An image region analysis method based on entropy-driven feature enhancement

Abstract

The application discloses an image region analysis method based on entropy-driven feature enhancement, and relates to the technical field of image analysis and feature enhancement.The method takes local channel information entropy as a core feature statistic quantity, realizes adaptive weighting and strengthening of features of different importance regions through explicit quantization of image feature information, dynamically adjusts weight distribution according to feature values, significantly enhances the response of high information-intensive regions, and effectively suppresses low information and noise interference regions.On this basis, a cross-layer attention mechanism based on entropy prior is designed, feature statistical information is embedded into the gate and weight generation process, and attention distribution guided by feature saliency is realized.Through the entropy-driven adaptive feature enhancement mechanism, the application effectively improves the perception ability, feature discrimination and analysis accuracy of the model to the salient regions of the image.

Description

Technical Field

[0001] This invention relates to the field of image analysis and feature enhancement technology, and in particular to an image region analysis method based on entropy-driven feature enhancement. Background Technology

[0002] Image region analysis, as a crucial component of computer vision and intelligent image understanding, is a core foundation for various tasks such as target recognition, scene segmentation, anomaly detection, target tracking, and semantic understanding. With the widespread availability of high-resolution, multimodal images, the spatial structure, texture details, and semantic information contained within images are becoming increasingly rich, providing a more robust foundation for fine-grained region recognition, structural analysis, and feature mining. However, how to efficiently and accurately perform image region analysis in complex backgrounds has become one of the key technical challenges that urgently needs to be overcome in the field of computer vision.

[0003] The core objective of image region analysis is to accurately extract discriminative regional features from large-scale complex images, supporting various downstream intelligent perception and decision-making tasks. Its research significance lies in: providing accurate regional boundaries and semantic information for automated recognition, segmentation, and detection; helping to mine potential spatial relationships and structural features in high-dimensional and diverse images, enhancing the model's comprehensive understanding of global and local scene information; and effectively improving the computational efficiency, robustness, and generalization performance of subsequent tasks by accurately identifying key regions and suppressing redundant information, while reducing resource consumption in large-scale system deployments.

[0004] In existing technologies, structures such as Convolutional Neural Networks (CNNs) and Transformers have made significant progress in multi-scale feature extraction and global context modeling. However, existing image region analysis methods still face significant challenges: First, regional information is highly heterogeneous, with significant differences in information entropy between different regions. A unified processing strategy weakens the representation ability of high-entropy key regions. Second, factors such as external illumination, sensor differences, noise, and acquisition conditions introduce a large number of pseudo-features, which interfere with the real regional features and reduce the accuracy and robustness of the analysis. Third, existing models lack uncertainty modeling and saliency identification based on information content. CNNs have limitations in modeling long-range dependencies, and although Transformers have global capabilities, their attention is evenly distributed, making it difficult to highlight high-entropy regions and consuming too many resources in low-entropy regions. Fourth, the models still have shortcomings in terms of parameter scale, training cost, and multi-scale feature consistency, making it difficult to meet the high-precision and high-efficiency requirements for region analysis in complex scenarios.

[0005] To address the above issues, there is an urgent need for a method that can explicitly quantify regional information entropy and perform differential modeling based on saliency levels, thereby further improving the accuracy, efficiency, and robustness of image region analysis. Summary of the Invention

[0006] To address the problem of insufficient region recognition accuracy in existing high-resolution image region analysis due to uneven distribution of feature information and interference from pseudo-features, this invention proposes an image region analysis method based on entropy-driven feature enhancement. By introducing an entropy-driven mechanism into the feature extraction, weight allocation, and multi-scale fusion stages, the method strengthens the feature representation of high-information regions and suppresses redundant interference in low-information regions, thereby achieving accurate, efficient, and robust analysis of image regions.

[0007] This invention adopts the following technical solution: an image region analysis method based on entropy-driven feature enhancement, comprising the following steps:

[0008] Step 1, Data Acquisition: Acquire multi-source image data as input samples and perform preprocessing operations;

[0009] Step 2, Feature Extraction and Representation: Use deep convolutional neural networks or transform network structures to extract features from the preprocessed image, obtain multi-layer feature maps, and realize hierarchical feature representation of the image;

[0010] Step 3: Construct an entropy-driven feature adaptive enhancement module. By calculating the multi-scale feature entropy distribution, an entropy weight mask is generated to strengthen high-entropy discriminative features, suppress low-entropy redundant features, and improve the discriminative ability and semantic representativeness of the feature layer.

