Deep learning driven image segmentation system

Abstract

The present application relates to the technical field of image segmentation, in particular to a deep learning driven image segmentation system, the system comprising an image analysis module, a feature encoding module, an incremental learning module, a segmentation strategy adjustment module, a refined segmentation execution module, a multi-scale fusion module, an edge optimization module, and a result calibration module.The present application, through a difference analysis submodule, can dynamically adjust based on the degree of difference between the feature vector set and the existing data set, and when the degree of difference exceeds the preset threshold, by marking as high difference and increasing the input proportion of new data, the system can quickly adapt to new or changing image features, thereby reducing the precision decline caused by outdated models.Through the effective cooperation of the learning strategy adjustment submodule and the model parameter update submodule, the system not only maintains the memory of old knowledge, but also quickly integrates new knowledge, ensuring the continuous optimization and updating of the model.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, and more particularly to a deep learning-driven image segmentation system. Background Technology

[0002] Image segmentation technology focuses on subdividing digital images into multiple parts or regions to simplify or alter their representation, making them easier to analyze. Image segmentation plays a crucial role in many applications, such as medical imaging, autonomous vehicles, robotic perception, and video surveillance. The core of the technology lies in identifying boundaries, objects, shapes, and specific regions within an image to facilitate further processing or analysis.

[0003] Deep learning-driven image segmentation systems utilize deep learning technology to automatically identify and segment different parts of an image. The system aims to learn from large amounts of labeled image data to automatically identify different objects and shapes within an image, achieving high accuracy and efficiency in image segmentation. This technology enables accurate segmentation of images in complex scenes without human intervention, significantly improving the automation level of image processing.

[0004] Compared to traditional image segmentation systems, conventional techniques suffer from insufficient adaptability and low segmentation accuracy when handling new data or complex scenes. Traditional systems lack effective mechanisms to balance learning from both new and old data, making them prone to forgetting previous knowledge when learning new tasks, leading to decreased model performance. They are also less refined in detail processing and edge optimization, making it difficult to handle varied image features and complex segmentation tasks. This results in significantly reduced practicality and reliability of segmentation results in demanding applications such as medical imaging and autonomous driving. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-driven image segmentation system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-driven image segmentation system, the system comprising an image parsing module, a feature encoding module, an incremental learning module, a segmentation strategy adjustment module, a fine segmentation execution module, a multi-scale fusion module, an edge optimization module, and a result calibration module;

[0007] The image parsing module parses the input image data, identifies the basic constituent elements in the image, including color blocks, shape outlines and texture patterns, and generates an element mapping table.

[0008] The feature encoding module encodes the element mapping table, assigns encoding weights based on the frequency and spatial distribution of elements, and creates a feature set to generate a set of feature vectors.

[0009] The incremental learning module analyzes the difference between the feature vector set and the existing dataset, adjusts the data input ratio according to the size of the difference, adopts an incremental learning strategy, dynamically modifies the weight allocation, updates the model parameters, and generates an updated model parameter set.

[0010] The segmentation strategy adjustment module evaluates the adaptability of the differentiated segmentation strategy to the current dataset based on the updated model parameter set, selects the strategy with the best matching degree, adjusts the segmentation parameters, and generates an optimized parameter configuration.

[0011] The refined segmentation execution module applies optimized parameter configuration to segment the image, uses judgment logic to process blurred or overlapping regions in the segmentation, adjusts the segmentation range, and generates a refined set of segmented images.

[0012] The multi-scale fusion module fuses the refined segmented image set at different resolutions, adjusts the fusion ratio and strategy, and optimizes the detail display and image coherence to generate a multi-scale fused image.

[0013] The edge optimization module analyzes and corrects the edges of the multi-scale fused image, adjusts the brightness and color of edge pixels, matches the surrounding area, optimizes edge blur and breakage, and generates an edge-optimized image.

[0014] The result calibration module verifies the edge-optimized image, checks the segmentation accuracy and detail performance, adjusts the output resolution and format, determines that it meets predetermined standards, and generates an image segmentation result set.

[0015] The present invention improves upon this invention by including the following: the element mapping table includes color distribution regions, key shape boundaries, and texture feature identifiers; the feature vector set includes color block weight distribution, shape contour encoding, and texture pattern frequency; the updated model parameter set includes adjusted inter-layer connection weights, neuron bias values, and activation function parameters; the optimized parameter configuration includes segmentation optimization threshold, region expansion control factor, and shape correlation evaluation criterion; and the image segmentation result set includes a segmentation effect comparison analysis chart and a detail restoration verification chart.

[0016] The present invention is improved in that the image parsing module includes a color analysis submodule, a shape parsing submodule, and a texture recognition submodule;

[0017] The color analysis submodule performs color spectrum analysis on the pixels in the image based on the input image data, counts the occurrence frequency and distribution location of each color, and determines the key color regions in the image to obtain color distribution data.

[0018] The shape parsing submodule identifies the basic shape contours in the image based on the color distribution data, divides the image into multiple shape feature regions, and obtains a shape contour dataset.

[0019] The texture recognition submodule analyzes the texture features of each shape region based on the shape contour dataset, including the identification of texture type, direction and repetition pattern, distinguishing feature regions of different textures in the image, and establishing an element mapping table.

[0020] The present invention is improved in that the feature encoding module includes a weight allocation submodule, an encoding conversion submodule, and a set creation submodule;

[0021] The weight allocation submodule performs key evaluations on various color, shape, and texture features based on the element mapping table, and assigns corresponding weight values ​​according to the visual impact and recognition difficulty of the features in the image, thereby generating weight data.

[0022] The encoding conversion submodule applies weight data to features, converts features into unique codes for representation, including color, shape and texture features, and assigns corresponding encoding segments to obtain encoded feature data;

[0023] The set creation submodule aggregates the encoded feature data and verifies the color, shape, and texture features of the images in the set, providing a data foundation for image processing and analysis tasks and generating a set of feature vectors.

