A remote sensing image target detection method and system based on an optimized SSD algorithm

Abstract

The application discloses a remote sensing image target detection method and system based on an optimized SSD algorithm, first, inputting a to-be-detected remote sensing image into an efficient remote sensing image target feature extraction network to extract key features, and obtaining a required feature map; then inputting the obtained feature map into a forward and reverse iterative fusion multi-scale feature network to perform feature fusion, and generating a new feature map; finally, utilizing an anchor box matching network based on a clustering algorithm to perform anchor box clustering matching on different categories of targets in the image according to the new feature generated by fusion, so that each to-be-detected target obtains the most suitable anchor box as a target detection result; the original feature extraction network VGG16 in the SSD algorithm is optimized, an improved RFB module and an efficient attention mechanism are added into the network, the feature extraction capability of the network is improved, the anchor box matching network based on the clustering algorithm is integrated into the algorithm, the accuracy of the anchor box matching target of the algorithm is improved, and therefore the detection capability is improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and target detection technology, and relates to a remote sensing image target detection method and system, particularly a remote sensing image target detection method and system based on an optimized SSD algorithm. Background Technology

[0002] Remote sensing technology is a non-contact detection technology that operates far from the ground. It utilizes the principles of electromagnetic waves, employing various sensors mounted on aircraft, satellites, and other spacecraft to collect and process electromagnetic wave information transmitted from ground objects, thereby creating remote sensing images. The information contained within these images is then analyzed and identified. The rapid development of the aerospace industry is also driving the rapid advancement of remote sensing technology, enabling people to better observe and detect land and oceans from multiple levels, aspects, and angles. The valuable remote sensing data obtained during Earth observation, combined with these data to form remote sensing images, greatly expands human vision and deepens our understanding of the Earth's surface. Currently, remote sensing images are widely used in various fields. For example, in urban management, they can detect the conditions of buildings, vehicles, and roads, helping relevant departments manage traffic and alleviate the increasingly serious traffic congestion problem; in marine monitoring, they can efficiently detect the situation of ships at sea and provide timely feedback on changes in sea conditions; in the military field, they can quickly locate and effectively strike enemy aircraft, armored vehicles, and other targets, providing crucial support for timely and accurate judgments by combatants. In summary, accurately and quickly detecting target objects in remote sensing images is of great value.

[0003] With the rapid development and progress of remote sensing technology, the resolution of remote sensing images is constantly improving, and the target information contained in the images is also increasing. The detection task is becoming increasingly arduous, and the current target detection algorithms can no longer meet the needs. If the current algorithms are directly applied to remote sensing images for the detection of small targets, it is difficult to achieve satisfactory results. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a remote sensing image target detection method and system based on an optimized SSD algorithm, which can effectively detect targets in remote sensing images and greatly improve the accuracy of target detection.

[0005] The technical solution adopted by the method of the present invention is: a remote sensing image target detection method based on an optimized SSD algorithm, comprising the following steps:

[0006] Step 1: Input the remote sensing image to be detected into the efficient remote sensing image target feature extraction network to extract key features and obtain the required feature map;

[0007] The efficient remote sensing image target feature extraction network comprises 13 sequentially connected layers. The first layer is a convolutional layer that receives the input data, transforming the resolution of the input image feature map to 300*300*3 with a stride of 2 and a kernel size of 3*3. The second layer is also a convolutional layer, transforming the resolution of the image feature map to 38*38*512 with a stride of 2. An efficient attention mechanism is added to the second layer, comprising global max pooling and global average pooling. After pooling, fully connected layers, ReLU activation functions, and another fully connected layer are added, and finally, a Sig... The `moid` function is used for processing, followed by multiplication of channel weights to obtain the result. This result is then convolved and combined with the upsampled features. After processing through an RFB structure, the final result is input into a Non-Maximum Suppression (NMS) layer for non-maximum suppression. The third layer is a MaxPooL layer with a stride of 2, outputting a feature map size of 19*19*512. The fourth layer is a Conv_3*3 convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*512. The fifth layer is a convolutional layer with a stride of 1, resulting in a feature map resolution of 19*1. The network consists of nine layers: the sixth layer is a double convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024 after convolution, followed by a 19*19*256 convolution, with the result fed into NMS; the seventh layer is an RFB structure layer for enhanced feature extraction, with the result fed into NMS; the eighth layer is an RFB structure layer for enhanced feature extraction, outputting a feature map resolution of 19*19*512, with the result fed into NMS; the ninth layer is an RFB structure layer for enhanced feature extraction, outputting a feature map resolution of 19*19*512, with the result fed into NMS. The network consists of three layers: the tenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 10*10*128; the eleventh layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 5*5*128, and the result is input into NMS; the twelfth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 3*3*128; the thirteenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 1*1*256, and the result is input into NMS; the output of NMS is the final extracted feature map.

