House spatial distribution feature extraction method based on intelligent deep search

Abstract

The invention provides a house space distribution feature extraction method based on intelligent deep search, and relates to the technical field of computer vision. According to the method, through feature point matching alignment, multi-modal data fusion, feature separation loss optimization, high-dimensional heat map generation and multi-magnification feature optimization, house space distribution features are accurately extracted, the analysis efficiency and precision are improved, and the method is suitable for the fields of urban planning and the like.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method for extracting house spatial distribution features based on intelligent deep search. Background Art

[0002] With the acceleration of urbanization, the extraction and analysis of housing spatial distribution features are of great significance in the fields of urban planning, real estate management, and geographic information systems. Traditional housing spatial distribution feature extraction methods usually rely on single-modal data, which is difficult to fully reflect the complexity and diversity of housing spatial distribution. In addition, when dealing with nonlinear regional features, existing methods often lack effective fusion of multimodal data, resulting in low accuracy and efficiency of feature extraction.

[0003] In recent years, with the development of deep learning and computer vision technology, multimodal data fusion and deep feature extraction technology have gradually become research hotspots. However, how to accurately align feature points in multimodal data, extract highly discriminative nonlinear regional features, and combine multi-scale image information for comprehensive analysis is still a technical problem that needs to be solved urgently. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a housing spatial distribution feature extraction method based on intelligent deep search, which significantly improves the accuracy and efficiency of housing spatial distribution feature extraction through technical means such as multimodal data fusion, feature separation loss optimization, high-dimensional deep heat map generation, neighborhood path feature fusion, and multi-rate image feature optimization. It has the advantages of comprehensiveness, accuracy, high efficiency and strong adaptability, and provides strong technical support for the analysis and application of housing spatiotemporal distribution characteristics.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A housing spatial distribution feature extraction method based on intelligent deep search, comprising:

[0007] Collecting original geographic space-time images of the same administrative area, preprocessing the original geographic space-time images to obtain preprocessed data, and processing the original geographic space-time images through a nonlinear transformation algorithm to obtain a nonlinear approximation image;

[0008] Extracting feature points from the preprocessed data and the nonlinear approximation image using a scale-invariant feature transformation algorithm, and matching and aligning the feature points to ensure feature consistency of the images of the two modalities, thereby obtaining an aligned image;

[0009] Extracting feature vectors from the aligned images using a neural network, calculating feature similarity between nonlinear regions and non-nonlinear regions, and optimizing feature extraction process using feature separation loss based on the feature similarity to obtain nonlinear region features;

[0010] Determine a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear region feature and the nonlinear approximation image;

[0011] Based on high-dimensional deep heatmaps, a neural network is used to extract the deep features of administrative region paths, and the enhanced features of neighborhood paths are calculated through neighborhood administrative region path feature fusion learning.

[0012] Based on the enhanced features of the neighborhood path, the feature vector and the feature point are input into a preset model, a table lookup process of computer search is simulated, different weights are assigned to image features of different magnifications, and optimized weight features are obtained;

[0013] Based on the optimized weight features, a deep convolutional network is used to classify the features, generate a weighted category activation map, and extract the spatiotemporal distribution area of ​​the house of interest according to the weighted category activation map.

[0014] Preferably, the process of preprocessing the original geographic spatiotemporal image includes denoising and contrast enhancement.

[0015] Preferably, the calculation formula of the feature separation loss is:

[0016]

[0017] Among them, F non-lesion 、F non-lesion and F ave-lesion They represent the non-linear region, non-linear region and average features of the non-linear region in the preprocessed data or non-linear approximation image; Sim(.) represents the cosine similarity calculation; Loss FS It is the feature separation loss, which means that the features of the non-linear region are as far away from the features of the non-linear region as possible, so as to achieve high discrimination extraction of non-linear features.

[0018] Preferably, determining a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear region feature and the nonlinear approximation image comprises:

[0019] In the nonlinear approximation mode, the nonlinear region is meshed;

[0020] Use high-power nonlinear approximation magnification mode to photograph the grid in the nonlinear area and generate small grid heat map blocks;

[0021] Switch the small grid heat map block to non-magnification mode, reduce the size to 1 / 50 of the original, and obtain the high-dimensional depth heat map block in non-magnification mode;

[0022] The statistical information is determined by pixel counting, connected domain analysis and mask analysis; the statistical information includes information on the area, proportion and spatial distribution of different high-dimensional depths.