[0024] The present invention is improved in that the incremental learning module includes a difference analysis submodule, a learning strategy adjustment submodule, and a model parameter update submodule;

[0025] The difference analysis submodule is based on the feature vector set and uses flexible weight sharing to compare the existing dataset, calculate the difference between the feature vector and the existing model. If the difference exceeds a preset threshold, it is marked as high difference; otherwise, it is marked as low difference, and a difference analysis record is generated.

[0026] The learning strategy adjustment submodule uses a dynamic data buffering mechanism based on the difference analysis record. If the difference is high, the proportion of new data input is increased, and the learning rate and weight allocation are dynamically adjusted. If the difference is low, the original learning rate is maintained, and the weight allocation is adjusted to obtain learning strategy adjustment data.

[0027] The model parameter update submodule adjusts the data based on the learning strategy, updates the model's weights, bias terms, and correlation parameters, matches new data features, continuously optimizes the model, and generates an updated model parameter set.

[0028] The present invention is improved in that the segmentation strategy adjustment module includes an adaptive evaluation submodule, a strategy selection submodule, and a parameter configuration submodule;

[0029] The adaptive evaluation submodule analyzes the compatibility of the current dataset features with multiple segmentation strategies based on the updated model parameter set, calculates the adaptive score for each strategy, and generates adaptive score data.

[0030] The strategy selection submodule compares the scores of multiple segmentation strategies based on the adaptive score data, selects the strategy with the best score as the key segmentation strategy, and determines that the selected strategy conforms to the characteristics of the current dataset, thus obtaining the selected optimal segmentation strategy.

[0031] The parameter configuration submodule adjusts the parameters required for the segmentation operation, including segmentation depth and edge detection sensitivity, based on the selected optimal segmentation strategy, to optimize the effect of the segmentation strategy and match new data features and model parameters to obtain an optimized parameter configuration.

[0032] The present invention is improved in that the refined segmentation execution module includes a preliminary segmentation submodule, a region processing submodule, and a range adjustment submodule;

[0033] The preliminary segmentation submodule performs preliminary segmentation processing on the image based on optimized parameter configuration, identifies key objects and background in the image, and marks areas with unclear boundaries as requiring processing to generate a preliminary segmentation image.

[0034] The region processing submodule, based on the preliminary segmented image, analyzes the marked blurred or overlapping regions that need to be processed using the backpropagation algorithm, and optimizes the regions and boundaries to obtain the processed region image.

[0035] The range adjustment submodule evaluates the segmentation effect of the image based on the processed region image. If the segmentation range is inaccurate or there are omissions, the Actor-Critic algorithm is used to dynamically adjust the range to cover all related objects and generate a refined segmented image set.

[0036] The present invention is improved in that the multi-scale fusion module includes a resolution analysis submodule, a fusion strategy determination submodule, and a detail and coherence optimization submodule;

[0037] The resolution analysis submodule examines each image one by one based on a refined segmented image set, analyzes the resolution characteristics of the image, identifies the resolution level of each image, and records its features to provide a basis for the fusion strategy and generate resolution feature data.

[0038] The fusion strategy determination submodule calculates the optimal ratio parameters for image fusion based on resolution feature data by analyzing the pixel size ratio and distribution differences between resolution images, and obtains fusion strategy parameters based on detail preservation and visual coherence between images.

[0039] The detail and coherence optimization submodule applies fusion strategy parameters to dynamically adjust the image, optimizes areas with lost details or poor visual coherence, adjusts the target area, and optimizes the quality and visual effect of the fused image to obtain a multi-scale fused image.

[0040] The present invention is improved in that the edge optimization module includes an edge detection submodule, a brightness and color adjustment submodule, and an edge smoothing submodule;

[0041] The edge detection submodule performs pixel contrast analysis on the multi-scale fused image, identifies edge regions that differ from the contrast of surrounding pixels, and marks the regions based on blurred or discontinuous boundary lines to generate edge detection records.

[0042] The brightness and color adjustment submodule analyzes the marked edge regions based on edge detection records, adjusts the brightness and color values ​​to match the surrounding areas, and determines the continuity and consistency between the edges and the surrounding environment through step-by-step fine-tuning, thereby reducing visual breaks and obtaining the adjusted image.

[0043] The edge smoothing submodule smooths the marked edges in the adjusted image, gradually reducing jaggedness, and adjusts the pixel gradients near the edges to optimize the smooth transition from the edges to the surrounding areas, resulting in an edge-optimized image.

[0044] The present invention is improved in that the result calibration module includes an accuracy verification submodule, a format adjustment submodule, and a standard confirmation submodule;

[0045] The accuracy verification submodule performs a step-by-step comparative analysis based on the edge-optimized image, compares the segmentation results with the original image, identifies the accuracy of the segmentation lines and the preservation of details, and verifies the consistency between the natural transition after edge processing and the original image, generating accuracy verification data.

[0046] The format adjustment submodule adjusts the image output resolution according to the accuracy test data to match the target output requirements. At the same time, it converts the image format to match the compatibility of the target platform or storage standard and formats the image to obtain the format-adjusted image.

[0047] The standard verification submodule performs quality checks on the format-adjusted images to determine whether each image meets the predetermined segmentation accuracy, detail representation, and format standards, including resolution, color depth, and file type, and establishes an image segmentation result set.

