CN109325947A - A SAR image tower target detection method based on deep learning - Google Patents

Abstract

The invention relates to a SAR image iron tower target detection method based on deep learning. Each sample slice in the sample set is processed to generate multiple different artificial sample slices, and the training sample set is added to achieve expansion after adding the sample label; the SSD model is constructed; the expanded training sample set is input into the constructed SSD model , use the gradient descent method to train the SSD model; cut the data graph to be tested into multiple slices to be detected with the same size as the sample slice, input the trained SSD model, and obtain the target detection result of the data graph to be tested. The invention has the advantages of strong robustness, fast running speed, high detection performance and easy migration, no need for high contrast between the target and the background during detection, and can be used for detection in complex scenes.

Description

A SAR image tower target detection method based on deep learning

technical field

The invention relates to the technical field of image processing, in particular to a deep learning-based SAR image tower target detection method.

Background technique

Spaceborne Synthetic Aperture Radar (SAR) belongs to a microwave imaging radar. It is widely used in many fields such as monitoring, resource detection, agriculture and forestry.

Convolutional Neural Network (CNN) is a commonly used deep, pre-feedback artificial neural network, which is a kind of deep learning method and has been successfully applied in the field of computer vision. With the recovery of deep learning technology, object detection methods based on deep learning have developed rapidly in recent years. The R-CNN method based on candidate regions, its improved versions FastR-CNN method and Faster R-CNN, as well as the YOLO method and SSD method based on the end-to-end idea have been proposed successively and have been widely used. However, the real-time performance of the Faster R-CNN model is poor, and the accuracy of the YOLO model is poor.

At present, many target detection methods for SAR images have been developed. Among them, the constant false alarm rate CFAR detection method is widely used in SAR image target detection because of its simplicity, rapidity and strong real-time performance. According to different types of targets, there are different representation forms on SAR images, and correspondingly different detection methods. However, these existing SAR image detection methods only utilize the statistical characteristics of the local area of the SAR image, and can only achieve pixel-level detection, requiring a high contrast between the target and the background, and the detection performance is good in simple scenes, but in complex The detection performance in the scene is poor.

SUMMARY OF THE INVENTION

(1) Technical problems to be solved

The technical problem to be solved by the present invention is to solve the problem that the target detection method for SAR images in the prior art requires a high contrast between the target and the background, and the detection performance is poor in complex scenes.

(2) Technical solutions

In order to solve the above technical problems, the present invention provides a SAR image tower target detection method based on deep learning to improve the detection performance in complex scenes, including:

S1. Randomly extract multiple SAR images from the SAR data set, segment to obtain sample slices, set a sample label for each sample slice to construct a training sample set, and the sample label includes the coordinate information and category information of the target in the sample slice;

S2, processing each sample slice in the training sample set to generate a plurality of different artificial sample slices, and adding the training sample set to realize expansion after adding the sample label;

S3. Build the SSD model, including:

A 15-layer convolutional neural network is used to initially extract the image features of the input image;

8-layer convolutional neural network for further extracting image features of different scales;

The multi-scale detection network detects the extracted image features of different scales;

S4, using the expanded training sample set in the step S2 as an input image and inputting the SSD model constructed in the step S3, and using the gradient descent method to train the SSD model;

S5. Cut the data graph to be tested into a plurality of slices to be detected with the same size as the sample slice, input the SSD model trained in step S4, and obtain the target detection result of the data graph to be tested.

Preferably, the step S1 includes:

S1-1. Randomly extract 100 SAR images from the MiniSAR dataset;

S1-2. Randomly divide and obtain multiple sample slices from each SAR image;

S1-3. Add a sample label to each sample slice, and the sample label includes the absolute path of the sample slice, the number of targets, the coordinates of the target frame, and the category of the target;

S1-4. Each of the sample slices with corresponding sample labels is formed into a training sample set.

Preferably, in the step S2, each sample slice in the training sample set is processed, and when multiple different artificial sample slices are generated, one or more methods of translation, flipping, rotation, and adding noise are randomly used to perform the processing. processing to generate 100 different artificial sample slices from each sample slice transformation.

Preferably, the step S4 includes:

S4-1. Input the expanded training sample set into the SSD model;

S4-2, use the forward propagation method to calculate the cost function of the input image to the current SSD model;

S4-3. Using the back-propagation method, respectively calculate the gradient values of the cost function for the parameters in each convolutional layer in the SSD model;

S4-4. Using the gradient descent method, update the parameters in each convolution layer according to the gradient value of the cost function for the parameters in each convolution layer;

S4-5. Repeat steps S4-2 to S4-4 to update the SSD model. When the number of cycles reaches the set value, or the parameters of each convolutional layer in the SSD model are no longer updated, the training is completed and the current SSD model is saved .