[0008] Step 2: Input the obtained feature map into the forward and reverse iterative fusion multi-scale feature network to perform feature fusion and generate a new feature map;

[0009] The forward and reverse iterative fusion multi-scale feature network includes T-Conv_1 layer, T-Conv_2 layer, T-Conv_3 layer, T-Conv_4 layer, T-Conv_5 layer and Conv layer. It is used to process the features T1, T2, T3, T4 and T5 output by the efficient remote sensing image target feature extraction network through a network and cascade to make all features have the same feature information scale. Then, in the order of T5, T4, T3, T2 and T1, they are passed through the upper network and cascaded for reverse fusion operation. Finally, forward and reverse fused multi-scale features are obtained. The multi-layer fused features are then passed through a convolutional layer with a kernel size of 3 and a stride of 2 to obtain the overall features.

[0010] The T-Conv_1 layer has a stride of 1, a kernel size of 5*5, and 1024 kernels; the T-Conv_2 layer has a stride of 1, a kernel size of 5*5, and 512 kernels; the T-Conv_3 layer has a stride of 1, a kernel size of 5*5, and 256 kernels; the T-Conv_4 layer has a stride of 1, a kernel size of 5*5, and 128 kernels; the T-Conv_5 layer has a stride of 1, a kernel size of 5*5, and 64 kernels; and the Conv layer has a stride of 2, a kernel size of 3*3, and 64 kernels.

[0011] Step 3: Using an anchor box matching network based on clustering algorithm, anchor boxes of different categories of targets in the image are clustered and matched according to the new features generated by fusion, so that each target to be detected gets the most suitable anchor box, which is used as the target detection result;

[0012] The anchor box matching network based on the clustering algorithm consists of a first convolutional layer, a second convolutional layer, a Canopy allocation layer, and a K-means algorithm aggregation anchor box layer. The first convolutional layer is a 7*7 kernel layer with a stride of 1. The second convolutional layer is a 3*3 kernel layer with a stride of 2. The Canopy allocation layer consists of a third convolutional layer, a Canopy layer, a residual block, and a third deconvolutional layer. The third convolutional layer has a stride of 1 and a kernel size of 5*5. The third deconvolutional layer has a stride of 1 and a kernel size of 5*5; the Canopy layer has a stride of 1, a kernel size of 5*5, and a ReLU activation function; the K-means algorithm aggregated anchor box layer consists of a fourth convolutional layer, a K-means convolutional layer, and a fourth deconvolutional layer; wherein the fourth convolutional layer has a stride of 2 and a kernel size of 5*5; the fourth deconvolutional layer has a stride of 2 and a kernel size of 5*5; the K-means convolutional layer has a stride of 2 and a kernel size of 5*5, and a ReLU activation function.

[0013] The technical solution adopted by the system of the present invention is: a remote sensing image target detection system based on an optimized SSD algorithm, comprising the following modules:

[0014] The first module is used to input the remote sensing image to be detected into the efficient remote sensing image target feature extraction network to extract key features and obtain the required feature map.

[0015] The efficient remote sensing image target feature extraction network comprises 13 sequentially connected layers. The first layer is a convolutional layer that receives the input data, transforming the resolution of the input image feature map to 300*300*3 with a stride of 2 and a kernel size of 3*3. The second layer is also a convolutional layer, transforming the resolution of the image feature map to 38*38*512 with a stride of 2. An efficient attention mechanism is added to the second layer, comprising global max pooling and global average pooling. After pooling, fully connected layers, ReLU activation functions, and another fully connected layer are added, and finally, a Sig... The `moid` function is used for processing, followed by multiplication of channel weights to obtain the result. This result is then convolved and combined with the upsampled features. After processing through an RFB structure, the final result is input into a Non-Maximum Suppression (NMS) layer for non-maximum suppression. The third layer is a MaxPooL layer with a stride of 2, outputting a feature map size of 19*19*512. The fourth layer is a Conv_3*3 convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*512. The fifth layer is a convolutional layer with a stride of 1, resulting in a feature map resolution of 19*1. The network consists of nine layers: the sixth layer is a double convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024 after convolution, followed by a 19*19*256 convolution, with the result fed into NMS; the seventh layer is an RFB structure layer for enhanced feature extraction, with the result fed into NMS; the eighth layer is an RFB structure layer for enhanced feature extraction, outputting a feature map resolution of 19*19*512, with the result fed into NMS; the ninth layer is an RFB structure layer for enhanced feature extraction, outputting a feature map resolution of 19*19*512, with the result fed into NMS. The network consists of three layers: the tenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 10*10*128; the eleventh layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 5*5*128, and the result is input into NMS; the twelfth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 3*3*128; the thirteenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 1*1*256, and the result is input into NMS; the output of NMS is the final extracted feature map.

[0016] The second module is used to input the obtained feature map into a forward and reverse iterative fusion multi-scale feature network for feature fusion to generate a new feature map.

[0017] The forward and reverse iterative fusion multi-scale feature network includes T-Conv_1 layer, T-Conv_2 layer, T-Conv_3 layer, T-Conv_4 layer, T-Conv_5 layer and Conv layer. It is used to process the features T1, T2, T3, T4 and T5 output by the efficient remote sensing image target feature extraction network through a network and cascade to make all features have the same feature information scale. Then, in the order of T5, T4, T3, T2 and T1, they are passed through the upper network and cascaded for reverse fusion operation. Finally, forward and reverse fused multi-scale features are obtained. The multi-layer fused features are then passed through a convolutional layer with a kernel size of 3 and a stride of 2 to obtain the overall features.

