Remote sensing image point-like independent house detection method and system based on YOLOX network

Abstract

The present application relates to the technical field of remote sensing image target recognition, and particularly relates to a remote sensing image point-like independent house detection method and system based on a YOLOX network, which labels and annotates rural or mountainous residential area monomer building houses in a collected point-like independent house image dataset, and expands the point-like independent house image sample set through a sample enhancement method; a point-like independent house target detection model is constructed based on the YOLOX network, and a target loss function for model training is established; the target detection model is trained using the sample set and the target loss function, and a trained target detection model is obtained; remote sensing image data to be detected is input into the trained target detection model, and the trained target detection model is used to obtain remote sensing image monomer building house targets. The present application can improve multi-scale feature recognition and fine-grained feature expression capability, and has certain universality in remote sensing image detection of point-like independent house and other scattered and small-scale ground object targets.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image target recognition, and particularly relates to a remote sensing image point-like independent house detection method and system based on a YOLOX network. BACKGROUND

[0002] Target detection of aerospace remote sensing images has important application value in topographic mapping, city planning and the like. Remote sensing images can clearly see mountains, lakes, houses and streets, and can objectively, dynamically and truly reflect the space-time information of production, life and ecology, and have become basic data resources in topographic mapping, city planning, earthquake disaster risk estimation and the like. House recognition based on remote sensing images can quickly find out the distribution and quantity of houses, and can provide fast and reliable data sources for topographic mapping, city planning, earthquake disaster risk estimation and the like in a macroscopic manner. Unlike traditional natural images, a large amount of target information in remote sensing images is relatively fragmented, and the background is more complex, thereby causing more interference information on a feature map, and individual point-like independent houses and the like are scattered, which aggravates the difficulty of remote sensing image target detection. The effect of a classical traditional algorithm on target segmentation and classification is poor, and the accuracy is low and easy to miss detection under complex conditions.

[0003] Deep learning technology can be used to solve the problem of remote sensing image target detection, and the detection effect on a plurality of open source target data sets has far exceeded that of a traditional algorithm, but there are still some problems in the practical application of a convolutional neural network algorithm, such as insufficient strengthening extraction of weak and small features in a high-level feature map in multi-scale feature extraction, and the target detail feature and semantic function do not play an obvious role; in dense target detection, the number and proportion of overlapping targets cannot be accurately analyzed and regressed; model network over-saturation causes gradient disappearance and poor generalization of over-fitting, which makes it difficult for the detection accuracy of the deep learning algorithm to meet the actual needs of remote sensing image detection of point-like independent houses and the like scattered distribution of ground object targets, and the generalization ability and recognition accuracy for different types of detection objects still need to be improved. SUMMARY

[0004] Therefore, the present application provides a remote sensing image point-like independent house detection method and system based on a YOLOX network, which improves multi-scale feature recognition and fine-grained feature expression capability, and has certain universality in remote sensing image detection of point-like independent houses and the like scattered distribution and small scale ground object targets.

[0005] According to the design scheme provided by the present application, a remote sensing image point-like independent house detection method based on a YOLOX network is provided, which comprises:

[0006] Labeling rural or mountainous residential areas of single building houses in the collected point-like independent house image data set, and expanding the data set after label annotation through a sample enhancement method to obtain a point-like independent house image sample set;

[0007] A point-shaped independent house target detection model is constructed based on a YOLOX network, and a target loss function for model training is established, wherein the point-shaped independent house target detection model takes a residual network and a Focus module as a backbone network, takes a PANet network as a feature pyramid extraction network, and takes a DecoupledHead as a YOLOHead detection network, and the target loss function uses a Swish activation function;

[0008] The point-shaped independent house target detection model is trained and optimized based on the established target loss function and the point-shaped independent house image sample set, and a trained point-shaped independent house target detection model is obtained.

[0009] The remote sensing image data to be detected is input into the trained point-shaped independent house target detection model, and the remote sensing image single building house target is obtained by using the trained point-shaped independent house target detection model.

[0010] As the point-shaped independent house detection method based on the YOLOX network, further, the data set after label annotation is expanded through a sample enhancement method, which includes:

[0011] First, the single building house label corresponding to the data set is traversed, and the single building house target of the label category attribute is valued;

[0012] Then, each image in the data set and its corresponding label information are read, the image data is converted into an RGB image, and the sample image range to be enhanced is framed by reading the image width and height and the label target width and height.

[0013] Then, different image data are mixed in proportion, and the framed sample image is expanded through a sample enhancement method, wherein the sample enhancement method includes geometric transformation, color transformation and discrete interpolation sampling.

[0014] As the point-shaped independent house detection method based on the YOLOX network, further, the data set after label annotation is expanded through a sample enhancement method, which also includes: the feature of the resident area of the image in the data set is enhanced to enhance the visual saliency of the region of interest.

[0015] As the point-shaped independent house detection method based on the YOLOX network, further, the feature of the resident area of the image in the data set is enhanced to enhance the visual saliency of the region of interest, which includes:

[0016] First, the texture feature component of the resident area information of interest in the data set image is extracted through edge detection, and the texture feature component is converted into a saliency heat map by using a Gaussian function.

[0017] Then, according to the estimation of the significant heat map of the residential area, the estimated significant heat map is fused into the original image at an adaptive parameter multiple, and the residential area of interest in the original image is hierarchically enhanced and the effect of the non-interest area is weakened.