[0048] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0049] In this invention, through the difference analysis submodule, the system can dynamically adjust based on the difference between the feature vector set and the existing dataset. When the difference exceeds a preset threshold, it is marked as high-discrepancy and the proportion of new data input is increased, enabling the system to quickly adapt to new or changing image features, thereby reducing the accuracy drop caused by model obsolescence. Through the effective cooperation of the learning strategy adjustment submodule and the model parameter update submodule, the system not only maintains its memory of old knowledge but also quickly integrates new knowledge, ensuring continuous model optimization and updates. The backpropagation algorithm and Actor-Critic algorithm used in the fine-grained segmentation execution module further optimize the accuracy and detail processing of the segmentation range, making the image segmentation results more refined and accurate. Attached Figure Description

[0050] Figure 1 This invention presents a block diagram of a deep learning-driven image segmentation system.

[0051] Figure 2 The present invention presents a system framework diagram of a deep learning-driven image segmentation system.

[0052] Figure 3 This invention provides a flowchart of the image parsing module in a deep learning-driven image segmentation system.

[0053] Figure 4 The present invention provides a flowchart of the feature encoding module in a deep learning-driven image segmentation system;

[0054] Figure 5 The present invention provides a flowchart of the incremental learning module in a deep learning-driven image segmentation system.

[0055] Figure 6 The present invention provides a flowchart of the segmentation strategy adjustment module in a deep learning-driven image segmentation system.

[0056] Figure 7The present invention provides a flowchart of the fine segmentation execution module in a deep learning-driven image segmentation system.

[0057] Figure 8 The present invention provides a flowchart of the multi-scale fusion module in a deep learning-driven image segmentation system.

[0058] Figure 9 The present invention provides a flowchart of the edge optimization module in a deep learning-driven image segmentation system.

[0059] Figure 10 The flowchart of the result calibration module in the deep learning-driven image segmentation system proposed in this invention is shown. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] In the description of this invention, it should be understood that the terms length, width, up and down, front and back, left and right, vertical and horizontal, top and bottom, inside and outside, etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, in the description of this invention, "multiple" means two or more, unless otherwise explicitly specified.

[0062] Example

[0063] Please see Figure 1 The present invention provides a technical solution: a deep learning-driven image segmentation system includes an image parsing module, a feature encoding module, an incremental learning module, a segmentation strategy adjustment module, a fine segmentation execution module, a multi-scale fusion module, an edge optimization module, and a result calibration module;

[0064] The image parsing module parses the input image data, identifies the basic constituent elements in the image, including color blocks, shape outlines and texture patterns, and generates an element mapping table.

[0065] The feature encoding module encodes the element mapping table, assigns encoding weights based on the frequency and spatial distribution of elements, and creates a feature set to generate a set of feature vectors.

[0066] The incremental learning module analyzes the differences between the feature vector set and the existing dataset, adjusts the data input ratio according to the magnitude of the differences, adopts an incremental learning strategy, dynamically modifies the weight allocation, updates the model parameters, and generates an updated model parameter set.

[0067] The segmentation strategy adjustment module evaluates the adaptability of the differentiated segmentation strategy to the current dataset based on the updated model parameter set, selects the strategy with the best matching degree, adjusts the segmentation parameters, and generates an optimized parameter configuration.

[0068] The fine segmentation execution module applies optimized parameter configurations to segment the image, uses judgment logic to process blurred or overlapping regions in the segmentation, adjusts the segmentation range, and generates a refined set of segmented images.

[0069] The multi-scale fusion module fuses the refined segmented image set at different resolutions, adjusts the fusion ratio and strategy, and optimizes the detail display and image coherence to generate a multi-scale fused image.

[0070] The edge optimization module analyzes and corrects the edges of the multi-scale fused image, adjusts the brightness and color of edge pixels, matches the surrounding area, optimizes edge blur and breakage, and generates an edge-optimized image.

[0071] The result calibration module verifies the edge-optimized image, checks the segmentation accuracy and detail performance, adjusts the output resolution and format, and determines that it meets the predetermined standards, generating an image segmentation result set.

[0072] The element mapping table includes color distribution regions, key shape boundaries, and texture feature identifiers. The feature vector set includes color block weight distribution, shape contour encoding, and texture pattern frequency. The updated model parameter set includes adjusted inter-layer connection weights, neuron bias values, and activation function parameters. The optimized parameter configuration includes segmentation optimization threshold, region expansion control factor, and shape correlation evaluation criteria. The image segmentation result set includes segmentation effect comparison analysis chart and detail restoration verification chart.

[0073] Please see Figure 2 and Figure 3 The image parsing module includes a color analysis submodule, a shape parsing submodule, and a texture recognition submodule;

[0074] The color analysis submodule performs color spectrum analysis on the pixels in the image based on the input image data, counts the occurrence frequency and distribution location of each color, and identifies the key color regions in the image to obtain color distribution data.

[0075] The shape parsing submodule identifies the basic shape contours in the image based on the color distribution data, divides the image into multiple shape feature regions, and obtains the shape contour dataset.

[0076] The texture recognition submodule analyzes the texture features of each shape region based on the shape contour dataset, including the identification of texture type, direction and repetition pattern, distinguishing feature regions with different textures in the image, and establishing an element mapping table.

[0077] In the color analysis submodule, based on the input image data, the histogram equalization method is used to optimize the image brightness distribution. Color features are extracted through HSV (hue, saturation, brightness) color space conversion. The pixels of each color are counted and their spatial positions in the image are analyzed. The color histogram is calculated using the calcHist function in Python's OpenCV library to determine the dominant colors in the image and their distribution patterns. This process not only counts color frequencies but also considers the distribution of colors in the image in order to more accurately capture key color regions in the image and obtain color distribution data.

[0078] In the shape parsing submodule, edge detection algorithms, such as the Canny algorithm, are used to determine the edge information in the image based on the color distribution data. Furthermore, the shape contours in the image are extracted using the findContours function of the OpenCV library through contour tracking technology. The extracted contours are analyzed to identify and distinguish different shape feature regions, including contour approximation and shape classification, so as to subdivide the image into regions containing specific shapes, thereby providing a foundation for subsequent texture analysis and obtaining a shape contour dataset.