[0023] Preferably, the calculation formula of the enhanced feature of the neighborhood path is:

[0024]

[0025] in, represents the characteristic Euclidean distance between administrative region path j and administrative region path i; F j NBI represents the depth feature of administrative region path j; F i NBI It represents the enhanced features after the fusion of the path features of the neighborhood administrative areas; p is the custom coefficient.

[0026] Preferably, the custom coefficient p=0.25.

[0027] Preferably, the preset model is a CNN optimization model using a joint regularization loss.

[0028] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0029] The present invention provides a method for extracting house spatial distribution features based on intelligent deep search, comprising: collecting original geographic spatiotemporal images of the same administrative area, preprocessing the original geographic spatiotemporal images to obtain preprocessed data, and processing the original geographic spatiotemporal images through a nonlinear transformation algorithm to obtain a nonlinear approximation image; extracting feature points in the preprocessed data and the nonlinear approximation image using a scale-invariant feature transformation algorithm, and matching and aligning the feature points to ensure feature consistency of images of two modalities to obtain an aligned image; extracting feature vectors from the aligned image using a neural network, and calculating feature similarity between nonlinear regions and non-linear regions, and optimizing the feature extraction process based on the feature similarity using feature separation loss. Obtain nonlinear regional features; determine the high-dimensional deep heat map block and the statistical information of the high-dimensional deep heat map block according to the nonlinear regional features and the nonlinear approximation image; based on the high-dimensional deep heat map block, use a neural network to extract the deep features of the administrative region path, and calculate the enhanced features of the neighborhood path through the neighborhood administrative region path feature fusion learning method; based on the enhanced features of the neighborhood path, the feature vector and the feature point are input into a preset model, the table lookup process of computer search is simulated, and different weights are given to different magnification image features to obtain optimized weight features; based on the optimized weight features, a deep convolutional network is used to classify the features, generate a weighted category activation map, and extract the target's interested house spatiotemporal distribution area according to the weighted category activation map. The present invention significantly improves the accuracy and efficiency of house spatial distribution feature extraction through technical means such as multimodal data fusion, feature separation loss optimization, high-dimensional deep heat map block generation, neighborhood path feature fusion, and multi-magnification image feature optimization, and has the advantages of comprehensiveness, accuracy, efficiency and strong adaptability, providing strong technical support for the analysis and application of house spatiotemporal distribution features. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 A flow chart of a method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] The purpose of the present invention is to provide a housing spatial distribution feature extraction method based on intelligent deep search, which significantly improves the accuracy and efficiency of housing spatial distribution feature extraction through technical means such as multimodal data fusion, feature separation loss optimization, high-dimensional deep heat map generation, neighborhood path feature fusion, and multi-rate image feature optimization. It has the advantages of comprehensiveness, accuracy, high efficiency and strong adaptability, and provides strong technical support for the analysis and application of housing spatiotemporal distribution characteristics.

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a housing space distribution feature extraction method based on intelligent deep search, comprising:

[0036] Step 100: collecting original geographic space-time images of the same administrative area, preprocessing the original geographic space-time images to obtain preprocessed data, and processing the original geographic space-time images through a nonlinear transformation algorithm to obtain a nonlinear approximation image;

[0037] Step 200: extracting feature points from the preprocessed data and the nonlinear approximation image using a scale-invariant feature transformation algorithm, and matching and aligning the feature points to ensure feature consistency of the images of the two modalities, thereby obtaining an aligned image;

[0038] Step 300: extracting feature vectors from the aligned images using a neural network, and calculating feature similarity between the nonlinear region and the non-linear region, and optimizing the feature extraction process based on the feature similarity using feature separation loss to obtain nonlinear region features;

[0039] Step 400: determining a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear regional features and the nonlinear approximation image;

[0040] Step 500: Based on the high-dimensional deep heat map, a neural network is used to extract the deep features of the administrative region path, and the enhanced features of the neighborhood path are calculated by a neighborhood administrative region path feature fusion learning method;

[0041] Step 600: inputting the enhanced features, feature vectors and feature points based on the neighborhood path into the preset model, simulating the table lookup process of computer search, assigning different weights to the features of images with different magnifications, and obtaining optimized weight features;

[0042] Step 700: Based on the optimized weight features, use a deep convolutional network to classify the features, generate a weighted category activation map, and extract the spatiotemporal distribution area of ​​the target house of interest based on the weighted category activation map.