[0018] As the point-shaped independent house detection method based on the YOLOX network of the present application, further, the residual network and the Focus module are used as the backbone network in the point-shaped independent house target detection model, the PANet network is used as the feature pyramid extraction network, and the DecoupledHead is used as the YOLOHead detection network, which includes: the PANet network in the feature pyramid extraction network, which uses the adaptive spatial feature fusion network and relies on the multiplication and addition of the weight parameters of each layer and the extracted features for feature mapping fusion, filters the features of other layers and retains the key local area information, so as to extract the fine-grained features of the input image; and the local cross-channel interaction ECA attention module is embedded in the residual network and the PANet network, and the local cross-channel interaction ECA attention module is used to strengthen the feature information of the residential area.

[0019] As the point-shaped independent house detection method based on the YOLOX network of the present application, further, the process of using the adaptive spatial feature fusion network and relying on the multiplication and addition of the weight parameters of each layer and the extracted features for feature mapping fusion is represented as: Wherein, represents the feature map obtained after the feature of the pixel point (i, j) is transformed from the 1st layer, the 2nd layer and the 3rd layer to the lth layer respectively, is the corresponding parameter weight learned by the lth layer feature map of the pixel point (i, j), and is the new feature map obtained after the lth layer adaptive spatial feature mapping of the pixel point (i, j).

[0020] As the point-shaped independent house detection method based on the YOLOX network of the present application, further, the local cross-channel interaction ECA attention module is used to strengthen the feature information of the residential area, which includes: using the parameter matrix W k The channel attention is learned to make each channel completely independent; and the grouping convolution is used to group the high-dimensional and low-dimensional channels to reduce the complexity of the common learning parameters in the model training process, and the convolution kernel size and the channel dimension are adaptively optimized to cover the local cross-channel interaction range, wherein the process of adaptively optimizing the local cross-channel interaction coverage is represented as k represents the convolution kernel size, C represents the channel dimension size, odd is the nearest odd number, and γ and b are preset adaptive proportion constants.

[0021] As the YOLOX network-based remote sensing image point-like independent house detection method of the application, further, the target loss function uses a Swish activation function, the Swish activation function is represented as S(x) = x sigma (G(x)), wherein x is input data, G(x) is a convolutional layer output, and sigma (·) represents a Sigmoid function.

[0022] Further, the application also provides a YOLOX network-based remote sensing image point-like independent house detection system, comprising a sample set establishing module, a detection model constructing module and a target detection module, wherein,

[0023] The sample set establishing module is used for labeling the rural or mountainous residential area monomer building houses in the collected point-like independent house image data set, and expanding the labeled data set through a sample enhancement method to obtain a point-like independent house image sample set;

[0024] The detection model constructing module is used for constructing a point-like independent house target detection model based on a YOLOX network, and establishing a target loss function for model training, wherein the point-like independent house target detection model takes a residual network and a Focus module as a backbone network, takes a PANet network as a feature pyramid extraction network, and takes a decoupled head DecoupledHead as a YOLOHead detection network, and the target loss function uses a Swish activation function; the point-like independent house target detection model is trained and optimized by using the point-like independent house image sample set and based on the established target loss function to obtain a trained point-like independent house target detection model;

[0025] The target detection module is used for inputting the remote sensing image data to be detected into the trained point-like independent house target detection model, and acquiring a remote sensing image monomer building house target by using the trained point-like independent house target detection model.

[0026] The application has the following beneficial effects:

[0027] The present application is aimed at the extraction of point-shaped houses in large-size remote sensing images. Considering the scattered distribution of point-shaped houses and the unclear semantic structural features on smaller scale images, a target detection model suitable for rural or mountainous point-shaped single building houses is constructed based on the YOLOX network, which reduces the parameter amount while maintaining the robustness and generalization performance of the model. High-quality sample data for model training is obtained by label annotation and sample enhancement using public remote sensing image datasets, which enables the model to fully learn the rural or mountainous residential feature information, effectively extract the robust and invariant features of point-shaped independent houses, reduce the influence of other interference factors, and improve the detection stability. Further experimental data show that the scheme has superior detection performance in algorithm design and complex background information, and can be applied to remote sensing image point-shaped independent house extraction and house distribution detection, providing strong support for surveying and mapping production. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The figure is a schematic diagram of the remote sensing image point-shaped independent house detection process based on the YOLOX network in the embodiment.

[0029] Figure 2 The figure is a schematic diagram of the algorithm technical route in the embodiment.

[0030] Figure 3 The figure is a schematic diagram of the target detection model network structure of the algorithm in the embodiment.

[0031] Figure 4 The figure is an example of point-shaped independent house dataset annotation in the embodiment.

[0032] Figure 5 The figure is a schematic diagram of the sample enhancement technical route in the embodiment.

[0033] Figure 6 The figure is a schematic diagram of the texture detection results of the original image by different algorithms in the embodiment.

[0034] Figure 7 The figure is a schematic diagram of the saliency heat map acquisition process in the embodiment.

[0035] Figure 8 The figure is a schematic diagram of the experimental results of different algorithms in the embodiment.

[0036] Figure 9 The figure is a schematic diagram of the comparison of detection and prediction effects before and after the optimization of dataset labels and YOLOX network in the embodiment.

[0037] Figure 10 The figure is a schematic diagram of the comparison of detection and prediction effects before and after the optimization of dataset labels and YOLOX network in the embodiment.

[0038] Figure 11 The figure is a schematic diagram of the incomplete detected house in the embodiment.

[0039] Figure 12 Fig. 1 is a schematic diagram of overlapping missing houses in the embodiments. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below with reference to the drawings and technical scheme.

