Video building detection method based on object measurement

Abstract

The invention discloses an object measurement-based video building detection method, which comprises the following steps of: firstly extracting an image from a video at a fixed frame rate and preprocessing the image, and then extracting a potential object region from the image by using a binarization norm gradient object measurement algorithm to obtain region data; secondly, a deep active learning algorithm is used, and scene feature learning is carried out based on regional data; and then selecting and classifying features obtained by feature learning by using a manifold regularization algorithm to obtain a video building detection result, and defining an optimization objective function. And finally, training and testing a deep active learning algorithm, a manifold regularization algorithm and a classifier based on the optimized objective function. According to the method, the important region with high object property can be efficiently and automatically extracted from the image, so that the interference of redundant information is avoided, and the semantic precision of building detection is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of video building detection and relates to a video building detection method based on objectivity measurement. Background Art

[0002] In recent years, with the rapid development of computer vision, artificial intelligence and deep learning technology, video analysis and target detection technology have made significant progress. In the field of building detection, especially in application scenarios such as urban planning, environmental monitoring, and disaster relief, real-time detection and accurate identification of buildings have become an important demand. Video surveillance systems and drone aerial photography provide a large amount of real-time image and video data, which contain rich spatial information and structural features. However, how to quickly and accurately detect building targets from these massive data still faces many challenges.

[0003] Traditional building detection methods mainly rely on manually designed features and rules. Although they can achieve certain results in some cases, the recognition effect will be greatly reduced in complex and dynamically changing environments, especially in low resolution, blurred or occluded conditions, and cannot meet the needs of the present invention. In addition, with the application of deep learning methods, target detection methods based on convolutional neural networks have gradually become mainstream, especially in the application of static images. Significant results have been achieved. However, in the real-time processing of video data, how to balance the contradiction between computational complexity and real-time performance is still a problem that needs to be solved.

[0004] In response to the above problems, the binary normed gradient objectness measure, as an efficient candidate region generation algorithm, has demonstrated its unique advantages in quickly identifying potential target regions in images, and can provide high-quality candidate regions for subsequent precise target detection. At the same time, robust deep active learning technology can effectively reduce the amount of labeled data by actively selecting the most informative samples for training, thereby improving the generalization ability and robustness of the model in different environments. However, most of the existing building detection methods ignore the temporal characteristics of video data and the diversity of buildings at different scales and angles, resulting in low detection accuracy and processing efficiency. Therefore, how to quickly and accurately detect buildings in large-scale video data and handle various changes in complex environments while ensuring high efficiency remains a problem that needs to be solved.

[0005] Therefore, designing a video building detection method that combines binary normed gradient objectness metric and robust deep active learning can effectively improve the accuracy, efficiency and robustness of building detection. Summary of the invention

[0006] In order to make up for the deficiencies of the prior art, the present invention provides a video building detection method based on objectivity measurement to solve the problem of accuracy in identifying different buildings in a video.

[0007] The video building detection method based on object measurement includes a model training phase and a testing phase, and the specific process is as follows:

[0008] S1. First, images are extracted from the video at a fixed frame rate and preprocessed. Then, the images extracted from the video are processed using a binary normed gradient objectness metric algorithm. A large number of potential object regions are extracted from these images to obtain region data.

[0009] S2. Use a deep active learning algorithm to learn the features of the scene based on the regional data extracted in stage S1.

[0010] S3. Use the manifold regularization algorithm to select and classify the learned features, obtain the video building detection results, and define the optimization objective function;

[0011] S4. Based on the optimization objective function, the deep active learning algorithm, manifold regularization algorithm and classifier are trained and tested.

[0012] Furthermore, in step S1, a binary normed gradient object metric algorithm is used, and the binary normed gradient algorithm identifies regions with higher object characteristics by calculating the object metric values ​​of each region in the image. These regions are considered to be the parts of the video scene that are most likely to contain buildings. Through the efficiency of the binary normed gradient algorithm, high-quality target regions can be quickly extracted in scenes of multiple scales. The extracted object regions will be transmitted to the next stage for further processing by the subsequent deep learning model.