[0018] The T-Conv_1 layer has a stride of 1, a kernel size of 5*5, and 1024 kernels; the T-Conv_2 layer has a stride of 1, a kernel size of 5*5, and 512 kernels; the T-Conv_3 layer has a stride of 1, a kernel size of 5*5, and 256 kernels; the T-Conv_4 layer has a stride of 1, a kernel size of 5*5, and 128 kernels; the T-Conv_5 layer has a stride of 1, a kernel size of 5*5, and 64 kernels; and the Conv layer has a stride of 2, a kernel size of 3*3, and 64 kernels.

[0019] The third module is used to use a clustering-based anchor box matching network to cluster and match anchor boxes for different categories of targets in the image based on the new features generated by fusion, so that each target to be detected gets the most suitable anchor box as the target detection result.

[0020] The anchor box matching network based on the clustering algorithm consists of a first convolutional layer, a second convolutional layer, a Canopy allocation layer, and a K-means algorithm aggregation anchor box layer. The first convolutional layer is a 7*7 kernel layer with a stride of 1. The second convolutional layer is a 3*3 kernel layer with a stride of 2. The Canopy allocation layer consists of a third convolutional layer, a Canopy layer, a residual block, and a third deconvolutional layer. The third convolutional layer has a stride of 1 and a kernel size of 5*5. The third deconvolutional layer has a stride of 1 and a kernel size of 5*5; the Canopy layer has a stride of 1, a kernel size of 5*5, and a ReLU activation function; the K-means algorithm aggregated anchor box layer consists of a fourth convolutional layer, a K-means convolutional layer, and a fourth deconvolutional layer; wherein the fourth convolutional layer has a stride of 2 and a kernel size of 5*5; the fourth deconvolutional layer has a stride of 2 and a kernel size of 5*5; the K-means convolutional layer has a stride of 2 and a kernel size of 5*5, and a ReLU activation function.

[0021] Compared to existing technologies, this invention optimizes the original feature extraction network VGG16 in the SSD algorithm by adding an improved RFB module and an efficient attention mechanism to enhance the network's feature extraction capabilities. It also incorporates a forward and reverse iterative fusion multi-scale feature network into the algorithm to fully integrate features and improve feature representation. Furthermore, it integrates the anchor box matching network from the clustering algorithm into the algorithm to improve the accuracy of anchor box matching targets, thereby enhancing detection capabilities.

[0022] This invention demonstrates superior detection results for remotely sensed targets in various scenarios, achieving higher accuracy than the initial algorithm. Compared to existing algorithms, this invention exhibits superior detection performance. In detecting targets in remotely sensed images against complex backgrounds, this invention demonstrates higher accuracy and better robustness. Attached Figure Description

[0023] Figure 1 This is a diagram of the efficient remote sensing image target feature extraction network structure according to an embodiment of the present invention;

[0024] Figure 2 This is a diagram of the forward and reverse iterative fusion multi-scale feature network structure according to an embodiment of the present invention;

[0025] Figure 3 This is a diagram of the anchor box matching network structure based on clustering algorithm according to an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the network training process according to an embodiment of the present invention;

[0027] Figure 5 This is a diagram showing the experimental detection results of an embodiment of the present invention. Detailed Implementation

[0028] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0029] This invention provides a remote sensing image target detection method based on an optimized SSD algorithm, comprising the following steps:

[0030] Step 1: Input the remote sensing image to be detected into the efficient remote sensing image target feature extraction network to extract key features and obtain the required feature map;

[0031] Please see Figure 1The efficient remote sensing image target feature extraction network in this embodiment includes 13 sequentially connected layers. The first layer is a convolutional layer, which receives the input data and transforms the resolution of the input image feature map to 300*300*3 with a stride of 2 and a kernel size of 3*3. The second layer is also a convolutional layer, transforming the resolution of the image feature map to 38*38*512 with a stride of 2. An efficient attention mechanism is added to the second layer, which includes global max pooling and global average pooling. After pooling, a fully connected layer, a ReLU activation function, and another fully connected layer are added. Finally, the signal is processed by the Sigmoid function, followed by a multiplication of the channel weights to obtain the result. The result is then convolved. The first layer combines the features with those after upsampling, processes them through an RFB structure, and finally inputs the processed result into a Non-Maximum Suppression (NMS) layer for non-maximum suppression. The third layer is a MaxPooL layer with a stride of 2, outputting a feature map size of 19*19*512. The fourth layer is a Conv_3*3 convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*512. The fifth layer is a convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024. The sixth layer is a double convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024, followed by a convolution of 19*19*256, and the result is input into NMS. The seventh layer... The network uses an RFB (Reactive Fusion) structure layer to enhance feature extraction, and the results are input into NMS (Non-Maximum Segmentation). The dilated convolutional layers in the RFB module expand the receptive field while maintaining the number of parameters, resulting in better discriminative power for the extracted features. In the bottleneck structure, 1×1 convolutions are first used to reduce channel features, then an n×n convolution layer is added, followed by computationally simple 1×3, 3×1, 1×7, and 7×1 convolutions to reduce computational cost, thereby obtaining a better non-linear parallel structure to offset the effects of convolution kernel degradation. This scheme reduces the number of channels in the convolutional layers by using 1×1 convolutional layers without activation functions. The eighth layer is also an RFB structure layer, performing enhanced feature extraction and outputting a feature map with a resolution of 19*19*512, which is then input into NMS. In NMS, the ninth layer is an RFB (Resource-Free) layer that enhances feature extraction, outputting a feature map with a resolution of 19*19*512, which is then input into NMS. The tenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 10*10*128. The eleventh layer is also a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 5*5*128, which is then input into NMS. The twelfth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 3*3*128. The thirteenth layer is another convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 1*1*256, which is also input into NMS.