[0041] For the problems of remote sensing image point independent house extraction and house distribution detection in surveying and mapping production, in the embodiments of the present application, referring to the figure Figure 1 As shown in the figure, a YOLOX network-based remote sensing image point independent house detection method is provided, which comprises:

[0042] S101, label annotation is performed on rural or mountainous residential area single building houses in the collected point independent house image dataset, and the dataset after label annotation is expanded through sample enhancement, to obtain a point independent house image sample set;

[0043] S102, a point independent house target detection model is constructed based on the YOLOX network, and a target loss function for model training is established, wherein the point independent house target detection model takes a residual network and a Focus module as a backbone network, takes a PANet network as a feature pyramid extraction network, and takes a decoupled head DecoupledHead as a YOLOHead detection network, and the target loss function uses a Swish activation function;

[0044] S103, the point independent house target detection model is trained and optimized based on the established target loss function and using the point independent house image sample set, to obtain a trained point independent house target detection model;

[0045] S104, the remote sensing image data to be detected is input into the trained point independent house target detection model, and the remote sensing image single building house target is obtained by using the trained point independent house target detection model.

[0046] YOLOX is a new single-stage target detection network proposed by PPL in August 2021. It follows the powerful extraction ability and deep network characteristics of the YOLO series, and innovatively uses Focus backbone network, DecoupledHead prediction branch decoupling, Mosaic data enhancement and AnchorFree mechanism. YOLOX network is representative and outstanding in the YOLO series, and is selected to improve the problem of remote sensing image target detection. The YOLOX network structure can be divided into four plates, namely the input end plate of strong augmentation data enhancement, the Backbone backbone network plate based on Darknet, the Neck plate based on FPN, and the Prediction plate based on DecoupledHead. In the embodiments of the present case, as shown in Figure 2 Residual network and Focus module are used as the backbone network, PANet network is used as the feature pyramid extraction network, DecoupledHead is used as the YOLOHead detection network, Swish activation function is used to construct the target loss function, public remote sensing image dataset is used to obtain high-quality sample data for model training through label annotation and sample enhancement, and the target detection model is trained based on the target loss function. While reducing the amount of model parameters, the robustness and generalization performance of the model can be maintained, and it can be applied to the extraction and detection of scattered point-like independent houses in remote sensing images.

[0047] As a preferred embodiment, further, the dataset after label annotation is expanded through sample enhancement, which can be designed to include the following contents:

[0048] First, traverse the single building house label corresponding to the dataset and assign the single building house target of the label class attribute;

[0049] Next, read each image and its corresponding annotation label information in the dataset, convert the image data to an RGB image, and frame the sample image range to be enhanced by reading the image width and height and the annotation target width and height;

[0050] Then, mix different image data in proportion, and expand the framed sample image through sample enhancement, wherein the sample enhancement includes geometric transformation, color transformation and discrete interpolation sampling.

[0051] Further, the method further comprises: performing feature enhancement on the resident area of the image in the data set to enhance the visual saliency of the region of interest, extracting a texture feature component of the resident area information of the region of interest in the data set image through edge detection, converting the texture feature component into a saliency heat map by using a Gaussian function; and estimating the resident area according to the saliency heat map, and fusing the estimated saliency heat map to the original image at an adaptive parameter multiple, to enhance the resident area of interest in the original image and weaken the effect of the non-interest region.

[0052] In the point-shaped independent house target detection model, a residual network and a Focus module are used as a backbone network, a PANet network is used as a feature pyramid extraction network, and a DecoupledHead is used as a YOLOHead detection network. Specifically, the PANet network in the feature pyramid extraction network uses an adaptive spatial feature fusion network, and relies on the multiplication and addition of the weight parameters of each layer and the extracted features for feature mapping fusion, filters the features of other layers and retains the key local area information, so as to extract the fine-grained features of the input image. A local cross-channel interaction ECA attention module is embedded in the residual network and the PANet network, and the local cross-channel interaction ECA attention module is used to strengthen the resident area feature information.

[0053] Referring to Figure 3 As shown in the figure, the backbone network CSP (Cross Stage Partial) Darknet is based on a residual network Residual and a Focus module, and constructs an FPN (Feature Pyramid Networks) feature pyramid extraction module based on a PANet network and a YOLOHead detection module based on a DecoupledHead. An attention mechanism based on the non-reduced dimension ECA local cross-channel interaction strategy is added to the feature map of the backbone network and the output part of the sampling layer, which effectively avoids the influence of dimension reduction on the learning effect of channel attention, and strengthens the extraction of weak and small features in the high-level feature map. Secondly, the ASFF module is fused in the FPN feature pyramid extraction network, which expands the perception domain and excavates high-level semantic features to balance the fine granularity in the bottom layer features. The global network activation function is adjusted, and the Swish activation function is used in the CSPL (Cross Stage Partial Layer) and CBS (Convolution Batch Normalization) layers to stabilize the gradient curve and speed up the convergence.

[0054] YOLOX uses the way of PANet for feature fusion, which cannot fully tap the multi-scale feature information. Therefore, in the embodiment of the case, the ASFF structure is introduced, relying on the mapping fusion of multiplying and adding the weight parameters of each layer with the features, efficiently filtering the features of other layers, suppressing negative information, and retaining important local area information, so that the extracted fine-grained features are more refined and hierarchical. It can be described as follows:

[0055] X 1,2,3→l For the feature map from the l-th layer, multiply and add the feature maps of different layers and the corresponding learnable weights l ,β l ,γ l to obtain the fused features, where is the parameter learned by the l-th layer feature map, The expression (1) is as follows ( Similarly, we can get):

[0056]

[0057] Among them, X 1,2,3→l is obtained through 1x1 convolution.