[0013] Furthermore, the robust deep active learning method used in step S2 is trained through a deep neural network, the goal of which is to select the most semantically meaningful areas and simulate the gaze trajectory of the human eye by generating a gaze transfer path (GSP). The robust deep active learning framework not only considers the spatial layout of the area in the scene, but also ensures the spatial and semantic consistency of the selected area by optimizing the reconstruction error. During the training process, the robust deep active learning method can effectively utilize limited annotated data through semi-supervised learning, and continuously improve the accuracy of the model through iterative optimization.

[0014] Furthermore, the feature selection and classification in step S3 uses a manifold learning method to compress and select the features of the image for the high-dimensional features generated during the deep learning process, and retains the key areas that best reflect the characteristics of the building. In this process, the similarity between samples is measured by constructing a graph Laplacian matrix, and a semi-supervised support vector machine classifier is used to optimize the classification effect. This process can not only effectively reduce the feature dimension, but also improve the accuracy of the classification model, especially when faced with a large amount of unlabeled data, it can still maintain efficient detection performance.

[0015] Furthermore, the specific test process in step S4 is as follows: by using the trained robust deep active learning model and manifold regularization classifier, the building area in the video is quickly and accurately identified. The detection result will be output as the coordinates and category information of the building location, and can be further used in scene analysis or automated building monitoring systems. This method is particularly suitable for building detection tasks with complex backgrounds and variable viewing angles, and can improve the accuracy of detection while ensuring high efficiency.

[0016] Compared with the prior art, the present invention has the following advantages: Through the binary normed gradient object measurement algorithm, it is possible to efficiently and automatically extract important areas with high objectivity from the image. These areas can accurately identify buildings and related areas, thereby avoiding the interference of redundant information and improving the accuracy of detection. The robust deep active learning method can automatically select the most semantically and visually important areas in the scene by simulating the gaze behavior of the human visual system. It can better capture the key areas that humans pay attention to when observing buildings, and improve the semantic accuracy of building detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the model training phase of the method of the present invention;

[0018] Figure 2 This is a flow chart of the testing phase of the method of the present invention. DETAILED DESCRIPTION

[0019] The following is a further description of a video building detection method based on objectivity measurement of the present invention in conjunction with the accompanying drawings.

[0020] A video building detection method based on objectness measurement includes a model training phase and a testing phase.

[0021] (I) Model training stage

[0022] like Figure 1 As shown in the figure, the specific process of the model training phase is as follows:

[0023] (1) First, extract images from video streams or image sequences and perform standardized preprocessing, including denoising (using Gaussian blur to remove noise in images), scale normalization (scaling images to meet the input size required by the binary normed gradient algorithm), and color space conversion (converting images from RGB color space to Lab color space for subsequent feature extraction). Gaussian blur is an image processing technique that blurs the details in an image to make it look softer. It is often used to reduce noise (random, irregular pixel value changes in an image) or create artistic effects. Lab color space is a color model based on human visual perception. It consists of brightness (L) and two color channels (a and b) and is used to describe the brightness and hue of colors.

[0024] Next, a sliding window is used to generate candidate regions of different sizes and positions. These regions are usually rectangular image regions. And in the multi-scale (different sizes or resolutions to analyze information) framework, the size of each candidate region is variable, the purpose is to capture the object features at different scales. After that, the "objectness" score of each candidate region is calculated, which measures the possibility that the region contains the target object. For each image candidate region, a convolutional neural network is applied to extract features and generate a deep feature vector for each region (the extracted image features are represented in the form of a vector). Assuming that N regions are extracted, the features of each region are represented as z 1 ,z 2 ,…,z N , the feature of the i-th region is expressed as:

[0025]

[0026] Where m is the feature dimension of the region feature.

[0027] For each candidate region R i , the binary normed gradient algorithm calculates its "objectness" metric according to the following steps:

[0028] 1.1 Calculate the objectness score: For the above features extracted, use an objectness model to calculate the objectness score of each candidate region. i , the calculation formula is as follows:

[0029]

[0030] in, represents the j-th feature (such as color, texture, edge, etc.) extracted from the i-th region, a j is the weight of feature j.