[0079] In the texture recognition submodule, based on the shape contour dataset, the texture features of each shape region are analyzed in detail. The Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) algorithms are used to calculate the statistical characteristics and pattern distribution of the texture, respectively. GLCM reveals the spatial relationship between pixels and the roughness of the texture, while LBP identifies the micro-texture patterns in the image. By combining the analysis results of these two methods, regions with different texture features in the image are distinguished and identified. The algorithm is implemented using Python and OpenCV libraries, and the element mapping table is further constructed and improved. This table records in detail the texture type, direction and repetition pattern of each region, thus establishing the element mapping table.

[0080] Please see Figure 2 and Figure 4 The feature encoding module includes a weight allocation submodule, an encoding conversion submodule, and a set creation submodule;

[0081] The weight allocation submodule evaluates various color, shape and texture features based on the element mapping table, assigns corresponding weight values ​​according to the visual impact and recognition difficulty of the features in the image, and generates weight data.

[0082] The encoding transformation submodule applies weighted data to features, transforms features into unique codes for representation, including color, shape and texture features, and assigns corresponding encoding segments to obtain encoded feature data;

[0083] The collection creation submodule aggregates coded feature data and verifies the color, shape, and texture features of the images in the collection, providing a data foundation for image processing and analysis tasks and generating a set of feature vectors.

[0084] In the weight allocation submodule, feature importance analysis is performed based on the element mapping table. This process analyzes the contribution of color, shape, and texture features to image recognition and uses a decision tree algorithm to determine the information gain value of each feature. Features with high information gain are considered more critical to image recognition and are assigned higher weight values ​​accordingly. In particular, this module uses Python's Scikit-learn library to implement the decision tree algorithm. The decision tree classifier calculates and ranks the information gain of all features, and then assigns weights to each feature according to this ranking. Features with higher weights will receive higher priority in subsequent encoding and transformation processes, generating weighted data.

[0085] In the encoding conversion submodule, weighted data is received, and each feature is converted into a unique encoded representation. This process encodes color, shape, and texture features using a hash function, ensuring that each feature is assigned to a unique encoding segment for fast retrieval and recognition. Specifically, the MD5 hash function is used to hash the weight value and feature identifier of each feature, generating a fixed-length encoded string. This operation is performed using Python's built-in hashlib library, using the md5() function to calculate the hash value. Each feature is assigned to a different encoding segment according to its importance. This not only optimizes the storage efficiency of features but also facilitates rapid feature matching and recognition, obtaining encoded feature data.

[0086] In the collection creation submodule, the encoded feature data is aggregated. This step integrates the encoded feature data into a unified feature vector space using vectorization methods to support efficient image processing and analysis tasks. Specifically, the NumPy library is used to construct a feature vector matrix, with each row representing the encoded feature set of an image. The feature matrix is ​​created using the numpy.array() function. In addition, the consistency and integrity of the generated feature vector set are verified to ensure that the color, shape, and texture features of each image are accurately represented and recorded, laying a solid data foundation for subsequent image analysis tasks and generating the feature vector set.

[0087] Please see Figure 2 and Figure 5The incremental learning module includes a differential analysis submodule, a learning strategy adjustment submodule, and a model parameter update submodule.

[0088] The difference analysis submodule is based on the feature vector set and uses flexible weight sharing to compare the existing dataset, calculate the difference between the feature vector and the existing model. If the difference exceeds the preset threshold, it is marked as high difference; otherwise, it is marked as low difference, and a difference analysis record is generated.

[0089] The learning strategy adjustment submodule uses a dynamic data buffering mechanism based on the difference analysis records. If the difference is high, the proportion of new data input is increased, and the learning rate and weight allocation are dynamically adjusted. If the difference is low, the original learning rate is maintained, and the weight allocation is adjusted to obtain learning strategy adjustment data.

[0090] The model parameter update submodule adjusts the data based on the learning strategy, updates the model's weights, bias terms, and associated parameters, matches new data features, continuously optimizes the model, and generates an updated model parameter set.

[0091] In the difference analysis submodule, based on the feature vector set, elastic weight sharing (EWC) is used to compare with the existing dataset to identify the impact of new information on the model. By calculating the Fisher information matrix between the feature vectors and the existing model parameters, this matrix is ​​used to determine which parameters are crucial to the previously learned task and are protected from the impact of the new learning task update. The importance of parameter changes is evaluated by setting a threshold. If the change exceeds the threshold, it is marked as high difference. This threshold is determined in advance through experiments on the validation set to balance the preservation and updating of new and old knowledge, and to generate difference analysis records.

[0092] In the learning strategy adjustment submodule, the learning strategy is adjusted using a dynamic data buffering mechanism based on the difference analysis records. If the records show high difference, it indicates that the new data has introduced significant new knowledge. In this case, the proportion of new data in the training batch is increased, and the learning rate and weight allocation are dynamically adjusted to promote the rapid learning of new knowledge. During this process, the adjustment of the learning rate is based on a preset learning rate adjustment strategy, such as learning rate decay. The weight allocation is adjusted according to the EWC results to ensure that the model does not lose memory of old tasks when learning new tasks. For low difference cases, the original learning rate is maintained, and only the weight allocation is fine-tuned to optimize the integration of existing knowledge and obtain learning strategy adjustment data.

[0093] In the model parameter update submodule, data is adjusted based on the learning strategy to update the model weights, bias terms, and related parameters, ensuring that the model can accurately match the new data features and continuously optimize. This is implemented through the backpropagation algorithm. The update amount of the weights and bias terms depends not only on the gradient of the current learning task but also on the parameter importance determined by EWC to prevent important parameters from changing drastically due to adapting to new data. This operation ensures that the model retains its memory of previously learned content while incorporating new knowledge, thereby generating an updated model parameter set. This provides an effective way for the system to maintain its adaptability and accuracy in a constantly changing data environment.