[0058] At the same time, the three parameters need to meet the following conditions (equation 2):

[0059]

[0060] The final output result of the l-th layer (equation 3):

[0061]

[0062] Among them, represents the feature map obtained by transforming the feature at pixel point (i, j) from the 1st, 2nd and 3rd layers to the l-th layer, is the corresponding parameter weight learned by the l-th layer feature map of pixel point (i, j), and is the new feature map obtained after adaptive spatial feature mapping fusion of the l-th layer of pixel point (i, j).

[0063] To better combine channel and spatial attention information, avoid the negative impact of dimension reduction operation on channel attention prediction, and reduce unnecessary and inefficient dependency acquisition, a local cross-channel interaction ECA attention module can be embedded in the PANet feature extraction module. Use a parameter matrix with diagonal elements different from each other to learn channel attention, so that each channel is completely independent; and use grouped convolution to group high and low dimensional channels to reduce the complexity of common learning parameters during model training, and use convolution kernel size and channel dimension to adaptively optimize the coverage range of local cross-channel interaction.

[0064] For the non-reduced dimension aggregation feature y∈R C The weight of each channel thereof can be expressed as follows:

[0065] ω=σ(C1D k (y)) (4)

[0066] The parameter matrix W in the ECA module k The diagonal elements are different, and each channel is completely independent, effectively avoiding the interaction between channels, wherein the parameter matrix can be expressed as follows:

[0067]

[0068] Group Convolutions are used to group high and low dimensional channels, share learning parameters and reduce model complexity. The size of the convolution kernel k is determined by the algorithm as follows:

[0069]

[0070] In the formula, C is the number of channels, γ=2, b=1, and odd is the nearest odd number.

[0071] The input feature map is globally averaged and pooled, and a 1D convolution operation with a convolution kernel size of k is performed. The weight values of different channels are calculated by the Sigmoid function. The required output feature map is obtained by multiplying the original feature map and the corresponding calculated weight. In the improvement strategy for the backbone network CSPDarknet and the feature extraction network PANet, the ECA attention mechanism is added to the output features of the three Res residual networks in the backbone network and the three Sampling sampling layers in the feature extraction network to optimize the network performance.

[0072] To solve the problems of gradient disappearance and weak activation effect that may occur during the training of the Silu activation function in YOLOX, the Swish activation function is used, which has stronger anti-overfitting ability in deep network structures and can effectively prevent the performance loss of neural networks caused by activation function oversaturation. The Swish function formula can be expressed as follows:

[0073] S(x)=x·σ(G(x)) (7)

[0074] In the formula, x is the input data, G(x) is the output of the convolution layer, and σ(·) represents the Sigmoid function.

[0075] The Swish function has no upper boundary and does not have the phenomenon of over-gradient saturation, has a lower boundary and can produce stronger regularization effect, and the characteristics of non-monotonicity, continuity and derivability everywhere make it easier to train. Using the Sigmoid function can stabilize the fluctuation amplitude of the numerical value, and ensure that the deep network calculation iteration can output more stable results.

[0076] Further, based on the above method, the embodiment of the application also provides a remote sensing image point-like independent house detection system based on a YOLOX network, comprising: a sample set establishing module, a detection model constructing module and a target detection module, wherein,

[0077] The sample set establishing module is used for labeling the rural or mountainous residential area monomer building houses in the collected point-like independent house image data set, and expanding the labeled data set through a sample enhancement method to obtain a point-like independent house image sample set;

[0078] The detection model constructing module is used for constructing a point-like independent house target detection model based on the YOLOX network, and establishing a target loss function for model training, wherein the point-like independent house target detection model takes a residual network and a Focus module as a backbone network, takes a PANet network as a feature pyramid extraction network, and takes a decoupled head DecoupledHead as a YOLOHead detection network, and the target loss function uses a Swish activation function; the point-like independent house target detection model is trained and optimized by using the point-like independent house image sample set and based on the established target loss function, to obtain a trained point-like independent house target detection model;

[0079] The target detection module is used for inputting the remote sensing image data to be detected into the trained point-like independent house target detection model, and acquiring a remote sensing image monomer building house target by using the trained point-like independent house target detection model.

[0080] To verify the effectiveness of the scheme, the following experimental data are used for further explanation and description:

[0081] The experimental operating system is Windows 11, the CPU version is 12th Gen Intel(R) Core(TM) i9-12900H, the GPU is NVIDIA GeForce RTX 3080TI, and the deep learning framework is Torch 1.11.0+cu115. The experimental application uses the idea of transfer learning, and uses the Pascal voc2007+2012 dataset to pre-train the weight. Through transfer learning, the network model relies on its unsaturated continuous learning ability to improve the accuracy of the detection result. The training process is divided into a frozen stage and an unfreezing stage. The frozen training parameters are: Epoch = 50, Batch Size = 16, learning rate is 0.001, weight decay rate is 0.0005, and the optimizer uses SGD. The unfreezing training parameters are: Epoch = 100, Batch Size = 8, the learning rate remains unchanged, and the training configuration information is shown in Table 1.

[0082] Table 1 Training configuration of detection algorithm

[0083]

[0084] The target detector has a strong dependence on the scale of the pre-training data. The amount of data required for point-like independent house training is large, therefore, in the experiment, satellite image samples are mainly based on EarthH meta-earth satellite images. The satellite image data is derived from the secondary development of Google Earth, and the main satellite data sources include SPOT series, Pleiades A / B, Geoeye-1, WorldView-1, WorldView-2, WorldView-3, etc. The data set contains 200 samples, and 5955 samples are manually labeled. The size of a single sample is 800 to 4000 pixels, and the spatial resolution is 1m, as shown in Figure 4 The images are sampled from multiple cities across the region, with high diversity.

