image semantic graph, pixel-level feature vector extraction is performed on the adaptive spatial feature map. Each pixel is treated as a graph node, and its corresponding feature vector is used as the initial node feature. The feature similarity between the initial node features of any two graph nodes is calculated, and node pairs with feature similarity greater than a preset threshold are identified as initial edge connections, thus forming the image semantic graph. This construction method enables semantically similar but spatially separable image regions to be connected in the graph structure, providing an efficient structural foundation for subsequent information propagation and semantic relationship reasoning on the graph structure.

[0012] In the node state iterative update operation, for each target graph node, all neighboring graph nodes with edge connections to it are obtained. The current node state vector of the target graph node and the current node state vectors of each neighboring graph node are input into the attention weight calculation function to obtain the attention interaction weights of each neighboring graph node relative to the target graph node. Based on this weight, the current node state vectors of all neighboring graph nodes are weighted and aggregated to form a neighborhood aggregated state vector. Then, a gating fusion operation is performed on the current node state vector of the target graph node and the neighborhood aggregated state vector. The gating fusion operation uses a learnable gating weight vector to weight and sum the two, where the gating weight vector is generated by nonlinear mapping from the current node state vector and the neighborhood aggregated state vector. The updated node state vector obtained in this way retains the original features while adaptively fusing neighborhood context information, effectively enhancing the semantic representation ability of node features.

[0013] Regarding the updating of the graph topology, the updated node state vector of each graph node in the current iteration is obtained and arranged in the order of the graph nodes to form the node state matrix of the current iteration. The covariance matrix of this matrix is ​​calculated to obtain the global state correlation matrix of the image semantic graph. For each target graph node, its correlation coefficient vector with all other nodes is extracted from the global state correlation matrix. A predetermined number of nodes with the highest ranking are selected as candidate reconnection nodes. All original edge connections of the target graph node are disconnected in the image semantic graph, and new edge connections are established between the target graph node and the candidate reconnection nodes, thereby obtaining the updated graph topology. In this way, the evolution of the graph structure is driven by the statistical correlation of the node features themselves, so that the updated graph topology can better reflect the true semantic associations of the current node state.

[0014] The preset iteration stopping condition can be set to the current iteration round reaching the preset maximum number of iterations, or the change in the graph topology of the image semantic map being less than a preset change threshold. The target image semantic map obtained after meeting the stopping condition has a stable and reasonable node feature representation and graph topology, providing an accurate basis for the subsequent generation of image semantic segmentation maps.

[0015] In the multi-level feature reconstruction process, the updated node state vector of each graph node in the semantic graph of the target image is mapped back to the original pixel position in the adaptive spatial feature map to obtain the reconstructed feature vector and form the reconstructed feature map. The reconstructed feature map is then subjected to a first-level upsampling operation and a second-level upsampling operation sequentially, ensuring that the spatial resolution of the second-level upsampled feature map matches the spatial resolution of the target image. Further, after the second-level upsampling, the input image size information of the target image at the first computation node is obtained. Based on this size information, an edge pixel cropping operation is performed on the second-level upsampled feature map to obtain an edge-cropped feature map whose spatial resolution is equal to the original spatial resolution of the target image. This eliminates any size deviation that may be introduced during the upsampling process and ensures a one-to-one correspondence between pixel positions.

[0016] In the category probability mapping operation, the feature map after edge cropping is input into a preset pixel-wise classifier, which includes a fully connected mapping layer and a normalized exponential function mapping layer. The fully connected mapping layer performs a linear mapping on the feature vector of each pixel location to obtain the initial category score vector corresponding to each preset image semantic category. Then, the normalized exponential function mapping layer performs a probability normalization operation on the initial category score vector to obtain the category probability value of the pixel location belonging to each preset image semantic category. The preset image semantic category corresponding to the maximum probability value is selected as the semantic category label of the pixel location, and finally, an image semantic segmentation map is generated to achieve high-precision pixel-level semantic segmentation of the target image.

[0017] This invention also provides a deep learning-based image processing system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned image processing method. Through the collaborative work of multiple computing nodes, this system can organically connect image reception, feature extraction, graph structure construction and updating, feature reconstruction, and semantic segmentation output, ensuring segmentation accuracy while possessing good computational parallelism and scalability.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The basic semantic feature map is divided into spatial grids to obtain multiple local spatial feature blocks. For each local spatial feature block, the spatial Euclidean distance between its center and the centers of the other local spatial feature blocks is calculated. After normalization mapping, a receptive field scale adjustment coefficient is generated for that local spatial feature block. This coefficient characterizes the effective neighborhood range of the local spatial feature block in subsequent feature extraction. Based on this coefficient, dilated convolution operations with different dilation rates are performed on each local spatial feature block. The dilation rate is positively correlated with the receptive field scale adjustment coefficient, allowing for the use of a smaller dilation rate in densely distributed spatial regions to preserve local details, and a larger dilation rate in sparsely distributed spatial regions to capture broader contextual information. In this way, each local spatial feature block obtains an adaptive receptive field that matches its spatial location characteristics. These are then concatenated to form an adaptive spatial feature map, effectively solving the problem of insufficient processing capability of fixed-receptive-field convolution for multi-scale targets, and improving the consistency and completeness of feature representation of objects at different scales. In the multiple iterations of node state updates in the image semantic graph, after each iteration updates the node state vector, the covariance matrix is ​​calculated using the node state matrix generated in the current iteration, resulting in the global state correlation matrix of the image semantic graph. For each target graph node, its correlation coefficient vector with all other graph nodes is extracted from the correlation matrix, and a predetermined number of graph nodes with the highest correlation coefficients are selected as candidate reconnection nodes based on this vector. Subsequently, all original edge connections of the target graph node are disconnected, and new edge connections are established between it and the candidate reconnection nodes, forming an updated graph topology. This dynamic graph topology update mechanism allows the graph structure to be continuously adjusted according to the correlation of node states during the iteration process. It re-establishes information transmission paths between nodes that may not have direct connections in the initial stage but have strong semantic relevance, thereby breaking the limitations of static graph structures, making long-distance semantic interactions more complete, significantly enhancing the global context modeling capability, and ultimately improving the object boundary and region consistency of the generated image semantic segmentation graph. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart of an image processing method based on deep learning; Figure 2 This is a flowchart of multi-level semantic feature extraction and dynamic receptive field adjustment; Figure 3 This is a schematic diagram of the image semantic graph construction and graph node state iterative update process; Figure 4 This is a flowchart of the iterative update process of the graph topology structure of the image semantic graph; Figure 5 This is a schematic diagram of the multi-level feature reconstruction and upsampling decoding process; Figure 6 It is a curve showing the relationship between the receptive field scale adjustment coefficient and the hole convolution dilation rate based on spatial distance; Figure 7 It is the curve showing the cross-entropy loss versus the learning rate during the training process. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See Figure 1This invention provides a deep learning-based image processing method applied to a first computing node in an image processing system. The image processing system includes multiple computing nodes and an image storage node, with the first computing node being any one of these nodes. The method includes: receiving a target image sent by the image storage node and performing multi-level semantic feature extraction operations on the target image to obtain a basic semantic feature map; performing a spatially aware dynamic receptive field adjustment operation on the basic semantic feature map to obtain an adaptive spatial feature map; constructing an image semantic map of the target image based on the adaptive spatial feature map, the image semantic map containing global contextual relationships between various image regions in the target image; performing multiple iterative update operations on the node state of the image semantic map, and updating the graph topology of the image semantic map based on the node state matrix of the current iteration after each iteration, until a preset iteration stopping condition is met to obtain the target image semantic map; and performing multi-level feature reconstruction operations on the features of each node in the target image semantic map to generate an image semantic segmentation map corresponding to the target image.