[0094] Please see Figure 2 and Figure 6 The segmentation strategy adjustment module includes an adaptive evaluation submodule, a strategy selection submodule, and a parameter configuration submodule.

[0095] The adaptive evaluation submodule analyzes the compatibility of the current dataset features with multiple segmentation strategies based on the updated model parameter set, calculates the adaptive score for each strategy, and generates adaptive score data.

[0096] The strategy selection submodule compares the scores of multiple segmentation strategies based on the adaptability score data, selects the strategy with the best score as the key segmentation strategy, and determines that the selected strategy conforms to the characteristics of the current dataset, thus obtaining the selected optimal segmentation strategy.

[0097] The parameter configuration submodule adjusts the parameters required for the segmentation operation, including segmentation depth and edge detection sensitivity, based on the selected optimal segmentation strategy, to optimize the effect of the segmentation strategy and match new data features and model parameters to obtain an optimized parameter configuration.

[0098] In the adaptive evaluation submodule, based on the updated model parameter set, a compatibility analysis is performed on the current dataset features and multiple segmentation strategies. A machine learning evaluation algorithm is adopted, specifically using Python language in conjunction with the accuracy_score function in the sklearn.metrics library to calculate the classification accuracy of each strategy for the current dataset. The accuracy scores of each segmentation strategy are weighted and averaged using the NumPy library, taking into account the sensitivity of each strategy to different categories of data, to generate adaptive score data.

[0099] In the strategy selection submodule, based on the adaptive scoring data, the scores of various segmentation strategies are compared. A decision tree algorithm is adopted, specifically using the decision tree classifier in the sklearn.tree library. By setting the criterion='entropy' parameter, the segmentation strategy with the highest information gain is selected as the key segmentation strategy. It is also determined that the selected strategy conforms to the characteristics of the current dataset. The model is trained by calling the fit function to obtain the optimal segmentation strategy.

[0100] In the parameter configuration submodule, the parameters required for the segmentation operation are adjusted according to the selected optimal segmentation strategy. A genetic algorithm is used, specifically the creator and base modules in the deap library. By defining the fitness function and evolution strategy, the initial population for segmentation depth and edge detection sensitivity is set. The optimal parameter configuration is searched through iterative optimization, taking into account the influence of new data features and model parameters, to obtain the optimized parameter configuration.

[0101] Please see Figure 2 and Figure 7 The refined segmentation execution module includes a preliminary segmentation submodule, a region processing submodule, and a range adjustment submodule;

[0102] The preliminary segmentation submodule performs preliminary segmentation on the image based on optimized parameter configuration, identifies key objects and background in the image, and marks areas with unclear boundaries as requiring processing to generate a preliminary segmentation image.

[0103] The region processing submodule, based on the preliminary segmented image, analyzes the marked blurred or overlapping regions that need to be processed using the backpropagation algorithm, and optimizes the regions and boundaries to obtain the processed region image.

[0104] The range adjustment submodule evaluates the segmentation effect of the image based on the processed region image. If the segmentation range is inaccurate or there are omissions, the Actor-Critic algorithm is used to dynamically adjust the range to cover all related objects and generate a refined segmented image set.

[0105] The preliminary segmentation submodule is based on optimized parameter configuration. This module first executes a deep learning model on the input image, specifically using an improved convolutional neural network (CNN) for image analysis. This network enhances the model's ability to identify key objects and backgrounds in the image by adding convolutional and pooling layers. During this process, the model determines the boundary between objects and backgrounds by analyzing the pixel intensity and color distribution of the image. If the algorithm detects areas with unclear boundaries, the system automatically marks the areas as requiring further processing to ensure that the areas receive special attention in subsequent steps. This process generates the preliminary segmentation image.

[0106] The region processing submodule receives the preliminary segmented image as input. For blurry or overlapping regions marked as requiring further processing, this module uses the backpropagation algorithm for detailed analysis and processing. In practice, this module optimizes the ability to identify blurry boundaries by adjusting network weights to improve the accuracy of segmentation. During this process, the algorithm focuses on adjusting pixel values ​​that differ significantly from the features of known objects to make the edges clearer, thereby optimizing the overall region and boundary definition and generating the processed region image.

[0107] The range adjustment submodule is based on the processed region image. This module uses the Actor-Critic algorithm to evaluate the segmentation effect. In this strategy, the Actor part is responsible for proposing adjustments to the segmentation range, while the Critic part evaluates the impact of the adjustments on the segmentation effect, thereby achieving dynamic adjustment of the segmentation range. In this way, the module can ensure that all relevant objects are accurately covered, especially those regions that were missed or inaccurately judged in the initial segmentation process, and finally generate a refined set of segmented images.

[0108] The execution formula for the range adjustment submodule is as follows:

[0109] Formula 1

[0110] Formula 2

[0111] in, and This indicates the update amount of parameters in the Actor and Critic, used to optimize the strategy and value function;

[0112] and The learning rate controls the step size for updating the Actor and Critic parameters;

[0113] The policy gradient indicates how to adjust the parameters. To increase the probability of the selected action;

[0114] A(s, a) is the dominance function, representing the state... Take action below The advantage relative to the average level is used to guide the actor's learning direction;

[0115] Dynamic weights are used to adjust the learning rate based on task complexity and learning progress.

[0116] This is the segmentation sensitivity coefficient, which enhances the model's sensitivity to segmentation boundaries;

[0117] For immediate reward, it indicates the execution of the action. The rewards given by the post-environmental environment;

[0118] This is a discount factor used to calculate the present value of future rewards, indicating the importance of future rewards to the current decision;

[0119] and These are the value estimates for the current state and the next state, provided by Critic, used to estimate the expected return starting from the current state;

[0120] For the value function with respect to the parameters The gradient indicates how to adjust To improve value estimation;

[0121] As a dynamic adjustment factor, the impact of historical adjustment effects on current action decisions is considered;

[0122] These are environmental response parameters that reflect the impact of environmental changes on decision-making, enabling the model to better adapt to different environmental conditions.