[0085] To effectively extract the robust and invariant features in point-like independent houses and fully learn the feature information of rural or mountainous residential areas, the size and accuracy of the data set required for training are optimized. Through sample enhancement, the data capacity is expanded, and through saliency enhancement, the feature information of point-like independent houses is enhanced.

[0086] The sample attribute diversity is improved through sample enhancement. By changing the brightness, gray scale, and flipping the image, etc. The morphology changes slightly, but the multi-level features of the sample are learned. In the case of limited actual data set, the existing data is enhanced to generate larger training value sample data, which can improve the stability of the algorithm, reduce the influence of other interference factors, and enhance the processing process as Figure 5 shown.

[0087] Data preprocessing: First, the point-like independent house label information corresponding to the sample image is assigned to the target of this category attribute. Second, read each sample image and its corresponding label information. Third, convert the sample image data into an RGB image. Fourth, read the width and height of the sample image and the width and height of the labeled target. Finally, frame the range of the sample image that needs to be enhanced. Thus, the preprocessing of the sample data input is completed.

[0088] Sample data augmentation: The main purpose is to mix different images in proportion. After mixing, the real box labeled on the original image is now located on one image. The sample data can be augmented through geometric transformation, color transformation, discrete interpolation sampling, etc. In the experiment, two methods, MixUp and Mosaic, are adopted. The single-process implementation idea of MixUp is as follows: (1) read two images; (2) perform data augmentation such as flipping, scaling, and color variation on the two images respectively; (3) stack the real boxes of the two images together. The single-process implementation idea of Mosaic is as follows: (1) read four images; (2) perform data augmentation such as flipping, scaling, and color variation on the four images respectively, and arrange them according to the four directions; (3) combine the images and the real boxes.

[0089] Since the geometric parameters and transformation coefficients in the sample augmentation algorithm can be set manually or determined adaptively, in the experiment, any multiple augmentation can be realized by parameter setting. However, when the number of samples is too small, the learned feature parameters tend to be uniform after high-multiple augmentation. Therefore, in practical applications, the augmentation coefficient should be adjusted according to the task requirements.

[0090] The human visual system has the ability to quickly and accurately locate objects or regions of interest from complex scenes, which is called visual attention mechanism. How to simulate this mechanism makes the residential area information the most visually attractive object or region in the given image, so as to quickly realize segmentation and extraction. Therefore, it is necessary to enhance the features of the residential area in the image, so that the visual saliency of the region of interest is enhanced and the feature information is more obvious. In the experiment, the edge detection operator can be used to accurately extract the feature components of the residential area information in the image of interest, and the pixel density of the feature components is converted into a heat map according to the Gaussian function, so that the saliency heat map generated according to the feature components can be used to estimate the accurate area of the residential area. The generated saliency heat map is fused into the original image at an adaptive parameter multiple, so as to realize the hierarchical enhancement of the residential area and the weakening effect on the non-interest region. As shown in Figure 6 , the texture features are extracted by performing saliency detection on the image; and as shown in Figure 7 , (a) is the original image, and (b) is the saliencyFigure 2 Value, (c) significance heat map, (d) image fusion after enhancement result, fuse the original image and the gray image / heat image to complete the significance enhancement of residential information.

[0091] Further combined with ablation test and comparative test, the effectiveness of the algorithm of the present scheme is explained:

[0092] 1. Ablation test

[0093] In order to qualitatively analyze the contribution degree of the present scheme to the network, the YOLOX is taken as the baseline system to test the experimental effect of the point-shaped house data set. The performance improvement effect of the model is evaluated through mAP (mean Average Precision), F1 score, Precision and RECAll recall rate, wherein the mAP refers to the average precision of the model recognition, which is a comprehensive evaluation index of the accuracy of the model prediction for all categories; the F1 score comprehensively considers the influence of precision and recall rate; the precision is the percentage of the model correctly predicted independent house to the number of real independent house; the recall rate is the percentage of the correctly predicted point-shaped independent house to the real independent house. The network training process is visualized, and the improved network effect is comprehensively analyzed. The parameter settings and experimental environment remain the same, and the results are shown in Table 2:

[0094] Table 2 Ablation experiment of different improvement strategies

[0095]

[0096] The above experiments can show that:

[0097] (1) The effect detection of small target independent house is improved, and the small target feature ability is enhanced. There are various types of ground objects with different scales on remote sensing images, and the apparent form, size of the same type of ground object at different scales also has great difference. Compared with large target and large and medium size ground objects, small target features are not obvious and easy to be disturbed, so the detection features are less and easy to be missed and misdetected. After the optimization and improvement of the algorithm, the detection rate of point-shaped independent house with different scales is obviously improved, and the mAP precision of the algorithm of the present scheme is optimally improved by 11.16% compared with the original algorithm. When there are single houses with different scales, the small size house feature capture ability is also improved, the detection effect of the improved algorithm for small size point-shaped independent house is obviously enhanced, the detection rate for small target ground object is improved, and the feature information of small size independent house can be better mined. The present scheme can be applied to the acquisition and information mining of multi-scale image point-shaped independent house of remote sensing image.

[0098] (2) The detection effect of dense houses is improved. From the original data set, it can be seen that the point-like independent house in the remote sensing image is difficult to detect under the premise of small target detection, and it also has the characteristics of single-class ground object dense distribution and background ground object mixed distribution. The feature information is more complex and difficult to locate and detect. The arrangement of point-like independent houses in remote sensing images often shows disorder, high overlap and dense complexity, which easily leads to high frequency of missed detection and misdiagnosis. Therefore, the algorithm of the case scheme: strengthen the sensitivity of the algorithm to dense houses; improve the utilization rate of the original features of point-like independent houses; effectively reduce the feature interference of non-key background information; reduce the missed detection and misdiagnosis rate of independent houses. After optimization, the detection effect of dense houses is improved, and the effectiveness and applicability of the algorithm of the case scheme in the dense distribution of point-like independent houses are verified.