[0031] 1.2 Suppress candidate regions with high overlap: In order to avoid redundant regions, the present invention uses non-maximum suppression technology (remove boxes with high overlap and only retain boxes that are most likely to contain targets, thereby reducing redundant detection.) Suppress candidate regions with high overlap, thereby retaining the best regions. Using non-maximum suppression technology, overlap suppression is performed. First, all candidate regions are sorted in descending order according to the objectness score (i.e., the possibility that each candidate region contains a building), and the region with the highest score is at the top of the list. Then, starting with the candidate region with the highest score, each region is processed one by one. For the currently selected candidate region, its overlap with the remaining candidate regions is calculated, and the intersection-and-union ratio is used as a measure. The intersection-and-union ratio refers to the ratio of the intersection area of ​​two regions to the union area, and the value is between 0 and 1. If the intersection-and-union ratio of the current candidate region and another candidate region exceeds a preset threshold (generally set between 0.5 and 0.7, adjusted according to the specific application scenario), it is considered that the two regions are highly overlapped in space and the same target may be detected. In this case, the candidate region with a higher score is retained, and the candidate region with a lower score is removed from the candidate list. This process effectively suppresses repeated detection of the same target and avoids the situation where multiple detection boxes point to the same building area. Next, move to the next highest-scoring and unprocessed candidate region, repeat the above steps, calculate the intersection-over-union ratio between it and the remaining candidate regions, and perform the corresponding suppression operation. This process is iterated until all candidate regions have been processed. Finally, the remaining candidate regions are the detection results after non-maximum suppression. These regions have higher objectivity scores and lower overlap with each other, and can accurately represent different potential building regions in the image.

[0032] 1.3 Output potential building regions: Sort the candidate regions after overlap suppression by objectness score, and select the top N regions as potential building regions. These regions are considered to be the most likely parts to contain buildings. Next, these regions will be used as input to the subsequent robust deep active learning method for further processing and selection.

[0033] (2) In step S1, the present invention has extracted a large number of potential building areas from the video using the binary normed gradient object metric algorithm. These object areas will be used as input to enter the robust deep active learning stage for scene feature learning. The specific process of the deep active learning method is described in detail below:

[0034] 2.1 Spatial structure analysis and reconstruction: Based on the potential building areas, in order to capture the spatial relationship between these areas, the present invention establishes the following spatial area optimization model as follows:

[0035]

[0036]

[0037] Where F is the reconstruction coefficient matrix, F ij For z j Right i Contribution of reconstruction; i and z j are the feature vectors of the ith and jth regions respectively. N represents the number of extracted regions. is with z i The goal of this optimization model is to minimize the difference between each region and the weighted reconstruction of its neighbors, thereby capturing the spatial structure of the scene.

[0038] 2.2 Semantic value of evaluation region: In order to select the L regions with the most semantic significance, based on the spatial region optimization model, the present invention introduces the following objective function to maximize the semantic features:

[0039]

[0040] where h i is z extracted using a convolutional neural network i Deep feature representation. is the selected regional feature with high semantic value. i is the regional feature representation obtained by reconstruction. γ is a regularization parameter used to balance the two errors. The first term of this objective function is to ensure that the selected region is close to the original region in the feature space, and the second term is to ensure that the reconstruction error of the region is minimized, thereby maintaining spatial and semantic consistency.