[0023] Example 1:

[0024] In specific implementation, please refer to Figure 2 The multi-level semantic feature extraction operation is performed by a pre-configured deep convolutional network in the first computing node. The deep convolutional network contains multiple sequentially connected convolutional modules, specifically a first-level convolutional module, intermediate convolutional modules, and high-level convolutional modules. The first-level convolutional module uses convolutional layers with a stride of a preset first value, which can be set to 2, to perform preliminary downsampling and local texture response extraction on the target image. The intermediate convolutional module consists of multiple stacked residual units, each containing two convolutional layers and an identity mapping connection, used for semantic abstraction of the input feature map. The high-level convolutional module contains dilated convolutional layers and global average pooling layers to capture global contextual information. After the target image is input into the deep convolutional network, the first-level convolutional module performs a first-level convolutional mapping operation on the target image, generating an initial shallow texture feature map. The initial shallow texture feature map is then fed into the intermediate convolutional module, which performs a second-level semantic abstraction operation on the initial shallow texture feature map, outputting an intermediate layer structure feature map. The intermediate layer structural feature map is fed into the high-level convolutional module, which performs a third-level global semantic encoding operation on the intermediate layer structural feature map, outputting a high-level semantic feature map. The cross-level feature overlay and fusion operation is as follows: the initial shallow texture feature map, intermediate layer structural feature map, and high-level semantic feature map are upsampled to the same spatial resolution using bilinear interpolation, which is set to half the spatial resolution of the target image; the three upsampled feature maps are concatenated along the channel dimension to obtain a concatenated feature map; the concatenated feature map is input into a fusion layer consisting of a 1×1 convolutional layer and a batch normalization layer to obtain the basic semantic feature map.

[0025] The implementation process of the spatially aware dynamic receptive field adjustment operation is as follows. In the spatial dimension meshing operation, the height of the basic semantic feature map is denoted as H, the width as W, and the number of channels as C. The basic semantic feature map is uniformly divided into non-overlapping local spatial feature blocks in the spatial dimension. The spatial size of each local spatial feature block is P×P, where P is a positive integer, typically set to 8. The total number of local spatial feature blocks obtained is N=(H / P)×(W / P), and each local spatial feature block contains P×P×C feature values. For each local spatial feature block, its spatial center pixel position is taken as the center position of that local spatial feature block. The coordinates of the center position are determined by the row and column index of the local spatial feature block in the basic semantic feature map. For the k-th local spatial feature block, the spatial Euclidean distance between the center position of the k-th local spatial feature block and the center positions of all other local spatial feature blocks in the basic semantic feature map is calculated, forming the spatial distance vector corresponding to the k-th local spatial feature block. The spatial distance vector contains N-1 distance values. The normalization mapping operation performed on the spatial distance vector is as follows: first, calculate the average of all distance values ​​in the spatial distance vector, denoted as . Then, the receptive field scale adjustment coefficient corresponding to the k-th local spatial feature block is obtained through a preset mapping function. The preset mapping function is:

[0026] in, This represents the receptive field scale adjustment coefficient corresponding to the k-th local spatial feature block. This represents the average of the distance values ​​in the spatial distance vector of the k-th local spatial feature block. The preset scale adjustment parameters, The value is a positive real number, and in this embodiment it is set to... The receptive field scale adjustment coefficient is used to characterize the effective neighborhood range corresponding to the k-th local spatial feature block in subsequent feature extraction. The larger the receptive field scale adjustment coefficient, the more contextual neighborhood is needed for feature extraction of the k-th local spatial feature block.

[0027] Multiple updated local spatial feature blocks are obtained through dilated convolution operations with different dilation rates. For each local spatial feature block, the dilation rate of the dilated convolution is determined based on the receptive field scale adjustment coefficient corresponding to that local spatial feature block. The dilation rate is calculated by multiplying the receptive field scale adjustment coefficient by a preset maximum dilation rate and rounding down. The maximum dilation rate can be set to 4. When the calculated dilation rate is less than 1, it is set to 1. For the k-th local spatial feature block, a dilated convolution operation is performed on it using a dilated convolution kernel with the dilation rate calculated above. The size of the dilated convolution kernel is fixed at 3×3, with C input channels and C output channels. After completing the dilated convolution operation on all local spatial feature blocks, N updated local spatial feature blocks are obtained, and the spatial size of each updated local spatial feature block remains P×P×C. Multiple updated local spatial feature blocks are concatenated according to their original spatial positions in the basic semantic feature map. That is, with a step size of P in the height dimension and P in the width dimension, the updated local spatial feature blocks are sequentially filled into the corresponding regions to obtain an adaptive spatial feature map with the same spatial size as the basic semantic feature map.