[0032] Depend on Figure 1 It is known that the output feature map has 7 layers. Layers 2, 6, 7, 8, and 9 are input into NMS, and layers 11 and 13 are fused and then input into NMS. The output of NMS is the final extracted feature map.

[0033] This embodiment's efficient remote sensing image target feature extraction network integrates an efficient attention mechanism, performing parallel operations of global max pooling and global average pooling, and then connecting fully connected layers to extract information between channels. Based on the interrelationships between channels, it assigns weights to each channel more reasonably, rapidly improving the extraction of target features. First, different information is added to generate a new feature vector. The two added vectors represent the Nth prediction layer (from shallow to deep) and the N-1th output layer (from deep to shallow) of the SSD algorithm model, respectively. Then, the generated feature maps are processed using global average pooling (GAP) and global max pooling (GMP). Subsequently, the two generated feature information are input into fully connected layers. The first fully connected layer reduces the number of channels, and the second fully connected layer increases the number of channels. The ReLU activation function handles the linear and non-linear relationships between channels. Finally, the two channel information are aggregated to generate a set of channel weights, where the weights reflect the correlation between channels.

[0034] The efficient attention function H(GA) is:

[0035]

[0036] In the formula, () represents the fully connected layer function, Q represents the fully connected coefficient, Relu() represents the activation function, μ represents the correlation coefficient, and m represents the m-th feature.

[0037] Finally, the results are processed using the Sigmoid function, followed by a multiplication of the channel weights to obtain the final result. Using this channel attention module in the network can enhance useful feature information, reduce feature confusion, and make the model focus more on the target features.

[0038] This embodiment extracts the relationship between each feature channel and its adjacent feature channels through global average pooling and fully connected layers, and then performs global average pooling to obtain the required global receptive field y. i ;

[0039]

[0040] Where i represents the horizontal parameter in the feature matrix, j represents the vertical parameter in the feature matrix, H represents the maximum height of the feature map, and W represents the maximum width of the feature map. j This represents the weight value of each channel, k is the number of convolutional kernels in the convolutional layer, and x...c (i, j) represents the longitudinal and transverse parameter matrices in each channel.

[0041] The efficient feature extraction network structure in this embodiment integrates an improved RFB structure. This structure uses a multi-branch structure, convolutional kernels of different sizes, and dilated convolution to increase the receptive field. All branches are added to multi-scale regular convolutional layers and dilated convolutional layers respectively. In the multi-scale regular convolutional layers, multiple perceptions in the receptive field of the group are simulated. The eccentricity in the dilated convolutional layer is used to simulate the ratio of eccentricity to scale in the receptive field of the group. Finally, the different channels are converged. The dilated convolutional layer of the RFB module expands the receptive field while maintaining the number of parameters, so that the extracted features have better discriminative power.

[0042] Step 2: Input the obtained feature map into the forward and reverse iterative fusion multi-scale feature network to perform feature fusion and generate a new feature map;

[0043] Please see Figure 2 The forward and reverse iterative fusion multi-scale feature network in this embodiment includes T-Conv_1 layer, T-Conv_2 layer, T-Conv_3 layer, T-Conv_4 layer, T-Conv_5 layer and Conv layer. It is used to process multiple features T1, T2, T3, T4 and T5 output by the efficient remote sensing image target feature extraction network through a network and cascade to make all features be within the same range of feature information scale. Then, in the order of T5, T4, T3, T2 and T1, they are passed through the upper network and cascaded for reverse fusion operation. Finally, forward and reverse fused multi-scale features are obtained. The multi-layer fused features are then passed through a convolutional layer with a kernel size of 3 and a stride of 2 to obtain the overall features. When the features do not meet the target required by the network, the above operation is performed again.

[0044] The T-Conv_1 layer has a stride of 1, a kernel size of 5*5, and 1024 kernels; the T-Conv_2 layer has a stride of 1, a kernel size of 5*5, and 512 kernels; the T-Conv_3 layer has a stride of 1, a kernel size of 5*5, and 256 kernels; the T-Conv_4 layer has a stride of 1, a kernel size of 5*5, and 128 kernels; the T-Conv_5 layer has a stride of 1, a kernel size of 5*5, and 64 kernels; the Conv layer has a stride of 2, a kernel size of 3*3, and 64 kernels.

[0045] This embodiment's efficient feature extraction network structure adds an improved RFB module to the SSD feature extraction network and incorporates an efficient attention mechanism to enhance feature extraction capabilities. In the improved structure, the high-resolution feature map convolutional layers are converted into RFB modules, and the 6th fully connected layer is converted into a downsampling convolutional layer. In this scheme, a single-structure setting is used to simulate eccentricity, and the module parameters are adjusted accordingly based on the receptive field size and eccentricity of the feature map.