[0011] Step 4: Construct an entropy-focused cross-space attention module. By combining entropy weights with multi-scale feature interactions, attention is allocated based on entropy priors in the spatial dimension, focusing on information-significant regions and enhancing the consistent perception and robust representation of real semantic regions.

[0012] Step 5: Output a black and white binary prediction image through the image transformation detection system, and perform image region analysis. Black represents areas where no changes were detected, and white represents areas where changes were detected.

[0013] Preferably, the multi-source image data includes: remote sensing images, medical images, and natural scene image data.

[0014] Perform uniform preprocessing operations on different types of image data, including: normalization and size standardization, channel alignment and color space conversion.

[0015] Preferably, in the multi-layer feature map, the lower-layer feature map contains local details of texture, edge, and brightness, while the higher-layer feature map contains semantic structure and spatial context information.

[0016] Preferably, the entropy-driven feature adaptive enhancement module quantifies the information density distribution of image features by calculating the channel information entropy of image data, as follows:

[0017] The feature values ​​of each channel are normalized by min-max to linearly map them to a unified numerical range;

[0018] feature map The spatial dimension is flattened into a one-dimensional vector, and the information uncertainty and complexity of the feature region are evaluated by quantifying the frequency of occurrence of each feature value;

[0019] Calculate the entropy value. The higher the entropy value, the greater the uncertainty of the corresponding channel information, and the richer the details and variation information. The lower the entropy value, the more uniform the distribution of the corresponding channel features, the higher the information redundancy, and the more singular the feature patterns.

[0020] Furthermore, based on the calculated entropy values ​​of each channel, the information complexity contained in different channels is quantified, and the feature maps are adaptively weighted.

[0021] The normalized entropy value is multiplied with the original feature map as a significance weight to enhance the discriminative features of low-entropy flat regions and suppress redundant background interference in high-entropy complex regions, thus highlighting key regions.

[0022] Preferably, the entropy-focused cross-space attention module uses local entropy prior as a feature-level guiding signal, embeds feature entropy values ​​into the attention weight generation and gating calculation process, and fuses them with deep semantic features to locate discriminative features in key regions. The method is as follows:

[0023] Step 4.1: Input the denoised and information entropy-weighted dual-temporal feature map, perform pooling processing, and perform max pooling and average pooling operations on each temporal feature to achieve a balance between high-entropy local regions and overall spatial distribution.

[0024] Step 4.2: Concatenate the dual-phase max pooling feature and average pooling feature to output the concatenated four-channel feature map;

[0025] Step 4.3: Apply a convolution operation to the stitched four-channel feature map to capture the correlation of neighborhood features;

[0026] Step 4.4: Perform a non-linear transformation on the convolution result using the Sigmoid function to generate a spatial attention weight map, which represents the importance of spatial locations;

[0027] Step 4.5: Multiply the spatial attention weight map element-wise with the original entropy weighted features to achieve adaptive weighting of the image spatial features.

[0028] Preferably, in step 4.1, max pooling is used to retain significant responses in local regions and highlight the significance of high-entropy regions; average pooling is used to extract global statistical features and reflect the overall information distribution.

[0029] Preferably, in step 4.1, the pooling process further includes: statistical histogram pooling, Norm pooling is used to enhance the diversity of feature representation.

[0030] Preferably, in step 4.3, the kernel size is 7×7, 3×3, or 5×5, or dilated convolution is used.

[0031] Preferably, in step 4.4, the Softmax function is used to replace the Sigmoid function to normalize the weight allocation of spatial locations.

[0032] Preferably, the image region analysis is performed using the image transformation detection system, as follows:

[0033] Step 5.1: Input dual-temporal image data, construct a dual-branch Siamese network using the first four stages of EfficientNet-V2-S, and extract dual-temporal image features through the decoder;

[0034] Step 5.2: Input the dual-temporal image features into the entropy-driven feature adaptive enhancement module, calculate the multi-scale feature entropy value, enhance the discriminative features related to real changes, and suppress feature interference corresponding to pseudo-changes;

[0035] Step 5.3: Distinguish between object edge and morphological change features using the entropy-focused cross-space attention module;

[0036] Step 5.4: Input the feature map into the Uni-FIRE module (Unified Fusion and Refinement Module) for interpolation processing:

[0037] The main branch passes through the first convolutional layer, the SiLU activation layer, and the batch normalization layer in sequence, and then through the second convolutional layer and the batch normalization layer to output the prediction map;

[0038] The skip branch uses 1×1 convolution for channel compression and fusion;

[0039] The skip branch is added to the main branch output to generate a multi-scale fused feature map.