[0043] Preferably, the process of preprocessing the original geographic spatiotemporal image includes denoising and contrast enhancement.

[0044] For the house search data in the same administrative area, the preprocessed data (PD) and nonlinear approximation (NQ) images of the area are collected as the initial input of this algorithm. The multimodal feature range prediction algorithm in 1) is used to predict the nonlinear area. In the nonlinear approximation (NQ) mode, the nonlinear area is gridded, and the small grid is photographed using the high-power nonlinear approximation (NQ) magnification mode. Through the 2) algorithm, the image analysis results are presented in the form of a heat map. The obtained small grid heat map block is switched to the non-magnification mode, and the size is reduced to 1 / 50 of the original size. The high-dimensional depth heat map block in the non-magnification mode is obtained and reorganized. By counting pixels, the area and proportion of different high-dimensional degrees in the feature are statistically analyzed; the spatial distribution of different high-dimensional types is analyzed through methods such as connected domain analysis and mask analysis, etc.

[0045] First, a neural network is used to extract feature information from the preprocessed data (PD) and nonlinear approximation (NQ) of the same area. These feature vectors contain texture, shape, intensity and other information in the original image data, as well as the high-level features learned by the neural network itself.

[0046] Due to the differences in resolution and angle between the preprocessed data (PD) and the nonlinear approximation (NQ) images, it is necessary to align the two modal images to achieve accurate nonlinear detection and table lookup. It is proposed to extract feature points from the preprocessed data (PD) and the nonlinear approximation (NQ) images and use the SIFT (Scale Invariant Feature Transform) algorithm to match them to find the correspondence between the two images. Secondly, the aligned features are fused and the fused key features are inferred to find the nonlinear regions in the preprocessed data (PD) and the nonlinear approximation (NQ).

[0047] Preferably, the calculation formula of the feature separation loss is:

[0048]

[0049] Among them, F non-lesion 、F non-lesion and F ave-lesion They represent the non-linear region, non-linear region and average features of the non-linear region in the preprocessed data or non-linear approximation image; Sim(.) represents the cosine similarity calculation; Loss FS It is the feature separation loss, which means that the features of the non-linear region are as far away from the features of the non-linear region as possible, so as to achieve high discrimination extraction of non-linear features.

[0050] Different from the existing methods, considering the heterogeneity and similarity of preprocessed data (PD) and nonlinear approximation (NQ) images, based on the important prior of similarity of non-linear region features, the project intends to further achieve more accurate feature alignment and nonlinear region prediction by calculating the similarity of features of nonlinear regions in preprocessed data (PD) and nonlinear approximation (NQ) images, that is, using the following feature separation loss:

[0051]

[0052] Among them, F non-lesion, F non-lesion, F ave-lesion They represent the non-linear region, non-linear region and average features of the non-linear region in the pre-processed data (PD) or non-linear approximation (NQ) image respectively; Sim(.) represents the cosine similarity calculation. From formula (1), we can see that Loss FS , namely feature separation loss, means that the features of the non-linear region are kept as far away from the features of the non-linear region as possible, thereby achieving high discrimination of non-linear features.

[0053] Subsequently, the nonlinear area is gridded in the nonlinear approximation (NQ) mode, and the small grid is photographed using the high-power nonlinear approximation (NQ) magnification mode. The administrative region path area in this mode is obtained through the administrative region path prediction algorithm, and then the neighborhood administrative region path feature fusion learning method is used to obtain the high discriminant features of the administrative region path.

[0054] Preferably, determining a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear region feature and the nonlinear approximation image comprises:

[0055] In the nonlinear approximation mode, the nonlinear region is meshed;

[0056] Use high-power nonlinear approximation magnification mode to photograph the grid in the nonlinear area and generate small grid heat map blocks;

[0057] Switch the small grid heat map block to non-magnification mode, reduce the size to 1 / 50 of the original, and obtain the high-dimensional depth heat map block in non-magnification mode;

[0058] The statistical information is determined by pixel counting, connected domain analysis and mask analysis; the statistical information includes information on the area, proportion and spatial distribution of different high-dimensional depths.