[0046] In this embodiment, the forward fusion of multi-scale features in the forward and reverse iterative fusion multi-scale feature network is achieved by performing forward feature fusion on the feature maps in the feature pyramid after the network generates the feature pyramid, i.e., transferring shallow features to deep features.

[0047] Please see Figure 2 In this embodiment, the forward and reverse iterative fusion multi-scale feature network structure, when forward propagating features, aims at N. x The feature map N1 is forward-propagated to feature map N2, and then fused with feature map N1. This process continues forward-propagation to feature map N3, and then fused with feature map N3. This process continues forward-propagation to feature map N4, and then fused with feature map N4. Finally, this process continues forward-propagation to feature map N5, and then fused with feature map N5, generating the target feature layer N. x When backpropagating features, the target is the feature map of layer Ny, which combines feature map N5 and feature map N. x After feature fusion, the features are backpropagated. The feature information of feature map N5 is backpropagated to feature map N4 and fused, then backpropagated to feature map N3, and then fused together with feature map N3. This process is repeated for feature map N2, and then backpropagated to feature map N1, where it is fused together to generate the target feature layer Ny. If Ny does not meet the detection requirements, the above steps are repeated iteratively for feature fusion. In this method, max pooling is used for feature downsampling, and then the number of channels in the downsampled feature map is transformed to be equal to the number of channels in the next level feature map. Specifically, this transformation is performed using a 1×1 convolution. Before feature fusion, the feature values ​​of different feature maps need to be unified to obtain a new feature map. Finally, the feature map is passed to the next layer through a feature fusion function. During the forward feature fusion process, 1×1 convolution is used to fuse the feature maps, enabling the network to perform feature fusion more effectively during autonomous learning.

[0048] Step 3: Using an anchor box matching network based on clustering algorithm, anchor boxes of different categories of targets in the image are clustered and matched according to the new features generated by fusion, so that each target to be detected gets the most suitable anchor box, which is used as the target detection result;

[0049] Please see Figure 3 The anchor box matching network based on clustering algorithm in this embodiment consists of a first convolutional layer, a second convolutional layer, a Canopy allocation layer, and a K-means algorithm aggregation anchor box layer. The first convolutional layer is a 7*7 kernel layer with a stride of 1. The second convolutional layer is a 3*3 kernel layer with a stride of 2. The Canopy allocation layer consists of a third convolutional layer, a Canopy layer, a residual block, and a third deconvolutional layer. The third convolutional layer has a stride of 1 and a kernel size of 5*5. The third inconvolutional layer has a stride of 1 and a kernel size of 5*5; the Canopy layer has a stride of 1, a kernel size of 5*5, and uses ReLU as the activation function; the K-means algorithm aggregated anchor box layer consists of a fourth convolutional layer, a K-means convolutional layer, and a fourth inconvolutional layer; the fourth convolutional layer has a stride of 2 and a kernel size of 5*5; the fourth inconvolutional layer has a stride of 2 and a kernel size of 5*5; the K-means convolutional layer has a stride of 2 and a kernel size of 5*5, and uses ReLU as the activation function.

[0050] The clustering algorithm in this embodiment selects cluster centers for a dataset containing n samples; for each sample in the dataset, the distance to each cluster center is calculated; the distances from each sample to each center are compared sequentially, and samples are assigned to cluster centers based on proximity according to distance, resulting in a series of clusters; the centers of the clusters are calculated and updated, and the centers of the new clusters are the centers of the samples; during clustering, Canopy uses a similarity metric to speed up the clustering process, specifying two distance thresholds t1 and T2 and forming a set U, where T1 > T2; an initial point is selected and added to the set center list C; if any point in the target dataset... With c i If the distance ∈C is greater than T1, then x will be... i Add x as a new subset to the central list C. If the interval is less than T1, then x will be added to the central list C. i Joining with c i In the list centered on x, if the distance between two points is greater than T2, then x will be... i Remove data from the dataset and no longer use it as the center of other subsets until the dataset is empty, allowing for duplicate subclasses.

[0051] Please see Figure 4 In this embodiment, the efficient remote sensing image target feature extraction network, the forward and reverse iterative fusion multi-scale feature network, and the anchor box matching network based on clustering algorithm are all pre-trained networks; the training process includes:

[0052] (1) Create a remote sensing image dataset using remote sensing images, and determine the training dataset and test dataset;

[0053] This embodiment focuses on collecting satellite remote sensing images and using image annotation tools to annotate the attributes of the targets to be detected in the images, including category, location, and shape. The size of the annotated satellite remote sensing images is adjusted to the size of the preset input image to form a satellite remote sensing image dataset. The prepared dataset is divided into a training set and a test set.