[0099] (3) The improvement of network detection effect is restricted by the complexity of house background. The original data set sample collection range is wide, and the topography and geomorphology characteristics are diverse, including mountains, hills and plains; and the difference of image background is large, the image information is complex and changeable, the regularity is poor, and the house information is easy to be mixed with other background information. Other types of ground objects also interfere with the extraction of house features, such as shielding houses and similar textures that are easy to confuse. However, the ability of the algorithm of the case scheme to suppress negative interference information has been greatly improved.

[0100] As can be seen from Table 2, by introducing ASFF, ECA and Swish mechanism through ablation experiment on the basis of the original network, mAP is increased by 3.16%, 2.57% and 2.67% respectively, rECAll is increased by 4.07%, 3.68% and 3.7% respectively, and F1 is increased by 2.94%, 2.21% and 2.27% respectively. The introduction of ASFF has a positive effect on the improvement of precision, and precision is increased by 1.15%. The introduction of ECA and Swish mechanism causes a slight decrease in precision, and precision is decreased by 0.15% and 0.02% respectively. Given that the precision value of the original model network can reach 93.19%, the precision is high, and after combining the mechanism, the precision reaches 96.32%. Although the precision value produces a slight loss in the single mechanism experiment, the optimal precision value is still increased by 3.28% compared with the original network. The precision is further optimized.

[0101] The experimental results show that several improved mechanisms have a positive promoting effect on the improvement of network effect, and the optimization of activation function has the most obvious improvement effect. It is found that after using Swish activation function, stronger regularization effect is produced, the expression of some information is enhanced, and part of the information is effectively suppressed. The non-monotonic and everywhere derivable characteristics make the value amplitude easy to control in the deep network, and the approximation ability of the network is more powerful, and the gain effect is obvious.

[0102] Therefore, the scheme of the present case focuses on local features, positive features, and improves the classification regression effect. Through improved mechanism combination and ablation experiment quantitative test, it is concluded that when combining ASFF, ECA and Swish modules, the improved network improves the mAP of the original YOLOX network model by 11.16%, and the combination mechanism achieves better results based on single improvement mode.

[0103] 2. Comparative test

[0104] (1) Attention algorithm comparison experiment

[0105] Attention mechanism (Attention Mechanism) is derived from the study of human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on part of all information while ignoring other visible information. In order to make rational use of limited visual information processing resources, humans need to select a specific part of the visual area and then focus on it. In summary, attention mechanism mainly has two aspects: deciding which part of the input needs to be focused on; allocating limited information processing resources to important parts. In order to select an attention mechanism algorithm suitable for point-like independent house feature information of remote sensing images, several mainstream attention mechanism algorithms such as SENet, CBAM and ECA are selected for comparison. Among them, SENet is a typical implementation of channel attention mechanism. For the input feature layer, the network can focus on the channel it needs to focus on. CBAM combines channel attention mechanism and spatial attention mechanism. For the input feature layer, it is processed by channel attention mechanism and spatial attention mechanism respectively. The comparison results are shown in Table 3:

[0106] Table 3 Comparison of gain effects of different attention mechanisms

[0107]

[0108]

[0109] The experimental results show that in the extraction of point-like independent houses on remote sensing images, the comparison of several attention mechanism algorithms can find that different attention algorithms on the YOLOX algorithm model all achieve good gain effect, among which the ECA cross-channel interaction attention method achieves better effect, and the mAP optimally improves by 2.57%. Compared with other algorithms, more feature information of the attention object is extracted, and the feature enhancement method is used. Under the premise of complex background information of remote sensing images, ECA reduces unnecessary and inefficient dependent attention interaction methods by suppressing the negative effects of unnecessary information in the network, and achieves better house extraction effect, verifying the effectiveness of the scheme of the present case.

[0110] (2) Activation function comparison experiment

[0111] Activation functions play a very important role in algorithm model learning, understanding very complex and nonlinear functions. They introduce nonlinear characteristics into neural networks. Enhance the approximation ability of the network, so that the deep neural network has stronger feature expression ability. In summary, several mainstream activation functions such as Relu, Leaky relu, Silu, Swish are selected for experiments. The Relu function can suppress all negative parameters to zero through the single-sided suppression feature to prevent parameter overfitting; The Leaky relu function can solve the zero gradient problem by weakening the negative parameters rather than parameter zero. Silu function solves the problem of non-monotonic parameters by introducing an exponential. By using multiple activation functions for experiments, an activation mode that can better improve the detection effect of remote sensing image point-like independent houses is selected. The comparison results are shown in Table 4:

[0112] Table 4 Activation function gain effect comparison

[0113]

[0114] Multiple activation function experiment comparison found that the use of activation function effectively activates the algorithm performance, enhances the deep expression ability and feature utilization rate of the network, among them, Swish activation function has stronger activation ability than other functions, promotes the regularization expression of algorithm network, mAP precision optimal improvement 6.07%, which verifies the effectiveness of Swish activation function for algorithm improvement.