[0059] Preferably, the calculation formula of the enhanced feature of the neighborhood path is:

[0060]

[0061] in, represents the characteristic Euclidean distance between administrative region path j and administrative region path i; F j NBI represents the depth feature of administrative region path j; F i NBI It represents the enhanced features after the fusion of the path features of the neighborhood administrative area; p is a custom coefficient. p is a custom coefficient. Preliminary experiments found that when p = 0.25, relatively better performance can be obtained.

[0062] Preprocess the geographic spatiotemporal image, including removing image noise and enhancing image contrast. Train a neural network to extract administrative region path features, and predict administrative region path types through classification algorithms. Use morphological operations (such as corrosion and expansion) to better display the location and area of ​​administrative region paths, correct the paths of adjacent administrative regions, and draw a heat map of administrative region path areas based on the predicted confidence.

[0063] After a series of operations such as small grid cutting, enlargement, and heat map generation, the color value of each pixel in the small grid heat map block is closely related to the high-dimensional depth of the corresponding area. In order to further analyze the high-dimensional situation of the feature area, it is necessary to switch the small grid heat map block to non-enlarged mode and reduce the size to 1 / 50 of the original size to obtain the high-dimensional depth heat map block in non-enlarged mode and reassemble it.

[0064] Finally, according to the actual situation, high-dimensional depth can be divided into different levels. For each level of high-dimensional depth, the area and proportion of different high-dimensional degrees in the feature are counted by counting pixels. By using common image segmentation algorithms in the field of computer vision, such as threshold segmentation, region growing and other methods, pixels of different high-dimensional depths can be automatically classified. The pixels of different high-dimensional depths in the heat map are divided into several connected domains. Through the connected domain analysis, the area, perimeter, center of gravity and other information of each connected domain can be obtained, which helps data retrieval to more comprehensively understand the spatial distribution of different high-dimensional types. Mask analysis helps data retrieval to more accurately understand the high-dimensional type and degree of features, and guide the formulation and adjustment of housing spatiotemporal data search plans.

[0065] First, according to the prediction algorithm in the previous step, the key regional features in the house spatiotemporal data search image data are constructed, and irrelevant redundant information and noise are further removed. A specific regularizer is proposed for digital prior knowledge, and the corresponding optimization problem is trained using CNN with joint regularization loss. In order to integrate the feature information of images at different magnifications into the learning process of the algorithm, the model input integrates the image information of multiple magnifications to comprehensively pay attention to the geographical distribution at high magnifications, medium and low magnifications, and fully consider the specificity and commonality of data samples to learn the key features of the data. At the same time, the table lookup process of computer search is simulated, and different attention weights are given to the image features of different magnifications, fully considering the data features of each magnification image. The above regularizer is learned and optimized through a deep convolutional network, so that the model can learn the data prior better and faster.

[0066] After the parametric model is learned, the features at multiple scales are fused and calculated according to the identified target category to obtain a weighted category activation map, which is finally processed to obtain the spatiotemporal distribution area of ​​the target house of interest.

[0067] The beneficial effects of the present invention are as follows:

[0068] (1) The present invention generates two modal data by preprocessing (PD) and nonlinear transformation (NQ) the original geographic spatiotemporal image, making full use of the heterogeneity and complementarity of the two types of data; the scale-invariant feature transform (SIFT) algorithm is used to match and align the feature points of the two modal images, ensuring the consistency of the features and laying the foundation for subsequent feature extraction and analysis.

[0069] (2) The present invention optimizes the feature extraction process through feature separation loss, which significantly improves the distinction between nonlinear region features and non-linear region features, thereby realizing accurate detection and extraction of nonlinear regions; by gridding the nonlinear region, high-dimensional depth heat maps are generated, and combined with pixel counting, connected domain analysis, mask analysis and other methods, the area, proportion and spatial distribution of different high-dimensional depths are statistically analyzed, providing comprehensive data support for the analysis of housing spatial distribution characteristics.

[0070] (3) The present invention adopts a neighborhood administrative region path feature fusion learning method and uses formula (2) to calculate the enhanced features of the neighborhood path, fully considering the similarities and differences between the paths and significantly improving the ability to discriminate the administrative region path features. It simulates the table lookup process of computer search, integrates the image feature information at different magnifications into the model learning process, conducts a comprehensive analysis of the specificity and commonality of high-magnification, medium-magnification and low-magnification images, and assigns different weights to optimize the effect of feature extraction.