[0123] The execution process is as follows:

[0124] The Actor part depends on the current environment. Propose an action This step utilizes a strategy That is, action In state The probability distribution and policy parameters are as follows. The update is performed using a gradient ascent method, with the goal of maximizing the expected reward. Furthermore, a segmentation sensitivity coefficient is incorporated. This enhances the model's sensitivity to segmentation boundaries, allowing for more detailed adjustments.

[0125] The Critic section evaluates the actions proposed by the Actor. In a given state The value is determined by the value function. To estimate the expected reward of the current state, and to calculate the actual reward. With future state The difference between the estimated returns is used to update the parameters of the value function. By dynamically adjusting factors and environmental response parameters The Critic component is better able to adapt to environmental changes and the effects of historical decisions.

[0126] Please see Figure 2 and Figure 8 The multi-scale fusion module includes a resolution analysis submodule, a fusion strategy determination submodule, and a detail and coherence optimization submodule.

[0127] The resolution analysis submodule examines each image one by one based on a refined segmented image set, analyzes the resolution characteristics of the image, identifies the resolution level of each image, and records its features to provide a basis for the fusion strategy and generate resolution feature data.

[0128] The fusion strategy determination submodule calculates the optimal ratio parameters for image fusion by analyzing the pixel size ratio and distribution differences between resolution images based on resolution feature data, and obtains the fusion strategy parameters based on detail preservation and visual coherence between images.

[0129] The detail and coherence optimization submodule applies fusion strategy parameters to dynamically adjust the image, optimizes areas with lost details or poor visual coherence, adjusts the target area, and improves the quality and visual effect of the fused image to obtain a multi-scale fused image.

[0130] In the resolution analysis submodule, based on the refined segmented image set, each image is examined one by one. Image processing algorithms are used, specifically the cv2.resize function in the OpenCV library, to adjust the image size to unify the image processing benchmark. By calculating the pixel density of the image, pixel point analysis is performed using the NumPy library to identify the resolution level of each image and record its features, generating resolution feature data.

[0131] In the fusion strategy determination submodule, an image fusion algorithm is adopted based on resolution feature data. Specifically, the Image.blend method in the PIL library is used to calculate the optimal ratio parameter for image fusion according to the pixel size ratio and distribution difference between resolution images. The pixel weights between images are balanced by adjusting the alpha fusion parameter to control the fusion intensity, maintain the details and visual coherence between images, and obtain the fusion strategy parameters.

[0132] In the detail and coherence optimization submodule, the fusion strategy parameters are applied to dynamically adjust the image. An image enhancement algorithm is used, specifically the `exposure.adjust_gamma` function from the `skimage` library, to adjust the gamma value of the image to optimize detail representation. For areas with lost details or poor visual coherence, the `cv2.GaussianBlur` method is used for local blurring to reduce visual tortuosity, adjust the target area, and optimize the quality and visual effect of the fused image, resulting in a multi-scale fused image.

[0133] Please see Figure 2 and Figure 9The edge optimization module includes an edge detection submodule, a brightness and color adjustment submodule, and an edge smoothing submodule;

[0134] The edge detection submodule performs pixel contrast analysis on the multi-scale fused image, identifies edge regions that differ from the contrast of surrounding pixels, and marks the regions based on blurred or discontinuous boundary lines to generate edge detection records.

[0135] The brightness and color adjustment submodule analyzes the marked edge regions based on edge detection records, adjusts the brightness and color values ​​to match the surrounding areas, and determines the continuity and consistency between the edges and the surrounding environment through step-by-step fine-tuning, reducing visual breaks and obtaining the adjusted image.

[0136] The edge smoothing submodule smooths the marked edges in the adjusted image, gradually reducing jagged edges, while adjusting the pixel gradients near the edges to optimize the smooth transition from the edges to the surrounding areas, resulting in an edge-optimized image.

[0137] In the edge detection submodule, pixel contrast analysis is performed based on multi-scale fused images. Image processing algorithms are adopted, specifically using Python language in conjunction with the cv2.Canny function from the OpenCV library to perform edge detection on the image. By setting minimum and maximum threshold parameters, the sensitivity of pixel contrast is determined, and edge regions with contrast differences from surrounding pixels are identified. The cv2.dilate function is used to enhance the detected edge regions, making the edges more obvious. Based on blurred or discontinuous boundary lines, the regions are marked to generate edge detection records.

[0138] In the brightness and color adjustment submodule, based on edge detection records, the marked edge regions are analyzed. A color space conversion algorithm is used, specifically the cv2.cvtColor function, to convert the image from the RGB color space to the HSV color space. The cv2.inRange method is used to set thresholds for brightness and color values, filtering out the edge regions that need adjustment. The cv2.addWeighted function is used to gradually fine-tune the brightness and color of the selected regions, matching them to the surrounding areas to obtain the adjusted image.

[0139] In the edge smoothing submodule, the marked edges in the adjusted image are smoothed using a pixel gradient adjustment algorithm. Specifically, the cv2.GaussianBlur function is used to apply Gaussian blur to the marked edge regions. By setting the kernel size parameter, the jagged edges are gradually reduced. At the same time, the cv2.addWeighted function is used to adjust the pixel gradient near the edge, optimizing the smooth transition from the edge to the surrounding area, resulting in an edge-optimized image.

[0140] Please see Figure 2 and Figure 10 The results calibration module includes an accuracy verification submodule, a format adjustment submodule, and a standard verification submodule.