[0054] In the solution provided by this embodiment of the invention, the method for determining the image training set and test set is as follows: The collected remote sensing images are divided into a training set and a test set according to a preset ratio of 7:3. The types to be detected in the datasets include aircraft and vehicles, etc. At least one target to be detected, such as a vehicle or aircraft, is set in the images of the training dataset. Then, the type and location of the target to be detected in the images are labeled automatically and manually. At least one target to be detected, such as a vehicle or aircraft, is set in the images of the test dataset. Then, the type and location of the target to be detected in the images are labeled automatically and manually. After completion, the training dataset and the test dataset are preprocessed. Preprocessing includes image resizing, image flipping, and color distortion. After processing, the data is read.

[0055] (2) Preprocess the created dataset;

[0056] In this embodiment, image data is enhanced without losing image information before network training and testing.

[0057] In this embodiment, the relationship between the number of image segments and the resolution after image data enhancement and segmentation is as follows:

[0058]

[0059] Y represents the number of images after segmentation, H represents the height of the unsegmented remote sensing image, W represents the width of the unsegmented remote sensing image, h represents the height of the segmented remote sensing image, w1 represents the width of the segmented remote sensing image, w2 represents the overlap width between the segmented images, and m represents the segmentation coefficient.

[0060] (3) Train the optimized remote sensing image target detection network on the preprocessed dataset;

[0061] The training data is input into the feature extraction network for feature extraction and fusion to obtain the required feature map; the obtained feature map is input into the anchor box matching network based on the clustering algorithm for anchor box matching to obtain the most suitable anchor box; the test data is input into the trained remote sensing image target detection network based on the optimized SSD algorithm, and the detection results are compared with the actual target. If the accuracy value mAP is at its maximum value, the training ends and the trained complex background remote sensing image target detection network is obtained; otherwise, step (3) is returned to the original execution.

[0062] In this embodiment, the pre-constructed detection model is trained using a labeled training dataset within the PyTouch deep learning framework until the network reaches its optimal state. The network parameters at this point are recorded and saved. The trained network model, the saved parameters, and the established detection network are then used to perform target detection on the test data, outputting the category and location information of the detected targets.

[0063] The invention will be further illustrated below through specific experiments.

[0064] Please see Figure 5 The dataset used in the experiment consisted of 5000 images, including self-collected images and the University of Chinese Academy of Sciences High-Resolution Aerial Object Dataset (UCAS-AOD). These images included samples of aircraft and vehicles, which were divided into two sets: one for training and one for testing. All methods used in the experiment were implemented in Python within the PyTouch deep learning framework.

[0065] The experiment compared the detection performance of five different object detection algorithms on the same dataset. The detection results of the present invention and the comparison method are shown in the table below:

[0066]

[0067] The table above shows that the algorithm of this invention significantly improves the detection accuracy of targets in remote sensing images. Among the target detection algorithms presented, the algorithm in this paper has an advantage.

[0068] Experimental results show that the target detection algorithm (SSD(ours)) of this invention has better detection results, proving that the improved algorithm has higher accuracy than the initial algorithm. Compared with several other classic algorithms, the SSD algorithm of this invention has superior detection performance.