[0041] 2.3 Optimization solution in matrix form: To facilitate the solution, the optimization problem of the objective function is transformed into a matrix form: define the characteristic matrix Z = [z 1 ,z 2 ,…,z N ], define the reconstruction feature matrix G = [g 1 ,g 2 ,…,g N ], introduce the indicator matrix Δ, where Δ ii =1 means the i-th region feature is selected, Δ ii =0 means not selected, and the matrix optimization objective function can be expressed as:

[0042]

[0043] Where tr represents the trace of the matrix, L1 is the Laplace matrix, which reflects the spatial relationship between regions. τ is a regularization parameter (a hyperparameter that controls the complexity of the model, which avoids overfitting of the model and improves the generalization ability of the model by adding a penalty term to the loss function). By taking the derivative of G and setting the gradient to zero, the optimal solution can be obtained:

[0044] (GZ)+τLG=0

[0045] G=(τL+Δ) -1 Z

[0046] 2.4 Select the optimal region set: Using the G obtained above, calculate the reconstruction error. The formula is as follows:

[0047]

[0048] where ||·|| F Represents the Frobenius norm. By minimizing the reconstruction error, L regions that best represent the scene features are selected.

[0049] 2.5 Build a deep model and extract features: Build a multi-layer deep learning model to extract features from the selected L regions. The structure of the model is as follows:

[0050] H=PQ R ,Q R =U R Q R-1 ,…,Q i =U i Q i-1 ,…,Q 1 =U 1 Y

[0051] Where Y is the feature matrix of the selected region. i is the transformation matrix of the i-th layer. i is the feature representation of the i-th layer. P is the potential semantic label matrix. R represents the total number of layers in the network. Through layer-by-layer linear transformation, the model can extract the deep features H of the region and capture more advanced semantic information.

[0052] (3) In step S3, the present invention needs to perform feature selection and classification on the high-dimensional features generated in the deep learning process. To this end, the present invention adopts a manifold regularization method, combined with a semi-supervised learning strategy, to effectively select and classify features. The following is the detailed process of this step:

[0053] 3.1 Constructing feature matrix and label matrix: In S2, processed deep features, namely deep GSP features, have been obtained from the robust deep active learning method. Although these features contain rich information, they are high in dimension, especially when the number of selected regions L is large, which will increase the complexity of classifier training. Therefore, the present invention needs to divide the deep features H extracted by S2 into labeled and unlabeled samples, which are specifically expressed as follows:

[0054] C s =[c 1 ,c 2 ,…,c s ]

[0055] M=[m 1 ,m 2 ,…,m s ]

[0056] C U =[c s+1 ,c s+2 ,…,c L ]

[0057] Among them C s is the feature matrix corresponding to the labeled sample, M is the label matrix, C U is the feature matrix of unlabeled samples.

[0058] 3.2 Constructing the graph structure between samples: In order to utilize the intrinsic structure and geometric relationship between samples, the present invention adopts the idea of ​​manifold learning (a nonlinear dimensionality reduction method that assumes that there is a low-dimensional manifold structure in high-dimensional data in order to reveal the essential geometric characteristics of the data, thereby retaining important local and global information. The manifold structure refers to the data in the high-dimensional space locally looking like part of the low-dimensional space) to construct the graph structure between samples. The following is the specific process:

[0059] a. Calculate the similarity matrix N

[0060] Define the similarity N between samples ij , measure sample c i and c j According to the deep feature matrix H extracted by S2, the similarity calculation method uses the Gaussian kernel function, and the specific formula is as follows:

[0061]

[0062] Among them, σ is the bandwidth parameter, which controls the decay speed of similarity. i Represents the deep feature representation of region i.

[0063] b. Calculate the degree matrix T

[0064] The degree matrix T is a diagonal matrix, and its diagonal elements are:

[0065]

[0066] Where T ii represents the degree of node i, that is, the total weight of the edges connected to it.

[0067] c. Calculate the graph Laplacian matrix B:

[0068] B=TN

[0069] Where B represents the global structural information of the graph, reflecting the geometric relationship between samples.

[0070] 3.3 Define the optimization objective function: Based on the graph structure, in the framework of semi-supervised learning, the present invention integrates feature selection and classifier training. The present invention implements this process by optimizing the objective function and outputs the video building detection results through the classifier. The specific process is as follows:

[0071] a. Define the objective function:

[0072]

[0073] Where q is the weight vector of the classifier. b is the bias of the classifier. k i is a slack variable that allows a certain degree of classification error. c is a regularization parameter that weighs the classification error and model complexity. t is a regularization parameter that controls the influence of the manifold regularization term. B is the graph Laplacian matrix that reflects the geometric structure of the sample.