[0040] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0041] 1. The method of the present invention introduces information entropy as a measure of feature saliency, thereby achieving self-enhancing weighting and suppression of discriminative features, which significantly improves the discriminative power and interpretability of feature expression.

[0042] 2. The method of this invention utilizes an entropy-driven cross-layer attention mechanism to enhance the response to key regional features at the feature level and effectively suppress spurious feature interference introduced by factors such as acquisition conditions, noise, and background complexity, thereby enhancing the model's regional analysis accuracy and robustness in complex scenarios.

[0043] 3. The method of this invention significantly improves the feature quality, model discrimination ability and overall interpretability in image region analysis tasks by introducing a physically guided adaptive weighting and interaction mechanism into the feature representation, providing effective technical support for intelligent analysis and understanding of high-resolution images. Attached Figure Description

[0044] Figure 1 This is a structural diagram of the entropy-driven feature adaptive enhancement module in this invention;

[0045] Figure 2 This is a structural diagram of the entropy-focused cross-space attention module in this invention;

[0046] Figure 3 This is a schematic diagram illustrating the calculation of entropy values ​​for features in this invention;

[0047] Figure 4 This is a comparative analysis of the feature adaptive enhancement results of two different images based on the entropy value of features according to the present invention;

[0048] Figure 5 This is a diagram showing the overall network structure of the image transformation detection system built in this invention to verify the image region analysis effect based on entropy-driven feature enhancement.

[0049] Figure 6 To verify the structure diagram of the Uni-FIRE module used in the network;

[0050] Figure 7 This is a visualization comparison of the transformation detection results of this invention with mainstream methods on different datasets. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0052] In one embodiment of the present invention, a feature-driven and adaptive weighted image region analysis method is proposed to address key issues in high-resolution image region analysis, such as blurred feature representation, significant pseudo-feature interference, and difficulty in effectively extracting discriminative features due to complex scene structures, highly heterogeneous spatial distribution, and differences in acquisition conditions.

[0053] First, by establishing a feature saliency quantification mechanism based on information entropy, feature-level differentiation and adaptive weighted enhancement of high and low information regions are achieved, thereby suppressing redundancy and interference in low feature information regions.

[0054] Then, a cross-layer attention fusion framework is designed to deeply fuse multi-scale structural information at the feature level, thereby reducing the pseudo-feature response caused by differences in acquisition conditions and complex backgrounds.

[0055] Ultimately, a regional analysis model was constructed that is physically interpretable for key areas and highly robust to complex environments, thereby improving the clarity of feature representation and the accuracy of regional identification.

[0056] The specific steps of the image region analysis method in this embodiment are as follows:

[0057] Step 1: Data acquisition and preprocessing.

[0058] Multi-source image data is collected as input samples, such as remote sensing images, medical images, or natural scene images. Uniform preprocessing operations are performed for different types of image data, mainly including:

[0059] Image normalization and size standardization;

[0060] Channel alignment and color space conversion (such as RGB→Lab or HSV).

[0061] In this embodiment, the acquired image data is multi-temporal high-resolution remote sensing image data. By calculating the local information entropy of each pixel channel and performing explicit quantization, adaptive weighting and enhancement of image feature information are achieved. The analysis results yield significantly enhanced feature maps and change response results, in which features of high-information-dense areas are highlighted, and low-information and noise areas are effectively suppressed. This provides a data foundation for subsequent applications such as remote sensing image change detection, land cover classification, and target recognition.

[0062] Step 2: Feature extraction and representation.

[0063] The preprocessed image is used to extract features by employing a deep convolutional neural network (CNN) or a transform network (Transformer) structure. After the network extracts features, multi-layer feature maps are obtained. This process realizes the hierarchical feature representation of the image, providing basic feature data for subsequent information entropy calculation.

[0064] In this embodiment, the multi-layer feature map includes a low-layer feature map and a high-layer feature map. The low-layer feature map mainly contains local detail information such as texture, edge, and brightness, while the high-layer feature map contains semantic structure and spatial context information.

[0065] Step 3: Construct an entropy-driven feature adaptive enhancement module (EFAE).

[0066] By introducing channel information entropy theory, saliency priors are calculated at the feature level and adaptive weighting is implemented, deeply integrating physical information entropy with deep learning feature selection mechanisms. This module can explicitly enhance the representational ability of high-entropy discriminative features, suppress the interference of low-entropy redundant features on the model, provide the network with interpretable and quantifiable criteria for evaluating feature importance, and significantly improve the expressive power and robustness of key changing features.