[0028] Example 2:

[0029] In specific implementation, please refer to Figure 3 In dilated convolution operations with different dilation rates, the relationship between the dilation rate of each local spatial feature block and the receptive field scale adjustment coefficient is set to be strictly positively correlated. That is, for the k-th local spatial feature block, its dilation rate is denoted as... , ,in This is the receptive field scale adjustment coefficient corresponding to this local spatial feature block. The pre-defined expansion rate mapping factor. The value is 4. This indicates rounding up. When the calculated result is... When it is less than 1, The value is forced to be 1. The larger the receptive field scale adjustment coefficient, the greater the corresponding expansion rate.

[0030] The process of constructing an image semantic map of the target image based on the adaptive spatial feature map is executed by the semantic map construction module deployed in the first computing node. The spatial size of the adaptive spatial feature map is... The number of channels is The pixel-level feature vector extraction operation performed on the adaptive spatial feature map is implemented through a 1×1 convolutional layer with an input channel number of... The number of output channels is , The value is 64. The kernel stride of this convolutional layer is 1, and no padding is used. After the convolution operation, the position of each pixel in the adaptive spatial feature map is... A corresponding one 3D pixel feature vector, denoted as ,in , Arrange all pixel feature vectors in row-major order to obtain the graph node set, and the total number of graph nodes. Each pixel position is treated as a graph node, and the pixel feature vector corresponding to that pixel position is used as the initial node feature of that graph node. The initial node feature vector is denoted as... , .

[0031] The feature similarity between the initial node features of any two graph nodes in the image semantic graph is measured using cosine similarity. For graph nodes... Graph Nodes Their feature similarity The calculation method is as follows The preset similarity threshold is set to... , The specific value is 0.6. When At that time, determine the graph nodes. With graph nodes There is an initial edge connection between them; if Then the graph nodes With graph nodes There are no initial edge connections between them. Based on the above judgment, traverse all graph node pairs to generate the initial adjacency matrix of the image semantic graph. ,in Represents graph nodes With graph nodes There are edges connecting them. This indicates no edge connections. The initial adjacency matrix and the initial node features of each graph node together constitute the image semantic graph of the target image.

[0032] Perform multiple iterations of node state update operations on the image semantic graph. In each node state update operation, perform the following steps.

[0033] For each target graph node in the image semantic graph Obtain the target graph nodes from the adjacency matrix of the current graph topology. The set of all adjacency graph nodes , Defined as the node in the target graph that is in the current adjacency matrix. All graph nodes that are connected by edges. (Target graph node) The current node state vector is denoted as Adjacency graph nodes The current node state vector is denoted as .Will and each Input the attention weight calculation function separately to obtain each adjacency graph node. Relative to the target graph node Attention interaction weights The attention weight calculation function is implemented using a learnable bilinear mapping: ,in It is a learnable attention weight matrix. This indicates the transpose operation.

[0034] Based on each adjacency graph node Attention interaction weights A weighted aggregation operation is performed on the current node state vectors of all adjacent graph nodes to obtain the target graph nodes. neighborhood aggregation state vector The calculation method is as follows .

[0035] Target graph nodes Current node state vector Aggregate state vector with neighborhood Perform gating fusion operations to obtain the target graph nodes. The updated node state vector after this node state iteration update operation is completed is gated and fused using the following formula:

[0036] in, This is the updated node state vector; This represents the sigmoid activation function. ; and These are learnable weight matrices, used to perform linear transformations on the current node's state vector and the neighborhood aggregated state vector, respectively. It is a learnable bias vector; This represents element-wise vector multiplication. In gating fusion operations, the gating weight vector is... The given gating weight vector is generated through a nonlinear mapping of the current node's state vector and the neighborhood aggregation state vector, and is used to control the fusion ratio of new and old information. After completing the above operations, the target graph nodes... The node state vector is updated to In each node state iteration update operation, the above update process is performed on all graph nodes in the image semantic graph, thereby completing one node state iteration update on the graph structure. All the above learnable parameters, including... , , and All parameters were determined by training on an image semantic segmentation dataset using an end-to-end backpropagation algorithm, with the cross-entropy loss function used during training.

[0037] See Figure 6 In the figure, the horizontal axis represents the average spatial distance vector of local spatial feature blocks, in pixels, with values ​​ranging from 0 to 5; the left vertical axis represents the receptive field scale adjustment coefficient. The range is from 0 to 1.0; the right vertical axis represents the dilation rate of dilated convolution. The range is 1 to 4. In the figure, the gray scatter dots represent the sample point locations corresponding to the average spatial distance of each local spatial feature block, and the black solid curve represents the receptive field scale adjustment coefficient calculated according to the preset mapping function. The trend curve showing the change with the average spatial distance exhibits a monotonically decreasing exponential decay, indicating that as the average distance between a local spatial feature block and other feature blocks increases, its receptive field scale adjustment coefficient significantly decreases. The black dashed step-like curve represents the relationship between the receptive field scale adjustment coefficient and the expansion rate, combined with the expansion rate mapping factor. The dilation rate of the dilated convolution calculated by rounding up. The dilation rate curve shows a step-like decreasing trend. When the receptive field scale adjustment coefficient is high and the average spatial distance is small, the dilation rate reaches a maximum of 4. As the average spatial distance increases, the dilation rate decreases to 3, 2, and 1 respectively. This reflects that different local spatial feature blocks automatically adjust the dilation rate of the dilated convolution according to the environmental distance of their spatial location, thereby realizing the spatial adaptive adjustment of the dynamic receptive field.