[0074] b. Constraints:

[0075]

[0076] where y i is the label of the i-th labeled sample. a is the feature selection vector, a i =1 means selecting the i-th feature, and M is the preset number of feature selections.

[0077] c. Introduction of feature selection matrix

[0078] In order to perform feature selection, the present invention introduces a diagonal matrix E(a), which is expressed as follows:

[0079] E(a)=diag(a 1 ,a 2 ,…,a T )

[0080] Adjust the diagonal matrix as follows:

[0081] C(a)=cE(a)

[0082] The present invention retains only selected features.

[0083] d. Combined with feature selection, the objective function is updated as:

[0084]

[0085] 3.4 Solving the optimization problem: Since the objective function is non-convex (it does not have a single minimum point, but may have multiple local minima and a global minimum) and contains a binary variable a, it is difficult to solve it directly. Usually, the alternating optimization method is used to iteratively solve two sub-problems.

[0086] Subproblem 1: Given feature selection a, optimize the classifier parameters q, b, κ.

[0087] Fixing h, the optimization problem becomes:

[0088]

[0089] Constraints:

[0090]

[0091] This is a convex optimization problem (an optimization problem in which both the objective function and the constraints are convex). The present invention uses the interior point method in the quadratic programming algorithm (by converting the constraints into a dual problem and performing iterative optimization to gradually approach the optimal solution) to solve it.

[0092] Subproblem 2: Given classifier parameters q, b, optimize feature selection a.

[0093] After fixing q and b, the optimization problem becomes:

[0094]

[0095] Constraints:

[0096]

[0097] This is a combinatorial optimization problem. The present invention evaluates the importance of features by calculating the contribution of each feature to the objective function. For the i-th feature, the present invention calculates its importance score according to the following formula:

[0098]

[0099] Among them, q i is the i-th element of the classifier weight, is the i-th feature of sample j. According to the importance score S iSort and select the M features with the highest scores, and assign the corresponding a i Set to 1 and the rest to 0.

[0100] The optimization of the above two sub-problems is performed alternately until convergence, that is, the objective function value no longer changes significantly or the preset number of iterations is reached.

[0101] ⑤Model application and classification

[0102] Through the above optimization process, the present invention obtains the optimal classifier parameters q*, b* and feature selection vector a*.

[0103] (II) Testing Phase

[0104] like Figure 2 As shown, the present invention uses the deep feature representation obtained in the training phase and the optimized classifier to detect buildings on new video data. This process includes the following detailed steps:

[0105] (1) Video preprocessing: Extract image frames from a new video at a fixed frame rate, for example, 5 frames per second, to ensure that all key frames in the video are captured. Preprocess the extracted images.

[0106] (2) Object measurement and candidate region extraction: Slide the window on the image at different scales and step sizes to generate a series of rectangular candidate regions R i For each candidate region R i , calculate its objectivity score O i , according to the objectivity score O i Sort candidate regions in descending order. Apply non-maximum suppression algorithm to eliminate highly overlapping candidate regions and keep only the regions with the highest scores. Select the top K candidate regions with the highest scores as potential building regions.

[0107] (3) Deep feature extraction: For each candidate region R i , use the deep model obtained in the training phase to extract features and obtain the deep feature vector Where A is the feature dimension. The feature vector is normalized using L2 normalization, and the formula is as follows:

[0108]

[0109] (4) Deep active learning processing: First, use the reconstruction coefficient matrix F learned in the training phase to evaluate the spatial relationship between candidate regions. Then calculate the semantic importance of each candidate region, and select the L candidate regions with the most semantic significance by minimizing the objective function to form a set {g q1 ,g q2 ,…,g qL}.

[0110] (5) Feature selection and classification: First, use the feature selection vector a* obtained in the training phase to screen the features of the candidate regions. Then use the optimized classifier parameters q* and bias b* to classify each candidate region.