[0141] The accuracy verification submodule performs a point-by-point comparative analysis based on the edge-optimized image, compares the segmentation results with the original image, identifies the accuracy of the segmentation lines and the preservation of details, verifies the consistency between the natural transition after edge processing and the original image, and generates accuracy verification data.

[0142] The format adjustment submodule adjusts the image output resolution based on the accuracy test data to match the target output requirements. At the same time, it converts the image format to match the compatibility of the target platform or storage standard and formats the image to obtain the format-adjusted image.

[0143] The standard verification submodule performs quality checks on the format-adjusted images to determine whether each image meets the predetermined segmentation accuracy, detail representation, and format standards, including resolution, color depth, and file type, and establishes an image segmentation result set.

[0144] In the accuracy verification submodule, a point-by-point comparative analysis is performed based on the edge-optimized image. Image processing algorithms are employed, specifically using Python in conjunction with the cv2.absdiff function from the OpenCV library to calculate the absolute difference between the segmentation result and the original image. The cv2.threshold method is then used to convert the difference value into a binary image to identify the accuracy of the segmentation lines and the preservation of details. The cv2.findContours function is then used to detect contours in the binary image to verify the consistency between the natural transition after edge processing and the original image, generating accuracy verification data.

[0145] In the format adjustment submodule, based on the accuracy test data, the image output resolution is adjusted by using an image scaling algorithm. Specifically, the Image.resize method in the PIL library is used to set new image size parameters to match the target output requirements. At the same time, the image format is converted by using the Image.save method to specify the output format parameters to match the compatibility of the target platform or storage standard. The image is then formatted to obtain the format-adjusted image.

[0146] In the standard verification submodule, the quality of the format-adjusted images is checked using an image quality assessment algorithm. Specifically, the cv2.cvtColor function from the OpenCV library is used to convert the image to grayscale to simplify the analysis process. The cv2.calcHist method is used to calculate the histogram of the grayscale image to evaluate the color depth of the image. Combined with file operation commands, the image file type is checked to determine whether each image meets the predetermined segmentation accuracy, detail representation, and format standards, including resolution, color depth, and file type, and to establish an image segmentation result set.

[0147] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-driven image segmentation system, characterized in that: The system includes an image parsing module, a feature encoding module, an incremental learning module, a segmentation strategy adjustment module, a fine segmentation execution module, a multi-scale fusion module, an edge optimization module, and a result calibration module. The image parsing module parses the input image data, identifies the basic constituent elements in the image, including color blocks, shape outlines and texture patterns, and generates an element mapping table. The feature encoding module encodes the element mapping table, assigns encoding weights based on the frequency and spatial distribution of elements, and creates a feature set to generate a set of feature vectors. The incremental learning module analyzes the difference between the feature vector set and the existing dataset, adjusts the data input ratio according to the size of the difference, adopts an incremental learning strategy, dynamically modifies the weight allocation, updates the model parameters, and generates an updated model parameter set. The segmentation strategy adjustment module evaluates the adaptability of the differentiated segmentation strategy to the current dataset based on the updated model parameter set, selects the strategy with the best matching degree, adjusts the segmentation parameters, and generates an optimized parameter configuration. The refined segmentation execution module applies optimized parameter configuration to segment the image, uses judgment logic to process blurred or overlapping regions in the segmentation, adjusts the segmentation range, and generates a refined set of segmented images. The refined segmentation execution module includes a preliminary segmentation submodule, a region processing submodule, and a range adjustment submodule; The preliminary segmentation submodule performs preliminary segmentation processing on the image based on optimized parameter configuration, identifies key objects and background in the image, and marks areas with unclear boundaries as requiring processing to generate a preliminary segmentation image. The region processing submodule, based on the preliminary segmented image, analyzes the marked blurred or overlapping regions that need to be processed using the backpropagation algorithm, and optimizes the regions and boundaries to obtain the processed region image. The range adjustment submodule evaluates the segmentation effect of the image based on the processed region image. If the segmentation range is inaccurate or there are omissions, the Actor-Critic algorithm is used to dynamically adjust the range to cover all related objects and generate a refined segmented image set. The execution formula for the range adjustment submodule is as follows: Set 1 Formula 2 in, and This indicates the amount of parameter updates in the Actor and Critic. and For learning rate, Let A(s, a) be the policy gradient, and A(s, a) be the advantage function. To learn dynamic weights, This is the segmentation sensitivity coefficient. For instant rewards, As a discount factor, and These are the value estimates for the current state and the next state, respectively. For the value function with respect to the parameters gradient, As a dynamic adjustment factor, For environmental response parameters; The multi-scale fusion module fuses the refined segmented image set at different resolutions, adjusts the fusion ratio and strategy, and optimizes the detail display and image coherence to generate a multi-scale fused image. The edge optimization module analyzes and corrects the edges of the multi-scale fused image, adjusts the brightness and color of edge pixels, matches the surrounding area, optimizes edge blur and breakage, and generates an edge-optimized image. The result calibration module verifies the edge-optimized image, checks the segmentation accuracy and detail performance, adjusts the output resolution and format, determines that it meets predetermined standards, and generates an image segmentation result set.

2. The deep learning-driven image segmentation system according to claim 1, characterized in that: The element mapping table includes color distribution regions, key shape boundaries, and texture feature identifiers. The feature vector set includes color block weight distribution, shape contour encoding, and texture pattern frequency. The updated model parameter set includes adjusted inter-layer connection weights, neuron bias values, and activation function parameters. The optimized parameter configuration includes segmentation optimization threshold, region expansion control factor, and shape correlation evaluation criteria. The image segmentation result set includes a segmentation effect comparison analysis chart and a detail restoration verification chart.