[0115] (3) Different algorithm comparison experiment

[0116] In the experiment, the label information of the original data set is visualized on the image, and the mAP is used as the index for quantitative analysis and comparison with the experimental results. In order to verify the practicability and effectiveness of the algorithm, several widely used deep learning target detection algorithms such as YOLOv3, YOLOv4, YOLOv5, yolov7 and the original YOLOX model are selected for comparison. The comparison results are as follows Figure 8As shown, (a) DOTA data (red box is the label); (b) YOLOv3; (c) YOLOv4; (d) YOLOv5; (e) YOLOX; (f) improved algorithm, which shows the prediction results of YOLOv3, YOLOv4, YOLOv5, original YOLOX algorithm and the algorithm of the case, based on the visualization results of the data set in (a), the performance of different algorithms is evaluated: (1) classification accuracy is improved: from the analysis of the prediction information, for single point independent house detection, the algorithm of the case improves the probability of judgment compared with other algorithms, which ensures the accuracy of house detection, and the prediction accuracy of the model network is higher. (2) The network is more suitable for small target classification detection: the algorithm performance of the case is balanced, which can identify and detect different scale targets, and can accurately classify and regress, and the classification detection performance is more comprehensive than other algorithms, and the applicability is stronger and the effect is better. (3) The detection performance is stronger: based on the original data set of single point independent house, the network in the algorithm of the case identifies the point independent house in the image which is not accurately labeled, compared with other algorithms, the number of labels predicted by the algorithm of the case is obviously increased, which shows that the number of targets detected by the algorithm of the case is more, the classification accuracy of point independent house is higher, and the performance of classification and detection of targets is stronger. (4) The detection ability in complex background is stronger: Figure 8 When the image tone tends to be consistent, and different colors of vegetation are mixed, the network in the algorithm of the case can identify more independent house information in the complex background object indicated by the arrow than other networks, and can focus on the information of the object of interest, and the effect is better in the image with complex background information. The specific comparison results are shown in Table 5.

[0117] Table 5 Comparison results of different algorithms

[0118]

[0119] The precision data in Table 5 shows that compared with YOLOv4, YOLOv5 and YOLOX, the algorithm of the case realizes the improvement of mAP precision by 11.16%-66.33%, and it can also be seen that the mainstream target detection algorithm still has a large space for improvement in the detection ability of small target objects. The mAP value of the original algorithm YOLOX model selected in the case reaches 83.67% in the detection of small target objects, which has obvious advantages. After optimizing the design of the model algorithm, the mAP value reaches 94.83, and the performance of detecting point independent house is further enhanced.

[0120] When applied to a dataset of point-like independent houses, the algorithm presented in this study achieved significant target detection performance even with complex and diverse background topography. Experiments show that the improved network based on the YOLOX algorithm model has significant advantages and strong applicability in multi-domain and multi-scene detection. Furthermore, it significantly improves detection accuracy for densely distributed target classes, such as a marked increase in the number of houses detected in overlapping clusters of point-like independent houses in the image. This demonstrates the strong practicality and clear advantages of the algorithm in current target detection networks. Compared to the original YOLOX algorithm (using default parameter settings), the algorithm in this study improves mAP by 12.75%, further illustrating the effectiveness of the YOLOX network model in this algorithm.

[0121] The visualization results of the original point-based independent house dataset sample labels and the experimental prediction results are comprehensively analyzed and compared. Figure 9 As can be seen from (b) and (d) in the figure, the improved algorithm can effectively classify and detect ground objects. First, it can achieve high-precision positioning and identification of point-like independent houses. Second, the probability of correctly classifying and identifying houses is relatively high, and it has good identification ability. Figure 10 As indicated by the arrows, the original dataset of isolated point houses suffers from inaccurate location labeling and mislabeling / missing densely packed houses. Optimizing the model network reduces the bias of the original data, resulting in more accurate regression classification and higher localization and detection precision. However, it also has significant shortcomings: First, the completeness of point house extraction is weak. When point houses of significantly different scales appear simultaneously in the same image, some houses can be completely extracted, while others can only have parts of their structures selected, with some still failing to be effectively detected. The completeness detection of houses is weak, with some missing elements. Figure 11 As shown, there is still significant room for improvement in the detection of dense targets. Due to factors such as satellite imagery capture or imaging angle, densely distributed point-like detached houses appear to overlap highly in remote sensing images, such as... Figure 12 When some houses are obscured or too close together, it is difficult to distinguish different house targets. The algorithm misses a lot of targets, the model detection effect is poor, and it fails to effectively identify the targets compared with the original data sample and the marked position.

[0122] The above data demonstrates that the detection model in this solution achieves better detection results and accuracy, and is easier to deploy. The trained target detection model exhibits better robustness and effectiveness, making it suitable for extracting point-like independent houses and detecting house distribution in remote sensing images, thus providing strong support for surveying and mapping production.

[0123] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0124] The various embodiments described in the specification are progressive in nature, and each embodiment highlights the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0125] The units and method steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been described in the above description in general terms. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation does not exceed the scope of the present application.

[0126] Those skilled in the art can understand that all or part of the steps of the above method can be instructed by a program to complete the relevant hardware, and the program can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk. Alternatively, all or part of the steps of the above embodiments can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiments can be implemented in the form of hardware or in the form of a software function module. The present application is not limited to any specific form of combination of hardware and software.