[0067] Specifically, different regions in an image (such as a target region with complex texture and a homogeneous background region) show obvious differences in feature response: the more uniform the feature response, the richer the information of its corresponding spatial location, and the higher the entropy value, reflecting that the features of that region have higher complexity and uncertainty.

[0068] Therefore, local information entropy can serve as an effective physical quantification index of image spatial information density, providing a theoretical basis for feature-level saliency modeling and adaptive weighting.

[0069] The structure of the entropy-driven feature adaptive enhancement module constructed in this implementation is as follows: Figure 1 As shown, the specific processing is as follows:

[0070] First, the feature values ​​of each channel are normalized using a min-max method, linearly mapping them to a uniform numerical range (usually [0,1]). This transformation is achieved using the following formula:

[0071] ;

[0072] in, The input are the original feature values. For feature map The minimum value of each channel. For feature map The maximum value of each channel. For smoothing terms, The output is the normalized eigenvalues.

[0073] Furthermore, to facilitate probability-based entropy calculation, such as Figure 2As shown, the spatial dimension of the feature map is flattened into a one-dimensional vector. The calculation of entropy depends on the probability distribution of the feature values ​​in the whole. By quantifying the frequency of each feature value, the information uncertainty and complexity of the feature region can be evaluated.

[0074] The probability estimation uses the following formula:

[0075] ;

[0076] in, Represents the th element in the characteristic matrix The probability of each position. The flattened feature matrix is The value at that location.

[0077] The entropy value is calculated using the following formula:

[0078] ;

[0079] in, Indicates the first Entropy of the channel The smoothing constant term is usually set to... , Scale-invariant entropy weights are used to eliminate the impact of feature map size changes on entropy calculation, ensuring that entropy measurements are consistent and comparable across different scales.

[0080] It is particularly important to note that: the higher the entropy value, the greater the uncertainty of the information in the corresponding channel, and the richer the details and variation information; conversely, a lower entropy value indicates that the feature distribution of the channel is more uniform, the information redundancy is higher, and the feature patterns contained are more singular.

[0081] By calculating the entropy value of each channel, the information complexity contained in different channels is quantified, thereby achieving adaptive weighted processing of the feature map. The weighting method is implemented according to the following formula:

[0082] ;

[0083] in, These are the weighted eigenvalues.

[0084] Furthermore, by multiplying the normalized entropy value as a saliency weight with the original feature map, adaptive enhancement of high-entropy regions and corresponding suppression of low-entropy regions are achieved.

[0085] This feature enhancement mechanism has a clear physical meaning. Its essence is a quantitative assessment of the spatial feature information content, which can guide subsequent network modules to focus on information-rich areas, thereby improving the feature representation ability and scene adaptability in subsequent tasks based on this image analysis.

[0086] Step 4: Construct the Entropy-Focused Cross-SpatialAttention Module (EnFocSA).

[0087] Building upon the aforementioned feature saliency prior module based on information entropy, this embodiment further proposes the EnFocSA module, which uses local entropy prior as a guiding signal at the feature level and deeply integrates it with deep semantic features. Feature saliency information is embedded in the attention weight generation, and attention allocation under physical constraints is achieved through cross-temporal feature interaction and spatial correlation modeling.

[0088] Compared with traditional channel aggregation methods, this module can enhance the response to real-world changes at the feature level, effectively suppress spurious change feature interference caused by factors such as imaging differences, alleviate the problems of blurred feature boundaries and loss of local structural information caused by single-channel weighting, and significantly improve the fine-grained characterization and structural integrity of target region features while maintaining physical interpretability. It also enhances the model's perception and discrimination accuracy of real-world changes, thereby achieving accurate localization and adaptive enhancement of discriminative features in key regions, and providing higher precision and robust feature representation for image region analysis.

[0089] Specifically, such as Figure 3 As shown, the denoised and entropy-weighted dual-temporal feature maps are first input into this module. For the features of each temporal phase, max pooling and average pooling operations are performed respectively.

[0090] Max pooling preserves the most significant responses in local regions, highlighting the salience of high-entropy areas; average pooling, on the other hand, extracts global statistical features, reflecting the overall information distribution. Combining the two achieves a balance between high-entropy local regions and the overall spatial distribution.