[0038] Example 3:

[0039] In specific implementation, please refer to Figure 4 The preset iteration stopping condition consists of two criteria. The first criterion is that the current iteration round reaches the preset maximum number of iterations, denoted as [the preset maximum number of iterations is denoted as 1]. , The specific value is preset to 10. The second criterion is that the change in the graph topology of the image semantic graph is less than a preset change threshold, which is denoted as [missing value]. , The specific value is preset to 0.01. After each node state iteration update operation, the above two judgment criteria are checked. When the current iteration round equals When, or when the graph topology of the image semantic graph changes by less than When the node state iteration update operation stops, the current image semantic graph is used as the target image semantic graph. The graph topology change of the image semantic graph is calculated as follows: obtain the first... The adjacency matrix generated after the secondary node state iteration update operation And the adjacency matrix generated after the previous round of node state iteration update operation. ,when hour This represents the initial adjacency matrix of the image semantic graph. The graph topology changes. The calculation formula is:

[0040] in, This represents the total number of graph nodes in the image semantic graph. and The index number of the graph node, with values ​​ranging from 1 to... Integers; Indicates the first After the next node's state iteration update operation, the adjacency matrix of the next node... Line 1 The elements of the column can only be 0 or 1; Indicates the first After the next node's state iteration update operation, the adjacency matrix of the next node... Line 1 The elements of the column. Calculated. Afterwards, Compared with the preset change threshold Compare them.

[0041] The process of updating the graph topology of the image semantic graph based on the node state matrix of the current iteration round is executed by the graph topology reconstruction module in the first computation node. After a node state iteration update operation is completed, the updated node state vector of each graph node in the image semantic graph in the current iteration round is obtained, and the dimension of the updated node state vector is denoted as . The updated node state vectors of all graph nodes are arranged row by row according to the fixed index order of the graph nodes, forming the node state matrix for the current iteration. The node state matrix is ​​denoted as […]. , , represents the total number of graph nodes. Each row of the node state matrix corresponds to the updated node state vector of a graph node.

[0042] Calculate the covariance matrix of the node state matrix to obtain the global state correlation matrix of the image semantic graph. The calculation method is as follows: First, calculate the mean of each column of the node state matrix. Subtract the mean of each column from the mean of the node state matrix to obtain the mean-reduced node state matrix. Transpose the mean-reduced node state matrix and multiply it with the mean-reduced node state matrix. Then divide the result by... The covariance matrix is ​​obtained, denoted as The covariance matrix is ​​the global state correlation matrix. The Middle Line 1 The elements of a column represent graph nodes. With graph nodes Correlation coefficients along the feature dimension.

[0043] For each target graph node in the image semantic graph, extract the complete correlation coefficient vector corresponding to that target graph node from the global state correlation matrix. For each target graph node... Extract the first global state correlation matrix All elements in the row form the correlation coefficient vector. , The length is , No. Each element is a graph node. With graph nodes The correlation coefficient. Based on the correlation coefficient vector, select nodes in the target graph. The preset number of correlation coefficient values ​​ranked first The graph node, as the target graph node Candidate reconnection nodes. Preset number. The preset value is 5, based on maintaining appropriate sparsity in the graph topology and the neighborhood size required for semantic information transmission. When selecting candidate reconnection nodes, if the target graph node... It appears at the top of the sort. If a node is selected from all nodes in the graph, it is removed and replaced by the next sorted node to ensure that the candidate reconnection node does not include the target graph node itself.

[0044] Disconnect the target graph node in the image semantic graph. All original edge connections: For the first edge in the adjacency matrix row and number In this column, change all elements with a value of 1 to 0, which indicates the removal of the target graph node. Connect the target graph node to all existing edges of all other graph nodes. In the adjacency matrix, connect the target graph node... Establish a new edge connection with each candidate reconnection node, that is, the th edge in the adjacency matrix Set the column element corresponding to the candidate reconnection node index in the row to 1, and simultaneously set the first... The row elements corresponding to the candidate reconnected node indices are set to 1, thus adding undirected edges to the graph topology. After performing the above operations of disconnecting the original edges and establishing new edges on all graph nodes in the image semantic graph, the updated adjacency matrix is ​​obtained, which in turn yields the updated graph topology.

[0045] After updating the graph topology, the updated adjacency matrix is ​​used as the graph topology for the next node state iteration update operation. A maximum number of iterations is preset in the above process. Preset change threshold and preset quantity All parameters are constants pre-set before the image processing method is executed and are not updated during the training process. The learnable parameters involved in the image processing method, including the attention weight matrix, the weight matrix in the gating fusion operation, and the bias vector, are all learned and determined through the backpropagation algorithm during the entire image semantic segmentation task training phase. The training dataset consists of the original images and the corresponding pixel-level semantic category annotations, and the loss function is pixel-wise cross-entropy loss.

[0046] Example 4:

[0047] In specific implementation, please refer to Figure 5 The process of performing multi-level feature reconstruction operations on the features of each node in the semantic map of the target image is jointly executed by the feature reconstruction module and the upsampling decoder module deployed in the first computing node. The semantic map of the target image contains... There are graph nodes, and each graph node corresponds to an updated node state vector. The dimension of the updated node state vector is denoted as . , The specific value is determined by the gating fusion output dimension in the aforementioned graph node state iteration update operation, and can be set to 64. The feature reconstruction module maps the updated node state vector of each graph node in the target image semantic map back to the original pixel position in the adaptive spatial feature map. There is a pre-recorded one-to-one correspondence between the graph nodes in the target image semantic map and the pixel positions in the adaptive spatial feature map. This correspondence is established when constructing the image semantic map, that is, the coordinates in the adaptive spatial feature map are... The pixel position is mapped to the graph node index. The mapping operation specifically involves: for each original pixel position in the adaptive spatial feature map... The corresponding graph node index is retrieved, the updated node state vector of that graph node is obtained, and the updated node state vector is directly assigned to the original pixel position as the reconstructed feature vector. The dimension of the reconstructed feature vector is... After obtaining the reconstructed feature vectors of all original pixel locations, they are arranged according to the spatial order of the original pixel locations, i.e., row-major order, to obtain a spatial dimension of [missing value]. The reconstructed feature map, where and These represent the height and width of the adaptive spatial feature map, respectively.