[0111] (6) Detection results: Merge adjacent candidate regions that are detected as buildings to generate a complete building detection frame. Calculate the confidence score of each detection result based on the decision value output by the classifier. Draw the building detection frame on the original image and annotate information such as confidence. Save the detection results to a specified file or database for subsequent analysis and processing.

[0112] In summary, the present invention aims at detecting buildings in videos. First, the binary normed gradient object metric algorithm is used to extract potential building areas in the image and generate candidate image areas. Then, the feature learning of these areas is performed through a robust deep active learning method, the spatial relationship and semantic value of the scene features are optimized, and the most semantically meaningful areas are selected. Then, the manifold regularization method is used for feature selection and classification, and the deep features and the optimized classifier are used to detect buildings in new videos. In the testing phase, the region is first extracted for each frame of the video, and these regions are classified by the deep feature model and classifier obtained in the training phase, and finally the detected building area and its confidence are output. Through this series of steps, the present invention effectively improves the accuracy and robustness of building detection, and can accurately identify and locate buildings in complex scenes.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A video building detection method based on object measurement, characterized in that: The following steps are involved: S1. extracting and preprocessing images from the video at a fixed frame rate, and then using a binary normed gradient objectness metric algorithm to extract potential object regions in the image to obtain region data; S2. Use deep active learning algorithm to learn scene features based on regional data; S3. Use the manifold regularization algorithm to select and classify the learned features, obtain the video building detection results, and define the optimization objective function; S4. Based on the optimization objective function, the deep active learning algorithm, manifold regularization algorithm and classifier are trained and tested.

2. The video building detection method based on objectivity measurement according to claim 1 is characterized in that: The specific process of the pretreatment is as follows: Use Gaussian blur method to remove noise, scale normalization and color space conversion in images; Candidate regions of different sizes and positions are generated by sliding windows. For each candidate region of the image, a convolutional neural network is applied to extract features and generate a feature vector for each region. Assuming that N regions are extracted, the features of each region are represented by z1, z2, …, z N , the feature of the i-th region is expressed as: m is the feature dimension of the regional feature.

3. The video building detection method based on objectivity measurement according to claim 2 is characterized in that: The specific implementation process of the binary normed gradient objectivity measurement algorithm is as follows: S1.1 Calculate the objectness score of each candidate region using the objectness model for the extracted features i : in, represents the j-th feature extracted from the i-th region, a j is the weight; S1.2 uses non-maximum suppression technology to suppress overlap: sort all candidate regions in descending order according to the objectness score, start from the candidate region with the highest objectness score, and process each region one by one: for the currently selected candidate region, calculate its overlap with the remaining candidate regions, and use the intersection-over-union ratio as a measurement standard; if the intersection-over-union ratio of the current candidate region and other candidate regions exceeds the preset threshold, retain the candidate region with the high score, and remove the candidate region with the low score from the candidate list; S1.3 sorts the candidate regions after the overlap is suppressed according to the objectness scores, and selects the first N regions as potential building regions.