3. The deep learning-driven image segmentation system according to claim 1, characterized in that: The image analysis module includes a color analysis submodule, a shape analysis submodule, and a texture recognition submodule; The color analysis submodule performs color spectrum analysis on the pixels in the image based on the input image data, counts the occurrence frequency and distribution location of each color, and determines the key color regions in the image to obtain color distribution data. The shape parsing submodule identifies the basic shape contours in the image based on the color distribution data, divides the image into multiple shape feature regions, and obtains a shape contour dataset. The texture recognition submodule analyzes the texture features of each shape region based on the shape contour dataset, including the identification of texture type, direction and repetition pattern, distinguishing feature regions of different textures in the image, and establishing an element mapping table.

4. The deep learning-driven image segmentation system according to claim 1, characterized in that: The feature encoding module includes a weight allocation submodule, an encoding conversion submodule, and a set creation submodule; The weight allocation submodule performs key evaluations on various color, shape, and texture features based on the element mapping table, and assigns corresponding weight values ​​according to the visual impact and recognition difficulty of the features in the image, thereby generating weight data. The encoding conversion submodule applies weight data to features, converts features into unique codes for representation, including color, shape and texture features, and assigns corresponding encoding segments to obtain encoded feature data; The set creation submodule aggregates the encoded feature data and verifies the color, shape, and texture features of the images in the set, providing a data foundation for image processing and analysis tasks and generating a set of feature vectors.

5. The deep learning-driven image segmentation system according to claim 1, characterized in that: The incremental learning module includes a difference analysis submodule, a learning strategy adjustment submodule, and a model parameter update submodule; The difference analysis submodule is based on the feature vector set and uses flexible weight sharing to compare the existing dataset, calculate the difference between the feature vector and the existing model. If the difference exceeds a preset threshold, it is marked as high difference; otherwise, it is marked as low difference, and a difference analysis record is generated. The learning strategy adjustment submodule uses a dynamic data buffering mechanism based on the difference analysis record. If the difference is high, the proportion of new data input is increased, and the learning rate and weight allocation are dynamically adjusted. If the difference is low, the original learning rate is maintained, and the weight allocation is adjusted to obtain learning strategy adjustment data. The model parameter update submodule adjusts the data based on the learning strategy, updates the model's weights, bias terms, and correlation parameters, matches new data features, continuously optimizes the model, and generates an updated model parameter set.

6. The deep learning-driven image segmentation system according to claim 1, characterized in that: The segmentation strategy adjustment module includes an adaptive evaluation submodule, a strategy selection submodule, and a parameter configuration submodule; The adaptive evaluation submodule analyzes the compatibility of the current dataset features with multiple segmentation strategies based on the updated model parameter set, calculates the adaptive score for each strategy, and generates adaptive score data. The strategy selection submodule compares the scores of multiple segmentation strategies based on the adaptive score data, selects the strategy with the best score as the key segmentation strategy, and determines that the selected strategy conforms to the characteristics of the current dataset, thus obtaining the selected optimal segmentation strategy. The parameter configuration submodule adjusts the parameters required for the segmentation operation, including segmentation depth and edge detection sensitivity, based on the selected optimal segmentation strategy, to optimize the effect of the segmentation strategy and match new data features and model parameters to obtain an optimized parameter configuration.

7. The deep learning-driven image segmentation system according to claim 1, characterized in that: The multi-scale fusion module includes a resolution analysis submodule, a fusion strategy determination submodule, and a detail and coherence optimization submodule; The resolution analysis submodule examines each image one by one based on a refined segmented image set, analyzes the resolution characteristics of the image, identifies the resolution level of each image, and records its features to provide a basis for the fusion strategy and generate resolution feature data. The fusion strategy determination submodule calculates the optimal ratio parameters for image fusion based on resolution feature data by analyzing the pixel size ratio and distribution differences between resolution images, and obtains fusion strategy parameters based on detail preservation and visual coherence between images. The detail and coherence optimization submodule applies fusion strategy parameters to dynamically adjust the image, optimizes areas with lost details or poor visual coherence, adjusts the target area, and optimizes the quality and visual effect of the fused image to obtain a multi-scale fused image.

8. The deep learning-driven image segmentation system according to claim 1, characterized in that: The edge optimization module includes an edge detection submodule, a brightness and color adjustment submodule, and an edge smoothing submodule; The edge detection submodule performs pixel contrast analysis on the multi-scale fused image, identifies edge regions that differ from the contrast of surrounding pixels, and marks the regions based on blurred or discontinuous boundary lines to generate edge detection records. The brightness and color adjustment submodule analyzes the marked edge regions based on edge detection records, adjusts the brightness and color values ​​to match the surrounding areas, and determines the continuity and consistency between the edges and the surrounding environment through step-by-step fine-tuning, thereby reducing visual breaks and obtaining the adjusted image. The edge smoothing submodule smooths the marked edges in the adjusted image, gradually reducing jaggedness, and adjusts the pixel gradients near the edges to optimize the smooth transition from the edges to the surrounding areas, resulting in an edge-optimized image.

9. The deep learning-driven image segmentation system according to claim 1, characterized in that: The result calibration module includes an accuracy verification submodule, a format adjustment submodule, and a standard confirmation submodule. The accuracy verification submodule performs a step-by-step comparative analysis based on the edge-optimized image, compares the segmentation results with the original image, identifies the accuracy of the segmentation lines and the preservation of details, and verifies the consistency between the natural transition after edge processing and the original image, generating accuracy verification data. The format adjustment submodule adjusts the image output resolution according to the accuracy test data to match the target output requirements. At the same time, it converts the image format to match the compatibility of the target platform or storage standard and formats the image to obtain the format-adjusted image. The standard verification submodule performs quality checks on the format-adjusted images to determine whether each image meets the predetermined segmentation accuracy, detail representation, and format standards, including resolution, color depth, and file type, and establishes an image segmentation result set.

Images