[0048] In practice, the first-level upsampling operation is implemented through a transposed convolutional layer, which has a number of input channels of [number missing]. The number of output channels is denoted as , The value is set to 32. The transposed convolution kernel size is 4×4, the stride is 2, and the side padding is 1. After inputting the reconstructed feature map into this transposed convolutional layer, the spatial resolution is doubled according to the stride of the transposed convolution, resulting in the first-level upsampled feature map. The spatial size of the first-level upsampled feature map is... A second-level upsampling operation is performed on the first-level upsampled feature map, using bilinear interpolation. The target output spatial size of the bilinear interpolation is set to match the spatial size of the target image, i.e., the height is [value missing]. Width is , and These are the original height and width recorded when the target image is input. The second-level upsampled feature map is obtained through bilinear interpolation, and the spatial resolution of the second-level upsampled feature map is... The number of channels remains at .

[0049] In practice, since there may be edge pixel alignment deviations during the size enlargement process of the first-level upsampling and the second-level upsampling, the input image size information of the target image when it is input to the first computing node is obtained. This size information includes the original image height. and width Based on the input image size information, perform edge pixel cropping on the second-level upsampled feature map: if the height or width of the second-level upsampled feature map is equal to the input image size... , If pixel deviation exists, redundant edge pixel rows and columns are symmetrically cropped to ensure that the spatial resolution of the cropped feature map is exactly equal to the original spatial resolution of the target image. When the bilinear interpolation target size is strictly set to... In this case, edge pixel cropping can be skipped. The final edge-cropped feature map will be used as the input feature map for subsequent class probability mapping operations, and its size is [size missing]. .

[0050] In practice, the class probability mapping operation is performed using a pre-defined pixel-wise classifier. The pixel-wise classifier consists of a fully connected mapping layer and a normalized exponential function mapping layer. The fully connected mapping layer is implemented using a 1×1 convolutional layer with an input channel number of... The number of output channels is equal to the total number of preset image semantic categories. , The specific value depends on the number of categories defined in the image semantic segmentation task; for example, it could be set to 19 in urban scene semantic segmentation. This fully connected mapping layer is applied to the feature map after edge cropping. For each pixel position in the feature map after edge cropping... Its corresponding eigenvector is denoted as The fully connected mapping layer performs a linear mapping operation to obtain the initial category score vector corresponding to each preset image semantic category at the pixel location. The parameters of the linear mapping are a learnable weight matrix and a bias vector.

[0051] The normalized exponential function mapping layer performs probability normalization on the initial class score vector, obtaining the class probability value of each pixel location belonging to each preset image semantic category. For pixel location... , No. The formula for calculating the category probability value of each preset image semantic category is:

[0052] in, Indicates pixel position Belongs to the The category probability values ​​of each preset image semantic category, with a value range of (0,1); Indicates pixel position The initial category score vector corresponding to the first The score values ​​of each preset image semantic category; The index is a preset image semantic category, with values ​​ranging from 1 to... Integers; To iterate through all preset image semantic categories in the summation operation, the indices also range from 1 to... Integers; This represents the natural exponential function. The pixel position is calculated using the above formula. Corresponds to all After constructing a vector from the category probability values ​​of each category, the preset image semantic category corresponding to the highest probability value in the vector is selected as the pixel location. The semantic category label is generated by performing the above operation on all pixel locations. This generates an image semantic segmentation map with the same spatial size as the target image. The value of each pixel location in the image semantic segmentation map is the semantic category label for that pixel location.

[0053] In practical implementation, the pre-configured deep convolutional network in the first computing node is used for multi-level semantic feature extraction. This deep convolutional network includes multiple encoder modules, which are cascaded sequentially. Each encoder module has a different downsampling stride, with the first encoder module having the smallest downsampling stride among all encoder modules. The internal structure of the first encoder module consists of: a convolutional layer with a preset first value stride, a batch normalization layer, and an activation function layer. The preset first value is set to 2, based on the principle that this stride can preserve sufficient shallow texture details while reducing spatial resolution. After the target image is input into the deep convolutional network, the first encoder module performs a convolution operation with a stride of 2 on the target image. The convolution kernel size is 3×3, the number of input channels is the same as the number of color channels in the target image (3), and the number of output channels is denoted as [missing information]. , This can be set to 64. The convolution operation yields the first-level convolutional feature map. The batch normalization layer calculates the mean and variance for each channel of the first-level convolutional feature map and normalizes them. The activation function layer uses a modified linear unit function to apply a non-linear mapping element-wise to the normalized feature map, resulting in an initial shallow texture feature map. The spatial resolution of the initial shallow texture feature map is half the spatial resolution of the target image. Subsequent intermediate and higher-level convolutional modules further perform downsampling operations, making the spatial resolution of the intermediate layer structure feature map smaller than that of the initial shallow texture feature map.

[0054] In practical implementation, all learnable parameters of the deep convolutional network and the attention weight matrix are used. Weight matrix in gating fusion operation and and bias vector The kernel parameters of the transposed convolutional layer, the weight matrix of the fully connected mapping layer, and the bias vector are all determined through an end-to-end supervised training process. The training dataset consists of multiple target image samples and corresponding pixel-level semantic category ground truth maps. Training uses the mini-batch stochastic gradient descent algorithm, with the loss function being the sum of the cross-entropy losses at all pixel locations. The optimizer momentum parameter is set to 0.9, the initial learning rate is set to 0.01, and a multinomial decay strategy is employed.

[0055] Example 5:

[0056] In practice, a deep learning-based image processing system includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory uses a non-volatile storage medium to store the computer program instructions, configuration parameters, pre-trained weights, and intermediate calculation results required during the image processing method's operation. The processor uses a graphics processing unit (GPU) or neural network processor with parallel computing capabilities to execute the computer program to implement the deep learning-based image processing method.