[0069] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A remote sensing image target detection method based on an optimized SSD algorithm, characterized in that, Includes the following steps: Step 1: Input the remote sensing image to be detected into the efficient remote sensing image target feature extraction network to extract key features and obtain the required feature map; The efficient remote sensing image target feature extraction network comprises 13 sequentially connected layers. The first layer is a convolutional layer that receives the input data, transforming the resolution of the input image feature map to 300*300*3 with a stride of 2 and a kernel size of 3*3. The second layer is also a convolutional layer, transforming the resolution of the image feature map to 38*38*512 with a stride of 2. An efficient attention mechanism is added to the second layer, comprising global max pooling and global average pooling. After pooling, fully connected layers, ReLU activation functions, and another fully connected layer are added, followed by a Sigmoid function. The function processes the data, then multiplies the channel weights to obtain the result. This result is then convolved and combined with the upsampled features. After processing through an RFB structure, the result is finally input into a Non-Maximum Suppression (NMS) layer for NMS. The third layer is a MaxPooL layer with a stride of 2, outputting a feature map size of 19*19*512. The fourth layer is a Conv_3*3 convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*512. The fifth layer is a convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024. The sixth layer is a double convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024, followed by a convolution to 19*19*256, with the result input into NMS. The seventh layer is an RFB structure layer for enhanced feature extraction, with the result input into NMS. The eighth layer is an RFB (Recursive Feature Buffer) layer, which enhances feature extraction and outputs a feature map with a resolution of 19*19*512. The result is then input into NMS (Non-Maximum Segmentation). The ninth layer is an RFB layer, which enhances feature extraction and outputs a feature map with a resolution of 19*19*512. The result is then input into NMS. The tenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 10*10*128. The eleventh layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 5*5*128. The result is then input into NMS. The twelfth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 3*3*128. The thirteenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 1*1*256. The result is then input into NMS. The output of NMS is the final extracted feature map. The efficient remote sensing image target feature extraction network integrates an efficient attention mechanism, performs global max pooling and global average pooling in parallel, and then extracts information between channels by connecting fully connected layers. Based on the interrelationship between channels, it assigns weights to each channel more reasonably, thereby rapidly improving the extraction of target features. The seventh layer network, in the bottleneck structure, first uses 1x1 convolution to reduce channel features, then adds a layer of convolution, and then uses 1x3, 3x1 convolution and 1x7, 7x1 convolution with simple calculation to reduce the amount of calculation; Step 2: Input the obtained feature map into the forward and reverse iterative fusion multi-scale feature network to perform feature fusion and generate a new feature map; The forward and reverse iterative fusion multi-scale feature network includes T-Conv_1 layer, T-Conv_2 layer, T-Conv_3 layer, T-Conv_4 layer, T-Conv_5 layer and Conv layer. It is used to process the features T1, T2, T3, T4 and T5 output by the efficient remote sensing image target feature extraction network through a network and cascade to make all features have the same feature information scale. Then, in the order of T5, T4, T3, T2 and T1, they are passed through the upper network and cascaded for reverse fusion operation. Finally, forward and reverse fused multi-scale features are obtained. The multi-layer fused features are then passed through a convolutional layer with a kernel size of 3 and a stride of 2 to obtain the overall features. The T-Conv_1 layer has a stride of 1, a kernel size of 5*5, and 1024 kernels; the T-Conv_2 layer has a stride of 1, a kernel size of 5*5, and 512 kernels; the T-Conv_3 layer has a stride of 1, a kernel size of 5*5, and 256 kernels; the T-Conv_4 layer has a stride of 1, a kernel size of 5*5, and 128 kernels; the T-Conv_5 layer has a stride of 1, a kernel size of 5*5, and 64 kernels; and the Conv layer has a stride of 2, a kernel size of 3*3, and 64 kernels. Step 3: Using a clustering-based anchor box matching network, anchor boxes of different categories of targets in the image are clustered and matched according to the new features generated by fusion, so that each target to be detected gets the most suitable anchor box, which is used as the target detection result. The anchor box matching network based on the clustering algorithm consists of a first convolutional layer, a second convolutional layer, a Canopy allocation layer, and a K-means algorithm aggregation anchor box layer. The first convolutional layer is a 7*7 kernel layer with a stride of 1. The second convolutional layer is a 3*3 kernel layer with a stride of 2. The Canopy allocation layer consists of a third convolutional layer, a Canopy layer, a residual block, and a third deconvolutional layer. The third convolutional layer has a stride of 1 and a kernel size of 5*5. The third deconvolutional layer has a stride of 1 and a kernel size of 5*5; the Canopy layer has a stride of 1, a kernel size of 5*5, and a ReLU activation function; the K-means algorithm aggregated anchor box layer consists of a fourth convolutional layer, a K-means convolutional layer, and a fourth deconvolutional layer; wherein the fourth convolutional layer has a stride of 2 and a kernel size of 5*5; the fourth deconvolutional layer has a stride of 2 and a kernel size of 5*5; the K-means convolutional layer has a stride of 2 and a kernel size of 5*5, and a ReLU activation function. 2.The method of claim 1, wherein: In step 1, the high-efficiency attention function is: In the formula, Here, Q represents the fully connected layer function, and Q represents the fully connected coefficients. This represents the activation function. Represents the correlation coefficient. m Let m be the m-th feature. 3.The method of claim 1, wherein: In step 1, the interrelation between each feature channel and adjacent feature channels in each channel is extracted by global average pooling and a fully connected layer, a global average pooling operation is performed, and the required global receptive field is obtained ; wherein, i represents a horizontal parameter in the feature matrix, j represents a vertical parameter in the feature matrix, H represents a maximum value of a feature map height, and W represents a maximum value of a feature map width, W j represents a weight value of each channel, k is a number of convolution kernels in the convolution layer, represents a vertical and horizontal parameter matrix in each channel. 4.The method of claim 1, wherein: The clustering algorithm described in step 3 is to select the clustering center for a data set containing n samples; for each sample in the data set, the distance to each clustering center is calculated respectively; The distance from each sample to each cluster center is compared sequentially, and then samples are assigned to cluster centers based on proximity, resulting in a series of clusters. The centers of these clusters are calculated and updated, and the centers of the new clusters become the centers of the samples. During clustering, Canopy uses a similarity metric to speed up the process, specifying two distance thresholds. And form a set ,in, Select an initial point and add it to the set center list. If any point in the target dataset and The distance is greater than , then Added to the central list as a new subset In the middle, if the spacing is less than Then Joining In the list centered on, if the distance between the two is greater than Then Remove data from the dataset and no longer use it as the center of other subsets until the dataset is empty, allowing for duplicate subclasses.

5. The method of claim 1, wherein the method further comprises: The efficient remote sensing image target feature extraction network, the forward and reverse iterative fusion multi-scale feature network, and the anchor box matching network based on clustering algorithm are all pre-trained networks. During training, the dataset is first created and preprocessed, and the training and test datasets are determined. The preprocessing is to enhance the image data without losing image information, that is, to perform image data segmentation and enhancement processing. The relationship between the number of image segments and resolution after image data enhancement and segmentation is as follows: wherein, Y represents the number of cut pictures, H represents the height of the remote sensing image before cutting, W represents the width of the remote sensing image before cutting, h represents the height of the remote sensing small image after cutting, w 1 represents the width of the remote sensing small image after cutting, w 2 represents the overlap width between the segmented images, m represents the cutting coefficient.