[0091] Subsequently, the max-pooling feature and the average-pooling feature of the two phases are concatenated to obtain a four-channel tensor:

[0092] ;

[0093] in, These represent the maximum pooling result and the average pooling result in both time phases, respectively. This is the spliced ​​four-channel feature map.

[0094] To further model local contextual information, this embodiment applies a convolution operation to the concatenated feature map:

[0095] ;

[0096] in, For convolution kernel, This represents the convolution operation. This is the result of the convolution.

[0097] Preferably, the kernel size is 7×7. This operation can capture the correlation of neighborhood features, enhance edge information and local structure expression, and avoid the boundary blurring phenomenon that occurs after feature fusion.

[0098] Next, the convolution result is non-linearly transformed using the Sigmoid function to generate a spatial attention weight map:

[0099] ;

[0100] in, For the Sigmoid function, This is a spatial attention weight map, ranging from 0 to 1. Each position in this weight map represents the importance of that spatial position and has clear physical interpretability.

[0101] Finally, this embodiment multiplies the spatial attention weight map element-wise with the original entropy-weighted features, thereby achieving adaptive weighting of image spatial features. Through this mechanism, the features of high-entropy complex regions are enhanced, while low-entropy or noisy regions are suppressed, achieving the highlighting of key regions and the suppression of redundant information.

[0102] Furthermore, we will conduct visualization analysis.

[0103] In this embodiment, a comparative analysis of the adaptive feature enhancement results of two different images based on the entropy value of the features is performed in the MATLAB environment.

[0104] First, a pre-trained ResNet-18 network is used to extract high-level feature maps from two images. Then, the local entropy value of each feature location is calculated using a sliding window to generate an entropy weight matrix representing the complexity of the region. Next, the normalized original features are multiplied by the entropy weights to achieve adaptive adjustment of the feature values—background suppression is performed on high-entropy complex regions, and feature enhancement is performed on low-entropy flat regions. Finally, by calculating the difference between the weighted features, the signal-to-noise ratio of change detection is significantly improved, allowing real change regions to stand out against the complex background.

[0105] like Figure 4 As shown, Figure 4 (a) in the image is a comparison of the original features and the enhanced features of the T1 image; Figure 4 (b) in the figure is a comparison of the original features and the enhanced features of the T2 image; Figure 4 (c) in the figure is a weighted comparison chart of differences.

[0106] As can be seen from the figure, adaptive enhancement of features is achieved through local entropy weighting. The weight matrix is ​​dynamically generated by utilizing the spatial complexity of deep features to enhance features in low-entropy smooth regions and suppress backgrounds in high-entropy complex regions. This significantly improves the signal-to-noise ratio of changing signals, allowing real changes to be accurately highlighted in backgrounds of varying complexity.

[0107] It should be noted that the 7×7 kernel size mentioned above is not the only option. In other embodiments, 3×3 or 5×5 kernels can be used to reduce computational complexity, or dilated convolution can be used to expand the receptive field. Regarding pooling methods, in addition to max pooling and average pooling, statistical histogram pooling can also be used. Methods such as norm pooling can be used to enhance the diversity of feature representation; the sigmoid function can also be replaced by the softmax function to achieve normalized weight allocation for spatial locations.

[0108] Through the above design, this embodiment's entropy-driven cross-spatial attention module can effectively fuse prior salient features and spatial distribution features in an image to generate a physically meaningful dense spatial attention map. This mechanism not only provides information-theoretic interpretability for feature selection and fusion, but also achieves precise localization and fine-grained enhancement of discriminative features in key regions through multi-scale feature compression and local correlation modeling, thereby significantly improving the accuracy, robustness, and clarity of feature representation in image region analysis.

[0109] Step 5: Construct an image transformation detection system and perform image region analysis.

[0110] To verify the effectiveness of image region analysis based on entropy-driven feature enhancement, this embodiment constructs an image transformation detection system based on the region analysis results. The system structure is as follows: Figure 5 As shown.

[0111] The experimental results and analysis are as follows:

[0112] Assume the two-phase image is Its size is 3×H×W, where H and W represent the height and width of the two images, respectively. As shown in the figure, this detection network uses the first four stages of EfficientNet-V2-S to construct a two-branch Siamese network, extracting features from both temporal images through a decoder. Subsequently, the extracted image features enter an entropy-driven feature enhancement module to perform region analysis. Based on this image analysis, subsequent change detection is achieved, and the entropy-focused cross-spatial attention module distinguishes between object edges and morphological change features.