[0057] When a computer program runs on a processor, it is organized into multiple functional modules. These modules include an image receiving module, a multi-level semantic feature extraction module, a dynamic receptive field adjustment module, an image semantic graph construction module, a node state iterative update module, a graph topology update module, and a multi-level feature reconstruction module. The connections between these modules are as follows: the output of the image receiving module is connected to the input of the multi-level semantic feature extraction module; the output of the multi-level semantic feature extraction module is connected to the input of the dynamic receptive field adjustment module; the output of the dynamic receptive field adjustment module is connected to the input of the image semantic graph construction module; the output of the image semantic graph construction module is connected to the input of the node state iterative update module; the output of the node state iterative update module is connected to the inputs of both the graph topology update module and the multi-level feature reconstruction module; the output of the graph topology update module is fed back to the node state iterative update module; and the output of the multi-level feature reconstruction module outputs the image semantic segmentation map.

[0058] In practical implementation, the processor receives the target image sent by the image storage node through the image receiving module. The image receiving module obtains the target image from the image storage node via a communication interface. The target image is stored as a tensor, with a tensor dimension of [missing information]. ,in Indicates the batch size, which is used when processing a single image. Setting it to 1 or 3 indicates the red, green, and blue color channels of the target image. Indicates the height of the target image. This indicates the width of the target image. The image receiving module transmits the target image to the multi-level semantic feature extraction module.

[0059] Multi-level semantic feature extraction is achieved through first-level convolutional mapping, second-level semantic abstraction, third-level global semantic encoding, and cross-level feature overlay and fusion. The multi-level semantic feature extraction module in the processor is configured with a deep convolutional network. The structure of the deep convolutional network is as follows: the first-level convolutional mapping part consists of a convolutional layer with a stride of 2, a batch normalization layer, and a modified linear unit activation layer connected sequentially. The convolutional kernel size is 3×3, with 3 input channels and 64 output channels. The second-level semantic abstraction part consists of three sequentially connected residual modules. Each residual module contains two 3×3 convolutional layers and an identity mapping branch. The first residual module has 6 input channels. 4. The output channel count is 128, and the downsampling stride is 2. The second residual module has 128 input channels, 256 output channels, and a downsampling stride of 2. The third residual module has 256 input channels, 512 output channels, and a downsampling stride of 2. The third-level global semantic encoding part consists of a dilated convolutional layer and a global average pooling layer. The dilated convolutional layer has an expansion rate of 2, 512 input channels, and 512 output channels. The global average pooling layer compresses the spatial dimension of the feature map to 1×1. Cross-level feature overlay and fusion are achieved through a feature concatenation layer and a 1×1 convolutional fusion layer. The feature concatenation layer upsamples the initial shallow texture feature map, intermediate structural feature map, and high-level semantic feature map to the same resolution and then concatenates them along the channel dimension. The 1×1 convolutional fusion layer has 64+256+512 input channels and 256 output channels. After the target image undergoes the processing described above, the multi-level semantic feature extraction module outputs a basic semantic feature map.

[0060] The dynamic receptive field adjustment operation obtains the receptive field scale adjustment coefficient through grid partitioning, spatial Euclidean distance calculation, and normalized mapping, and performs dilated convolution operations with different dilation rates accordingly. The dynamic receptive field adjustment module in the processor uniformly divides the basic semantic feature map space into 8×8 local spatial feature blocks. For each local spatial feature block, the spatial coordinates of its center pixel are determined, and the spatial Euclidean distances from the center coordinates to the center coordinates of the other local spatial feature blocks are calculated to form a spatial distance vector. After averaging all distance values ​​in the spatial distance vector, the receptive field scale adjustment coefficient is obtained through a preset mapping. The mapping method is: the receptive field scale adjustment coefficient is equal to the natural constant... The base is the average distance value divided by the scale adjustment parameter. The opposite of the exponent is the result of exponentiation. The value is fixed at 1.0. Based on the receptive field scale adjustment coefficient, a dilated convolution operation is performed on each local spatial feature block, with the dilation rate positively correlated with the receptive field scale adjustment coefficient. The dilated convolution operation for each local spatial feature block is performed by independent 3×3 dilated convolution kernels. The number of input channels and output channels are equal to the number of channels in the basic semantic feature map. All updated local spatial feature blocks are concatenated according to their original spatial positions, and the dynamic receptive field adjustment module outputs an adaptive spatial feature map.

[0061] Image semantic graph construction is achieved by determining edge connectivity through pixel-level feature vector extraction, feature similarity calculation, and threshold comparison. The image semantic graph construction module in the processor uses a 1×1 convolutional layer to perform pixel-level feature vector extraction on the adaptive spatial feature map. The number of input channels of the 1×1 convolutional layer is the same as the number of channels in the adaptive spatial feature map, and the number of output channels is... Set to 64 to generate the value for each pixel position. 3D pixel feature vectors. Each pixel position is treated as a graph node, and the pixel feature vector serves as the initial node feature. The feature similarity between graph nodes is calculated using cosine similarity, which is the inner product of two initial node feature vectors divided by the product of their magnitudes. A preset similarity threshold is then applied. When the feature similarity between two graph nodes is set to 0.6, an initial edge connection is established between these two graph nodes. The image semantic graph construction module outputs an image semantic graph containing the adjacency matrix and the initial node features.

[0062] The node state iterative update operation is implemented through attention interaction weight calculation, weighted aggregation, and gated fusion. For each target graph node in the image semantic graph, the node state iterative update module in the processor obtains all neighboring graph nodes from the current adjacency matrix. It inputs the current node state vector of the target graph node and the current node state vector of each neighboring graph node into an attention weight calculation function implemented by a learnable bilinear mapping to obtain the attention interaction weight of each neighboring graph node to the target graph node. After softmax normalization, the attention interaction weights are used to perform a weighted summation on the current node state vectors of all neighboring graph nodes to obtain the neighborhood aggregated state vector. The gated fusion operation performs linear transformations on the current node state vector of the target graph node and the neighborhood aggregated state vector through the learnable weight matrix, adds them together, applies a bias, and then uses a sigmoid activation function to generate a gated weight vector. Finally, the updated node state vector is obtained by adding 1 to the element-wise product of the gated weight vector and the current node state vector, plus 1 to the element-wise product of the gated weight vector and the neighborhood aggregated state vector. The node state iteration update module performs the above operations on all graph nodes to complete one node state iteration update.