4. The video building detection method based on objectivity measurement according to claim 3 is characterized in that: The specific implementation process of the deep active learning algorithm is as follows: S2.1 Based on the potential building area, the following spatial area optimization model is established: F is the reconstruction coefficient matrix, F ij For z j Right i Contribution of reconstruction; i and z j are the feature vectors of the i-th and j-th regions respectively; N is the number of regions extracted, is with z i A collection of regions that are spatially adjacent to the region to which it belongs; S2.2 Select the L regions with the most semantic meaning: Based on the spatial region optimization model, the following objective function for maximizing semantic features is introduced: where h i is z extracted using a convolutional neural network i Deep feature representation; is the selected regional feature; g i It is the regional feature representation obtained by reconstruction; γ is the regularization parameter; S2.3 Define the characteristic matrix Z = [z1,z2,…,z N ], define the reconstruction feature matrix G = [g1, g2, ..., g N ], introduce the indicator matrix Δ, where Δ ii =1 means the i-th region feature is selected, Δ ii =0 means not selected, the optimization problem of the objective function is converted into a matrix form, and the matrix optimization objective function is expressed as: Where tr represents the trace of the matrix, L1 is the Laplace matrix, and τ is the regularization parameter. By taking the derivative of G and setting the gradient to zero, we get the optimal solution: (GZ)+τLG=0 G=(τL+Δ) -1 Z S2.4 Select the optimal region set: Use G obtained above to calculate the reconstruction error. The formula is as follows: where ||·|| F represents the Frobenius norm; by minimizing the reconstruction error, L regions that best represent the scene features are selected; S2.5 Build a deep model and perform feature extraction: Build a multi-layer deep learning model and perform feature extraction on the selected L regions. The structure of the model is as follows: H=PQ R ,Q R =U R Q R-1 ,…,Q i =U i Q i-1 ,…,Q1=U1Y Where Y is the feature matrix of the selected area; U i is the transformation matrix of the i-th layer; Q i is the feature representation of the i-th layer; P is the potential semantic label matrix; R represents the total number of layers in the network; through layer-by-layer linear transformation, the model extracts the deep features H of the region.

5. The video building detection method based on objectivity measurement according to claim 4 is characterized in that: The specific implementation process of the manifold regularization algorithm is as follows: S3.1 Construct feature matrix and label matrix: Divide the deep feature H into labeled and unlabeled samples: C s =[c1,c2,…,c s ] M=[m1,m2,…,m s ] C U =[c s+1 ,c s+2 ,…,c L ] C s is the feature matrix of the labeled samples, M is the label matrix, C U is the feature matrix of unlabeled samples; S3.2 uses the idea of ​​manifold learning to construct a graph structure between samples: S3.3 is based on the graph structure and defines the optimization objective function in the framework of semi-supervised learning, and outputs the video building detection results through the classifier: S3.4 adopts alternating optimization and divides the optimization objective function into two sub-problems for iterative solution.

6. The video building detection method based on objectivity measurement according to claim 5 is characterized in that: The specific process of constructing the graph structure between samples is as follows: Define the similarity N between samples ij , measure sample c i and c j The similarity between them; According to the deep feature matrix H, the similarity calculation method uses the Gaussian kernel function to obtain the similarity matrix N; The degree matrix T is a diagonal matrix, and its diagonal elements are: Where T ii represents the degree of node i, that is, the total weight of the edges connected to it; Calculate the graph Laplacian matrix B: B = TN; B represents the global structural information of the graph, reflecting the geometric relationship between samples.

7. The video building detection method based on objectivity measurement according to claim 6 is characterized in that: The specific process of defining the optimization objective function is as follows: Define the objective function: Where q is the weight vector of the classifier, b is the bias of the classifier, and κ i is the slack variable, c and t are regularization parameters, and B is the graph Laplacian matrix; Constraints: where y i is the label of the i-th labeled sample, a is the feature selection vector, a i =1 means selecting the i-th feature, M is the preset number of feature selections; Introduce the diagonal matrix E(a), E(a) = diag(a1, a2, …, a T ); Adjust the diagonal matrix to: C(a) = cE(a); Combined with feature selection, the objective function is updated as:

8. The video building detection method based on objectivity measurement according to claim 7 is characterized in that: The specific process of step S3.4 is as follows: Subproblem 1: Given feature selection a, optimize classifier parameters q, b, κ; fixed h, the optimization problem becomes: Constraints: Use the interior point method in the quadratic programming algorithm to solve; Sub-problem 2: Given classifier parameters q, b, optimize feature selection a; after fixing q and b, the optimization problem becomes: Constraints: Evaluate the importance of features by calculating the contribution of each feature to the objective function; For the i-th feature, its importance score is calculated according to the following formula: Among them, q i is the i-th element of the classifier weight, is the i-th feature of sample j; Score according to importance i Sort and select the M features with the highest scores, and assign the corresponding a i Set to 1, and the rest to 0; The optimization of the above two sub-problems is performed alternately until convergence, that is, the objective function value no longer changes significantly or the preset number of iterations is reached.

Images