[0113] Finally, the feature map is input into the Uni-FIRE module, interpolated, and then passed through the first convolutional layer, SiLU activation layer, and batch normalization layer in sequence. After passing through the second convolutional layer and batch normalization layer, the prediction map is output.

[0114] like Figure 6 As shown, after stitching together each input image, it is interpolated to 3×H×W, where H and W represent the height and width of the target interpolation, respectively.

[0115] Subsequently, the interpolated image is input into the main branch and jump connection branches:

[0116] In the main branch, the process sequentially passes through the first convolutional layer, the SiLU activation layer, and the batch normalization layer, and then through the second convolutional layer and the batch normalization layer.

[0117] In the skip connection branch, 1×1 convolution is used for channel compression and fusion.

[0118] Finally, the skip branch is added to the main branch output to generate a multi-scale fused feature map.

[0119] Through the above steps, this embodiment can finally obtain a highly saliency feature representation map after entropy-driven feature enhancement and cross-layer attention fusion optimization.

[0120] The results have the following characteristics and applications:

[0121] (1) Characteristics of the results

[0122] The output feature map significantly enhances the response of high-information-dense regions in the spatial dimension, while suppressing low-information and noisy regions;

[0123] It enhances the discriminative power of key target areas and background in the semantic dimension, making the area boundaries clearer and the response more focused;

[0124] Feature distributions have higher separability and information integrity, and can provide robust input features for subsequent vision tasks.

[0125] (2) Application of the results:

[0126] It can be used as input features or intermediate representations for tasks such as saliency detection, target recognition, change detection, image segmentation, and medical lesion extraction. For example, in remote sensing image analysis, it can be used to highlight high-information areas such as building clusters and road intersections to achieve more accurate detection of ground feature changes. In medical image analysis, it can be used to enhance the features of abnormal areas (such as lesions and tumors) and improve the accuracy of automatic diagnosis.

[0127] Furthermore, the method of this invention was compared with several representative RSCD methods as baseline models, including FC-Siam-Conc, FC-Siam-Diff, SNUNet, ChangeFormer, DASNet, STANet, BiT, and SEIFNet. All baseline methods were reimplemented under the same network configuration to ensure a fair and consistent comparison.

[0128] The results are as follows:

[0129] (1) The effectiveness and complementarity of the entropy plus cross-attention mechanism: When the information entropy saliency prior module is used alone, the F1 score on the LEVIR-CD and WHU-CD datasets has reached 92.04% / 93.49%, respectively, which is 0.15% / 0.49% higher than the baseline network. After the entropy module and the cross-space attention module work together, the F1 score is further improved to 92.25% / 94.27%, which is 0.21% / 0.78% higher than the baseline network, respectively, which verifies the effective synergy of the two modules in image analysis.

[0130] (2) Visual verification is intuitive and effective: This invention demonstrates significant advantages in complex scenarios. The entropy-driven feature adaptive enhancement module can effectively enhance discriminative features related to real changes while suppressing feature interference corresponding to pseudo-changes. Combined with the entropy-focused cross-space attention mechanism, the model excels in the continuous expression of features at change boundaries and the characterization of fine-grained features. It can more clearly distinguish features related to building edges and morphological changes, fully verifying the effectiveness and application potential of this method in extracting high-discriminative features and detecting interpretable changes in complex scenarios.

[0131] Experimental results on the LEVIR-CD and WHU-CD datasets are summarized in Tables I and II, respectively, with the best-performing results highlighted in bold. The results show that the proposed network achieves F1 scores of 92.25% and 94.27% on the LEVIR-CD and WHU-CD datasets, respectively, outperforming several state-of-the-art methods.

[0132] Table I:

[0133]

[0134] Table II:

[0135]

[0136] Specifically, on the WHU-CD dataset, the proposed model outperforms other models on all evaluation metrics; on the LEVIR-CD dataset, although the recall (Rec) is slightly lower than the BIT model, it achieves the highest values ​​on Pre, F1, OA and IoU metrics, demonstrating excellent overall performance.

[0137] To further verify the effectiveness of each key module in this invention, ablation experiments were conducted on two public datasets, LEVIR-CD and WHU-CD, as shown in Tables III and IV below.

[0138] Table III:

[0139]

[0140] Table IV:

[0141]

[0142] The experimental baseline model consists of an encoder built on an efficient convolutional neural network and a decoder and prediction head composed of reusable unified fusion and refinement modules. The innovative modules of this invention are gradually introduced into this baseline model to evaluate their independent contributions and synergistic effects.