[0063] The graph topology update operation is implemented through the calculation of the covariance matrix of the node state matrix, the extraction of correlation coefficients, and the selection and reconnection of candidate reconnection nodes. After each node state iteration update operation, the graph topology update module in the processor arranges the updated node state vectors of all graph nodes in the order of the graph nodes to form a node state matrix. The number of rows in the node state matrix is ​​equal to the total number of graph nodes. The number of columns is equal to the dimension of the node state vector. After calculating the mean of each column of the node state matrix and removing the mean, calculate the product of the mean-removed node state matrix and its transpose, and divide by . ,get The covariance matrix of dimension 1 is used as the global state correlation matrix. For each target graph node, the correlation coefficient vector corresponding to that target graph node is extracted from the global state correlation matrix, and a predetermined number of nodes with the highest correlation coefficient values ​​are selected. Each graph node is considered a candidate reconnection node. The value is fixed at 5, based on balancing the sparsity of the graph topology and the sufficiency of semantic information transmission. All existing edge connections to the target graph node are disconnected in the adjacency matrix, and new edge connections are established with candidate reconnection nodes. The graph topology update module feeds back the updated graph topology to the node state iteration update module for use in the next iteration. The alternating execution of node state iteration updates and graph topology updates continues until a preset iteration stop condition is met. The preset iteration stop condition is that the number of iterations reaches a preset maximum of 10, or the change in graph topology is less than a preset change threshold of 0.01. The change in graph topology is obtained by calculating the sum of the absolute values ​​of the differences between corresponding elements in the two adjacency matrices and dividing by the square of the total number of graph nodes.

[0064] Multi-level feature reconstruction is implemented through node state vector mapping, upsampling, edge pixel cropping, and a pixel-by-pixel classifier. The multi-level feature reconstruction module in the processor first assigns the updated node state vector of each graph node in the target image semantic map to the corresponding pixel position based on the correspondence between graph nodes and the pixel positions of the adaptive spatial feature map, thus obtaining the reconstructed feature map. The upsampling operation is divided into two levels: the first-level upsampling uses transposed convolution with a kernel size of 4×4, a stride of 2, padding of 1, and an input channel number of... The output channel count is 32, which doubles the spatial resolution of the reconstructed feature map; the second-stage upsampling uses bilinear interpolation to directly adjust the feature map resolution to the original height of the target image. and width The edge pixel cropping operation is performed when the spatial size of the second-level upsampled feature map deviates from the original size of the target image due to the rounding operation. It symmetrically crops redundant edge pixel rows and columns, making the feature map resolution exactly equal to that of the target image. The pixel-wise classifier consists of a 1×1 convolutional layer as a fully connected mapping layer and a softmax layer as a normalized exponential function mapping layer, connected sequentially. The fully connected mapping layer has 32 input channels and output channels equal to the preset total number of image semantic categories. , The value is the number of categories set according to the specific application scenario. The softmax layer performs exponential normalization on the score vector of each pixel position. Belongs to the The probability value of each category is calculated using the following formula:

[0065] in, Indicates pixel position Belongs to the The category probability values ​​of a preset image semantic category, This is the pixel row coordinate index, with values ​​ranging from 1 to... , This is the pixel column coordinate index, with values ​​ranging from 1 to... ; Pixel positions in the output of a fully connected mapping layer Corresponding to the The score values ​​for each category; This is an image semantic category index, with values ​​ranging from 1 to... Integers; The index for traversing all image semantic categories in the summation operation also takes values ​​from 1 to... An integer. Select pixel position. The category corresponding to the maximum probability value among all categories is used as the semantic category label. After performing this operation on all pixel positions in the entire image, the multi-level feature reconstruction module outputs the semantic segmentation map of the image.

[0066] See Figure 7 In the graph, the horizontal axis represents the number of training iterations, ranging from 0 to 500; the left vertical axis represents the training cross-entropy loss, ranging from approximately 0.4 to 2.5; and the right vertical axis represents the learning rate, ranging from approximately 0 to 0.01. The solid curve in the graph shows the trend of training cross-entropy loss as the number of training iterations changes, and the dashed curve shows the trend of learning rate as the number of training iterations changes.

[0067] As shown in the figure, in the early stages of training (approximately 0 to 50 iterations), the training cross-entropy loss decreases rapidly from approximately 2.5 to approximately 1.0, indicating that the deep convolutional network and subsequent graph structure node state iterative update module described in this embodiment have initially learned effective image semantic segmentation features. With increasing iteration count, the rate of decrease in training cross-entropy loss gradually slows, stabilizing within the range of approximately 150 to 500 iterations, fluctuating between 0.5 and 0.7, indicating that the model gradually converges and the semantic segmentation performance steadily improves.

[0068] The learning rate curve exhibits a monotonically decreasing trend, gradually decreasing from an initial value of 0.01 to near 0, which aligns with the polynomial decay learning rate strategy employed in this embodiment. As the learning rate decreases, the fluctuation amplitude of the training cross-entropy loss curve diminishes, contributing to the model's stable optimization towards the end of convergence.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based image processing method, characterized in that, The method involves applying a first computing node to an image processing system, the image processing system comprising multiple computing nodes and image storage nodes, wherein the first computing node is any one of the computing nodes. The system receives the target image sent by the image storage node and performs multi-level semantic feature extraction on the target image to obtain the basic semantic feature map of the target image. Perform a spatially perceptual dynamic receptive field adjustment operation on the basic semantic feature map to obtain an adaptive spatial feature map; An image semantic map of the target image is constructed based on the adaptive spatial feature map, and the image semantic map contains the global contextual relationships between various image regions in the target image; Perform multiple iteration update operations on the node state of the graph structure on the image semantic graph. After each iteration update operation is completed, update the graph topology of the image semantic graph according to the node state matrix of the current iteration round until the preset iteration stop condition is reached to obtain the target image semantic graph. A multi-level feature reconstruction operation is performed on each node feature in the semantic map of the target image to generate an image semantic segmentation map corresponding to the target image.