6. A remote sensing image target detection system based on an optimized SSD algorithm, characterized in that, Includes the following modules: The first module is used to input the remote sensing image to be detected into the efficient remote sensing image target feature extraction network to extract key features and obtain the required feature map. The efficient remote sensing image target feature extraction network comprises 13 sequentially connected layers. The first layer is a convolutional layer that receives the input data, transforming the resolution of the input image feature map to 300*300*3 with a stride of 2 and a kernel size of 3*3. The second layer is also a convolutional layer, transforming the resolution of the image feature map to 38*38*512 with a stride of 2. An efficient attention mechanism is added to the second layer, comprising global max pooling and global average pooling. After pooling, fully connected layers, ReLU activation functions, and another fully connected layer are added, followed by a Sigmoid function. The function processes the data, then multiplies the channel weights to obtain the result. This result is then convolved and combined with the upsampled features. After processing through an RFB structure, the result is finally input into a Non-Maximum Suppression (NMS) layer for NMS. The third layer is a MaxPooL layer with a stride of 2, outputting a feature map size of 19*19*512. The fourth layer is a Conv_3*3 convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*512. The fifth layer is a convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024. The sixth layer is a double convolutional layer with a stride of 1, resulting in a feature map resolution of 19*19*1024, followed by a convolution to 19*19*256, with the result input into NMS. The seventh layer is an RFB structure layer for enhanced feature extraction, with the result input into NMS. The eighth layer is an RFB (Recursive Feature Buffer) layer, which enhances feature extraction and outputs a feature map with a resolution of 19*19*512. The result is then input into NMS (Non-Maximum Segmentation). The ninth layer is an RFB layer, which enhances feature extraction and outputs a feature map with a resolution of 19*19*512. The result is then input into NMS. The tenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 10*10*128. The eleventh layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 5*5*128. The result is then input into NMS. The twelfth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 3*3*128. The thirteenth layer is a convolutional layer with a kernel size of 3*3 and a stride of 1, outputting a feature map with a resolution of 1*1*256. The result is then input into NMS. The output of NMS is the final extracted feature map. The efficient remote sensing image target feature extraction network integrates an efficient attention mechanism, performs global max pooling and global average pooling in parallel, and then extracts information between channels by connecting fully connected layers. Based on the interrelationship between channels, it assigns weights to each channel more reasonably, thereby rapidly improving the extraction of target features. The seventh layer network, in the bottleneck structure, first uses 1x1 convolution to reduce channel features, then adds a layer of convolution, and then uses 1x3, 3x1 convolution and 1x7, 7x1 convolution with simple calculation to reduce the amount of calculation; The second module is used to input the obtained feature map into a forward and reverse iterative fusion multi-scale feature network for feature fusion to generate a new feature map. The forward and reverse iterative fusion multi-scale feature network includes T-Conv_1 layer, T-Conv_2 layer, T-Conv_3 layer, T-Conv_4 layer, T-Conv_5 layer and Conv layer. It is used to process the features T1, T2, T3, T4 and T5 output by the efficient remote sensing image target feature extraction network through a network and cascade to make all features have the same feature information scale. Then, in the order of T5, T4, T3, T2 and T1, they are passed through the upper network and cascaded for reverse fusion operation. Finally, forward and reverse fused multi-scale features are obtained. The multi-layer fused features are then passed through a convolutional layer with a kernel size of 3 and a stride of 2 to obtain the overall features. The T-Conv_1 layer has a stride of 1, a kernel size of 5*5, and 1024 kernels; the T-Conv_2 layer has a stride of 1, a kernel size of 5*5, and 512 kernels; the T-Conv_3 layer has a stride of 1, a kernel size of 5*5, and 256 kernels; the T-Conv_4 layer has a stride of 1, a kernel size of 5*5, and 128 kernels; the T-Conv_5 layer has a stride of 1, a kernel size of 5*5, and 64 kernels; and the Conv layer has a stride of 2, a kernel size of 3*3, and 64 kernels. The third module is used to use a clustering-based anchor box matching network to cluster and match anchor boxes for different categories of targets in the image based on the new features generated by fusion, so that each target to be detected gets the most suitable anchor box as the target detection result. The anchor box matching network based on the clustering algorithm consists of a first convolutional layer, a second convolutional layer, a Canopy allocation layer, and a K-means algorithm aggregation anchor box layer. The first convolutional layer is a 7*7 kernel layer with a stride of 1. The second convolutional layer is a 3*3 kernel layer with a stride of 2. The Canopy allocation layer consists of a third convolutional layer, a Canopy layer, a residual block, and a third deconvolutional layer. The third convolutional layer has a stride of 1 and a kernel size of 5*5. The third deconvolutional layer has a stride of 1 and a kernel size of 5*5; the Canopy layer has a stride of 1, a kernel size of 5*5, and a ReLU activation function; the K-means algorithm aggregated anchor box layer consists of a fourth convolutional layer, a K-means convolutional layer, and a fourth deconvolutional layer; wherein the fourth convolutional layer has a stride of 2 and a kernel size of 5*5; the fourth deconvolutional layer has a stride of 2 and a kernel size of 5*5; the K-means convolutional layer has a stride of 2 and a kernel size of 5*5, and a ReLU activation function.

Images