[0127] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present application, which are used to illustrate the technical solutions of the present application, and are not limiting. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can make modifications or easily think of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some of the technical features. Such modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for detecting point-like independent houses in remote sensing images based on YOLOX networks, characterized in that, Include: The individual buildings in rural or mountainous residential areas in the collected point-like detached house image dataset are labeled, and the labeled dataset is expanded by sample augmentation to obtain the point-like detached house image sample set. A point-based independent house target detection model is constructed based on the YOLOX network, and a target loss function is established for model training. In the point-based independent house target detection model, the residual network and Focus module are used as the backbone network, the PANet network is used as the feature pyramid extraction network, and the DecoupledHead is used as the YOLOHead detection network. The target loss function uses the Swish activation function. The point-based independent house target detection model was trained and optimized using a sample set of point-based independent house images and based on the established target loss function, resulting in the trained point-based independent house target detection model. The remote sensing image data to be detected is input into the trained point-based detached house target detection model, and the trained point-based detached house target detection model is used to obtain the single building targets in the remote sensing image. The method of augmenting the labeled dataset through sample enhancement also includes: enhancing the visual salience of regions of interest by performing feature enhancement on residential areas in the images of the dataset; Feature enhancement is performed on residential areas in the dataset to improve the visual saliency of regions of interest, including: First, texture feature components of residential information in the region of interest in the dataset image are extracted by edge detection, and the texture feature components are transformed into a salient heatmap using a Gaussian function; Then, residential areas are estimated based on the saliency heatmap, and the estimated saliency heatmap is fused onto the original image with an adaptive parameter scaling factor. The residential areas of interest in the original image are then layered and enhanced, and the effects of non-interest areas are reduced.

2. The method for detecting point-like independent houses in remote sensing images based on YOLOX network according to claim 1, characterized in that, The labeled dataset is augmented using sample augmentation techniques, including: First, iterate through the individual building labels in the dataset and assign values ​​to the individual building targets for the category attributes of those labels; Next, the image data is converted into RGB images by reading the image width and height and the width and height of the labeled target. The range of the sample image to be enhanced is defined by reading the image width and height and the labeled target width and height. Then, the different image data are mixed proportionally, and the framed sample images are augmented using sample enhancement methods, including geometric transformation, color transformation, and discrete interpolation sampling.

3. The method for detecting point-like independent houses in remote sensing images based on YOLOX network according to claim 1, characterized in that, The point-based independent house target detection model uses a residual network and a Focus module as the backbone network, a PANet network as the feature pyramid extraction network, and a decoupled head as the YOLO head detection network. It includes: the PANet network in the feature pyramid extraction network, an adaptive spatial feature fusion network, and feature mapping fusion by multiplying and adding the extracted features with the weight parameters of each layer, filtering features from other layers and retaining key local region information to extract fine-grained features from the input image; Furthermore, a local cross-channel interaction ECA attention module is embedded in the residual network and PANet network to enhance the regional feature information of the residents.

4. The method for detecting point-like independent houses in remote sensing images based on YOLOX network according to claim 3, characterized in that, The process of using an adaptive spatial feature fusion network and multiplying and adding the extracted features with the weight parameters of each layer to perform feature mapping and fusion can be represented as follows: in, This represents the feature map obtained after the features at pixel (i, j) are transformed from layers 1, 2, and 3 to layer l, respectively. Let (i, j) be the weights of the corresponding parameters learned from the l-th layer feature map, and The new feature map is obtained after fusing the adaptive spatial feature mapping of the l-th layer for pixel (i, j).

5. The method for detecting point-like independent houses in remote sensing images based on YOLOX network according to claim 3, characterized in that, The local cross-channel interactive ECA attention module is used to enhance the regional feature information of residents, including: using a parameter matrix W with distinct diagonal elements. k Channel attention is learned to make each channel completely independent; grouped convolutions are used to group high- and low-dimensional channels to reduce the complexity of co-learning parameters during model training; and the local cross-channel interaction coverage is adaptively optimized by utilizing the convolution kernel size and channel dimension. The process of adaptively optimizing the local cross-channel interaction coverage is expressed as follows: k represents the kernel size, C represents the channel dimension, and |*| odd γ is the nearest odd number, and b is a preset adaptive scaling constant.

6. The method for detecting point-like independent houses in remote sensing images based on YOLOX network according to claim 1, characterized in that, The target loss function uses the Swish activation function, which is expressed as S(x)=x·σ(G(x)), where x is the input data, G(x) is the output of the convolutional layer, and σ(·) represents the Sigmoid function.

7. A remote sensing image point-based independent house detection system based on YOLOX network, characterized in that, It includes: a sample set establishment module, a detection model construction module, and an object detection module, among which, The sample set building module is used to label individual buildings in rural or mountainous residential areas in the collected point-like independent house image dataset, and to expand the labeled dataset through sample augmentation to obtain the point-like independent house image sample set. The detection model construction module is used to build a point-based independent house target detection model based on the YOLOX network and establish a target loss function for model training. The point-based independent house target detection model uses a residual network and a Focus module as the backbone network, a PANet network as the feature pyramid extraction network, and a DecoupledHead as the YOLOHead detection network. The target loss function uses the Swish activation function. The point-based independent house target detection model is trained and optimized using a set of point-based independent house image samples and based on the established target loss function, resulting in the trained point-based independent house target detection model. The target detection module is used to input the remote sensing image data to be detected into the trained point-based detached house target detection model, and use the trained point-based detached house target detection model to obtain the single building targets in the remote sensing image. The method of augmenting the labeled dataset through sample enhancement also includes: enhancing the visual salience of regions of interest by performing feature enhancement on residential areas in the images of the dataset; Feature enhancement is performed on residential areas in the dataset to improve the visual saliency of regions of interest, including: First, texture feature components of residential information in the region of interest in the dataset image are extracted by edge detection, and the texture feature components are transformed into a salient heatmap using a Gaussian function; Then, residential areas are estimated based on the saliency heatmap, and the estimated saliency heatmap is fused onto the original image with an adaptive parameter scaling factor. The residential areas of interest in the original image are then layered and enhanced, and the effects of non-interest areas are reduced.

8. An electronic device, characterized in that, The system includes a memory and a processor, which communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the steps of the method as described in any one of claims 1 to 6 by calling the program instructions.

Images