2. The image processing method based on deep learning according to claim 1, characterized in that, Perform multi-level semantic feature extraction operations on the target image to obtain the basic semantic feature map of the target image, including: The first-level convolutional mapping operation is performed on the target image by calling the pre-configured deep convolutional network in the first computing node to obtain the initial shallow texture feature map of the target image; The initial shallow texture feature map is input into the intermediate convolutional module of the deep convolutional network to perform the second-level semantic abstraction operation, thereby obtaining the intermediate layer structure feature map of the target image; The intermediate layer structure feature map is input into the high-level convolutional module of the deep convolutional network to perform the third-level global semantic encoding operation, thereby obtaining the high-level semantic feature map of the target image. A cross-level feature overlay and fusion operation is performed on the initial shallow texture feature map, the intermediate layer structure feature map, and the high-level semantic feature map to obtain the basic semantic feature map.

3. The image processing method based on deep learning according to claim 2, characterized in that, Performing a spatially aware dynamic receptive field adjustment operation on the basic semantic feature map to obtain an adaptive spatial feature map includes: Perform a spatial dimension grid partitioning operation on the basic semantic feature map to obtain multiple local spatial feature blocks; For each local spatial feature block, calculate the spatial Euclidean distance between the center position of the local spatial feature block and the center positions of the other local spatial feature blocks in the basic semantic feature map to obtain the spatial distance vector corresponding to the local spatial feature block; A normalization mapping operation is performed on the spatial distance vector to obtain the receptive field scale adjustment coefficient corresponding to the local spatial feature block. The receptive field scale adjustment coefficient is used to characterize the effective neighborhood range of the local spatial feature block in the subsequent feature extraction process. Based on the receptive field scale adjustment coefficient corresponding to each local spatial feature block, dilated convolution operations with different dilation rates are performed on each local spatial feature block to obtain multiple updated local spatial feature blocks; The multiple updated local spatial feature blocks are concatenated according to their original spatial positions in the basic semantic feature map to obtain the adaptive spatial feature map.

4. The image processing method based on deep learning according to claim 3, characterized in that, In the dilated convolution operations with different dilation rates, the dilation rate corresponding to each local spatial feature block is positively correlated with the receptive field scale adjustment coefficient corresponding to that local spatial feature block.

5. The image processing method based on deep learning according to claim 3, characterized in that, Constructing an image semantic map of the target image based on the adaptive spatial feature map includes: Perform pixel-level feature vector extraction on the adaptive spatial feature map to obtain the pixel feature vector corresponding to each pixel position in the adaptive spatial feature map; Each pixel position is used as a graph node in the image semantic graph, and the pixel feature vector corresponding to each pixel position is used as the initial node feature of the graph node. Calculate the feature similarity between the initial node features of any two graph nodes in the image semantic graph; The connection relationship between two graph nodes whose feature similarity is greater than a preset similarity threshold is determined as the initial edge connection relationship of the image semantic graph; Based on the initial edge connection relationships and the initial node features of each of the graph nodes, the image semantic graph of the target image is constructed.

6. The image processing method based on deep learning according to claim 5, characterized in that, Perform multiple iterative update operations on the node states of the graph structure on the image semantic graph, including: In each node state iteration update operation, for each target graph node in the image semantic graph, all adjacent graph nodes of the target graph node are obtained from the image semantic graph. The all adjacent graph nodes are all graph nodes in the image semantic graph that have edge connection relationships with the target graph node. The current node state vector of the target graph node and the current node state vector of each adjacent graph node are respectively input into the attention weight calculation function to obtain the attention interaction weight of each adjacent graph node relative to the target graph node. A weighted aggregation operation is performed on the current node state vector of all the adjacent graph nodes according to the attention interaction weight of each of the adjacent graph nodes to obtain the neighborhood aggregated state vector of the target graph node. Perform a gating fusion operation on the current node state vector of the target graph node and the neighborhood aggregated state vector to obtain the updated node state vector of the target graph node after the node state iteration update operation is completed.

7. The image processing method based on deep learning according to claim 6, characterized in that, The gating fusion operation uses a learnable gating weight vector to perform a weighted summation of the current node state vector and the neighborhood aggregated state vector of the target graph node. The gating weight vector is generated by nonlinear mapping of the current node state vector and the neighborhood aggregated state vector.

8. The image processing method based on deep learning according to claim 1, characterized in that, The preset iteration stop condition is that the current iteration round reaches the preset maximum number of iterations or the graph topology change of the image semantic graph is less than the preset change threshold.

9. The image processing method based on deep learning according to claim 6, characterized in that, The step of updating the graph topology of the image semantic graph based on the node state matrix of the current iteration round includes: Obtain the updated node state vector of each graph node in the image semantic graph in the current iteration round, and arrange the updated node state vectors of all graph nodes in the order of graph nodes to obtain the node state matrix of the current iteration round. Calculate the covariance matrix of the node state matrix to obtain the global state correlation matrix of the image semantic graph; For each target graph node in the image semantic graph, extract the correlation coefficient vector between the target graph node and all other graph nodes from the global state correlation matrix; Based on the correlation coefficient vector, a predetermined number of graph nodes with the highest correlation coefficients to the target graph node are selected as candidate reconnection nodes for the target graph node. Disconnect all original edge connections of the target graph node in the image semantic graph, and establish new edge connections between the target graph node and each of the candidate reconnection nodes to obtain the updated graph topology.

10. A deep learning-based image processing system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based image processing method according to any one of claims 1 to 9.

Images