[0143] In the first set of experiments, a saliency-adaptive enhancement module based on information entropy was applied to skip connections. This enabled the encoder output features to highlight high-information regions and suppress low-information regions through an entropy-driven weighting mechanism during cross-scale propagation. Results show that this module significantly improves the model's response capability in truly variable regions while effectively reducing false detections caused by noise and spurious differences, resulting in a significant performance improvement over the baseline model.

[0144] In the second set of experiments, an entropy-focused cross-spatial attention module was further introduced on top of the baseline model. This module utilizes entropy priors to guide the generation process of spatial attention weights, promoting cross-spatial interaction and fusion between deep features of dual-temporal images, thereby enhancing the model's ability to locate change-related discriminative features. Experimental results show that this module significantly improves feature boundary continuity preservation, fine-grained feature characterization, and feature robustness in complex backgrounds, fully demonstrating the role of entropy information in feature optimization and enhancement within the spatial attention mechanism.

[0145] Furthermore, to investigate the performance of the entropy-driven cross-space attention module under different convolutional kernel sizes, this embodiment conducted additional comparative experiments. The results show that when the convolutional kernel size is set to... At this time, the model performed best in both detection accuracy and boundary preservation. This conclusion indicates that a larger receptive field can better capture neighborhood correlations and preserve structural details. Therefore, as a preferred option, the kernel size of this module is set to 7.

[0146] The ablation experiments described above demonstrate that the proposed feature-adaptive saliency enhancement module based on information entropy and the entropy-focused cross-space attention module can effectively improve feature discrimination capabilities when used individually. When working together, they achieve complementary enhancement at the feature level: on the one hand, they strengthen the expression of high-entropy discriminative features; on the other hand, they optimize the interaction and attention distribution of cross-temporal features, thereby achieving stable detection and accurate localization of change-related features in complex scenarios. The experimental results verify the rationality of the design concept and the effectiveness of the overall architecture of this invention from a feature perspective.

[0147] Figure 7 The results of change detection by this system on different image samples are shown, and compared with other models. It is evident that the method proposed in this invention demonstrates significant advantages in the identification and enhancement of key region features.

[0148] Specifically, information entropy-based modules, as physically interpretable feature prior mechanisms, can effectively capture discriminative features of potentially changing regions and achieve high-fidelity boundary characterization in information-dense regions (such as...). Figure 7 (c) and (e) in the text significantly improve the semantic representation quality of change features. Simultaneously, the cross-spatial attention mechanism enhances the feature response of truly changing regions and suppresses low-information or noise interference (such as...). Figure 7 (b) and (d) in the model effectively reduce the false alarm rate and improve the reliability of feature selection. Furthermore, this framework demonstrates excellent discriminative ability in preserving edge structure and extracting features from small-scale regions.

[0149] The visualization results further demonstrate that the proposed method can generate feature-coherent and structurally complete change maps, highlighting the robustness of region feature perception and change detection accuracy in complex image scenes.

[0150] In summary, through the entropy-driven feature enhancement and saliency-guided fusion mechanism of this invention, the system can generate image region analysis results with high information fidelity, high saliency response and strong discriminative ability, thereby providing more physically interpretable and robust feature support for complex visual tasks.

[0151] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An image region analysis method based on entropy-driven feature enhancement, characterized in that, The method comprises the following steps: Step 1, data acquisition: collect multi-source image data as input samples and perform preprocessing operations; Step 2, feature extraction and representation: use a deep convolutional neural network or a transformation network structure to extract features from the preprocessed image, obtain a multi-layer feature map, and realize hierarchical feature representation of the image; Step 3, construct an entropy-driven feature adaptive enhancement module, calculate the multi-scale feature entropy distribution, generate an entropy weight mask, strengthen high-entropy discriminative features, suppress low-entropy redundant features, and improve the discriminability and semantic representativeness of the feature layer; Step 4, construct an entropy-focused cross-space attention module, combine entropy weights and multi-scale feature interactions, and perform entropy-prior-based attention allocation in the spatial dimension to focus on areas with significant information and enhance consistent perception and robust representation of real semantic regions; The entropy-focused cross-space attention module uses local entropy prior as a feature layer guide signal, embeds feature entropy values into attention weight generation and gating calculation processes, fuses deep semantic features, and locates discriminative features in key regions, as follows: Step 4.1, input the de-noised and entropy-weighted dual-phase feature map, perform pooling processing, and perform maximum pooling and average pooling operations on each phase feature to balance between high-entropy local regions and overall spatial distribution; Step 4.2, concatenate the maximum-pooled features and the average-pooled features of the dual-phase to obtain a four-channel tensor: ； wherein, respectively represent the maximum pooling result and the average pooling result of the double time phase, is the four-channel feature map after splicing, is a feature splicing function; Step 4.3, apply a convolution operation on the four-channel feature map after concatenation to capture the correlation of neighboring features: ； wherein, is a convolution kernel, denotes a convolution operation, is a convolution result; Step 4.4, perform nonlinear transformation on the convolution result using a Sigmoid function to generate a spatial attention weight map representing the importance of spatial positions: ； wherein, and denote the height and width of the image, is a Sigmoid function, is a spatial attention weight map, ranging between 0 and 1; Step 4.5, multiply the spatial attention weight map and the original entropy-weighted feature element by element to achieve adaptive weighting of the image spatial features; Step 5, output a black and white binary prediction map through the image transformation detection system, perform image region analysis, and black represents areas where changes are not recognized, while white represents areas where changes are recognized.

2. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, The multi-source image data includes remote sensing images, medical images, and natural scene image data; Uniform preprocessing operations are performed on different types of image data, including normalization and size standardization, channel alignment, and color space conversion.

3. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, In the multi-layer feature map of step 2, the low-level feature map contains texture, edge, and brightness local detail information, and the high-level feature map contains semantic structure and spatial context information.

4. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, The entropy-driven feature adaptive enhancement module in step 3 calculates the channel information entropy of the image data to quantify the information density distribution of the image features, as follows: Perform min-max normalization on each channel feature value to linearly map it to a uniform numerical interval, with the formula being: ； wherein, is the input raw feature map, is the feature map is the minimum value of each channel in the feature map, is the feature map is the maximum value of each channel in the feature map, is the smoothing term, is the normalized feature value output; The spatial dimension of the feature map is flattened into a one-dimensional vector, and the information uncertainty and complexity of the feature region are evaluated by quantifying the frequency of each feature value. The probability estimation formula is: ； wherein, denotes the probability of the th position in the feature matrix, is the value of the flattened feature matrix at , and denotes the height and width of the image; Calculate the channel information entropy value with the formula being: ； wherein, represents the first channel's entropy, is a smoothing constant term, is a scale-invariant entropy weight used to eliminate the influence of feature map size changes on entropy value calculation.

5. The entropy-driven feature enhancement based image region analysis method according to claim 4, characterized in that, The higher the entropy value, the greater the uncertainty of the corresponding channel information, the richer the detail and variation information; the lower the entropy value, the more uniform the corresponding channel feature distribution, the higher the information redundancy, and the more single the feature mode; Based on the calculated channel entropy value, the complexity of information contained in different channels is quantified, and the feature map is adaptively weighted as follows: ； wherein, is the weighted feature value; The normalized entropy value is multiplied by the original feature map as the saliency weight to strengthen the discriminative features of low-entropy flat regions and suppress the redundant background interference of high-entropy complex regions, highlighting the key regions.

6. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, In step 4.1, the max pooling is used to retain the significant response in the local region and highlight the saliency of high-entropy regions; the average pooling is used to extract global statistical features and reflect the overall information distribution. The pooling process further includes: statistical histogram pooling, Norm pooling is used to enhance the diversity of feature representation.

7. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, In step 4.3, the convolution kernel with a size of 7x7 or 3x3 or 5x5, or with a dilated convolution.

8. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, In step 4.4, the Sigmoid function is replaced by the Softmax function to normalize the weight distribution of the spatial position.

9. The entropy-driven feature augmentation based image region analysis method of claim 1, wherein, The image transformation detection system in step 5 is processed as follows: Step 5.1, input the dual-phase picture data, use the first four stages of EfficientNet-V2-S to construct a dual-branch twin network, and extract dual-phase picture features through the decoder respectively; Step 5.2, input the dual-phase picture features into the entropy-driven feature adaptive enhancement module, calculate the multi-scale feature entropy value, enhance the discriminative features related to real changes, and suppress the feature interference corresponding to false changes; Step 5.3, distinguish the object edge and morphological change features through the entropy focusing cross-space attention module; Step 5.4, input the feature map into the unified fusion and refinement module for interpolation processing: The main branch passes through the first convolutional layer, SiLU activation layer and batch normalization layer in turn, and then passes through the second convolutional layer and batch normalization layer for processing, and outputs the prediction map; The skip branch uses 1×1 convolution for channel compression and fusion; The output of the skip branch and the main branch is added to generate a multi-scale fusion feature map.

Images