[0071] (3) Based on the optimized weight features, the present invention uses a deep convolutional network to classify the features, generate a weighted category activation map, and extract the spatiotemporal distribution areas of the houses of interest, thereby ensuring the accuracy and reliability of the final results. Through the introduction of neural networks and deep learning technology, the entire process from feature extraction to target area identification is automatically completed, reducing manual intervention and improving processing efficiency.

[0072] (4) The method of the present invention fully considers the complexity and diversity of the spatial distribution of houses, can adapt to geographic spatiotemporal images of different resolutions and modalities, and is suitable for various application scenarios such as urban planning, real estate management, and geographic information systems.

[0073] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0074] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A housing spatial distribution feature extraction method based on intelligent deep search, characterized in that: include: Collecting original geographic space-time images of the same administrative area, preprocessing the original geographic space-time images to obtain preprocessed data, and processing the original geographic space-time images through a nonlinear transformation algorithm to obtain a nonlinear approximation image; Extracting feature points from the preprocessed data and the nonlinear approximation image using a scale-invariant feature transformation algorithm, and matching and aligning the feature points to ensure feature consistency of the images of the two modalities, thereby obtaining an aligned image; Extracting feature vectors from the aligned images using a neural network, calculating feature similarity between nonlinear regions and non-nonlinear regions, and optimizing feature extraction process using feature separation loss based on the feature similarity to obtain nonlinear region features; Determine a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear region feature and the nonlinear approximation image; Based on high-dimensional deep heatmaps, a neural network is used to extract the deep features of administrative region paths, and the enhanced features of neighborhood paths are calculated through neighborhood administrative region path feature fusion learning. Based on the enhanced features of the neighborhood path, the feature vector and the feature point are input into a preset model, a table lookup process of computer search is simulated, different weights are assigned to image features of different magnifications, and optimized weight features are obtained; Based on the optimized weight features, a deep convolutional network is used to classify the features, generate a weighted category activation map, and extract the spatiotemporal distribution area of ​​the house of interest according to the weighted category activation map.

2. The housing space distribution feature extraction method based on intelligent deep search according to claim 1 is characterized in that: The process of preprocessing the original geographic spatiotemporal image includes denoising and contrast enhancement.

3. The housing space distribution feature extraction method based on intelligent deep search according to claim 1 is characterized in that: The calculation formula of the feature separation loss is: Among them, F non-lesion 、F non-lesion and F ave-lesion They represent the non-linear region, non-linear region and average features of the non-linear region in the preprocessed data or non-linear approximation image; Sim(.) represents the cosine similarity calculation; Loss FS It is the feature separation loss, which means that the features of the non-linear region are as far away from the features of the non-linear region as possible, so as to achieve high discrimination extraction of non-linear features.

4. The housing space distribution feature extraction method based on intelligent deep search according to claim 1 is characterized in that: Determining a high-dimensional depth heat map block and statistical information of the high-dimensional depth heat map block according to the nonlinear region feature and the nonlinear approximation image includes: In the nonlinear approximation mode, the nonlinear region is meshed; Use high-power nonlinear approximation magnification mode to photograph the grid in the nonlinear area and generate small grid heat map blocks; Switch the small grid heat map block to non-magnification mode, reduce the size to 1 / 50 of the original size, and obtain the high-dimensional depth heat map block in non-magnification mode; The statistical information is determined by pixel counting, connected domain analysis and mask analysis; the statistical information includes information on the area, proportion and spatial distribution of different high-dimensional depths.

5. The housing space distribution feature extraction method based on intelligent deep search according to claim 1 is characterized in that: The calculation formula of the enhanced feature of the neighborhood path is: in, represents the characteristic Euclidean distance between administrative region path j and administrative region path i; F j NBI represents the depth feature of administrative region path j; F i NBI It represents the enhanced features after the fusion of the path features of the neighborhood administrative areas; p is the custom coefficient.

6. The housing space distribution feature extraction method based on intelligent deep search according to claim 5 is characterized in that: The custom coefficient p=0.

25.

7. The housing space distribution feature extraction method based on intelligent deep search according to claim 1 is characterized in that: The preset model is a CNN optimization model using joint regularization loss.

Images