Preferably, the step S5 includes:

S5-1, cutting the data graph to be tested into a plurality of slices with a size similar to the sample slice in a grid shape, as the slices to be detected;

S5-2. Input the slice to be detected into the SSD model trained in step S4, and use the SSD model to perform target detection on each slice to be detected, and obtain the target detection result of each slice to be detected;

S5-3. Merge the target detection results of each slice to be detected at the corresponding positions of the original data map to be tested to obtain the target detection results of the data map to be tested.

Preferably, the size of the input image of the SSD model constructed in the step S3 is 300×300 pixels;

The 15-layer convolutional neural network includes:

The first group of convolutional layers: 1-2 layers of convolutional layers, each layer uses 64 convolution kernels with a window size of 3 × 3, the window movement step size is 1, the edge is filled with 1 pixel, and the output size is 64 300×300 feature map;

The second group of convolutional layers: 3-4 layers of convolutional layers, each layer uses 128 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 128 150×150 feature map;

The third group of convolutional layers: 5-7 layers of convolutional layers, each layer uses 256 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 256 75×75 feature map;

The fourth group of convolutional layers: 8-10 layers of convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 38×38 feature map;

The fifth group of convolutional layers: the 11-13th convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 19×19 feature map;

The sixth group of convolutional layers: the 14th convolutional layer, using 1024 convolution kernels with a window size of 3 × 3, a window shift step size of 1, and edge padding of 1 pixel, outputting 1024 features of size 19 × 19 picture;

The seventh group of convolutional layers: the 15th convolutional layer, using 1024 convolution kernels with a window size of 1×1, a window movement step size of 1, no edge padding, and outputting 1024 feature maps with a size of 19×19 ;

Among them, the feature maps output by the 2nd, 4th, 7th, 10th, and 13th convolutional layers also need to go through the largest downsampling layer for dimensionality reduction, using a downsampling kernel with a window size of 2 × 2, and a window moving step size of 2.

Preferably, the 8-layer convolutional neural network in step S3 includes:

The eighth group of convolutional layers: the 16-17th convolutional layer, the 16th convolutional layer uses 256 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer Use 512 convolution kernels with a window size of 3 × 3, a window movement step size of 2, and edge padding of 1 pixel; output 512 feature maps with a size of 10 × 10;

The ninth group of convolutional layers: 18-19 layers of convolutional layers, the 18th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer uses 256 A convolution kernel with a window size of 3×3, a window movement step size of 2, and edge padding of 1 pixel; output 256 feature maps with a size of 5×5;

The tenth group of convolutional layers: 20-21 layers of convolutional layers, the 20th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 21st layer uses 256 A convolution kernel with a window size of 3×3, a window moving step size of 1, and no edge padding; output 256 feature maps with a size of 3×3;

The eleventh group of convolutional layers: 22-23 layers of convolutional layers, the 22nd layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 23rd layer uses 256 A convolution kernel with a window size of 3 × 3, the window movement step size is 1, and the edges are not filled; 256 feature maps with a size of 1 × 1 are output.

Preferably, the multi-scale detection network in the step S3 performs the following operations on the feature maps output by the 10th, 15th, 17th, 19th, 21st, and 23rd convolutional layers:

Normalize the feature map;

Use a convolution kernel with 4k windows of size 3 × 3, the window movement step size is 1, and the edge is filled with a convolution layer of 1 pixel, and the coordinate information of k default borders is extracted; among them, k=1, 2, 3 , 4, 5, and 6 correspond to the 10th, 15th, 17th, 19th, 21st, and 23rd floors respectively;

Use a convolution kernel with (classes+1)×k window size of 3×3, window shift step size 1, edge padding 1 pixel convolution layer, and extract the confidence for each class of k default borders Score, where classes is the number of categories of the target to be detected;

The coordinate information and confidence scores extracted from the default borders of the feature maps of each layer are spliced together to generate detection results.

Preferably, for the 10th, 15th, 17th, 19th, 21st, and 23rd layers, the default frame size values are sequentially s ₁ to s ₆ in the following formulas:

In the above formula, s _k is the size value of the default frame, s _min =0.2, s _max =0.9;

For each layer of feature maps, set different aspect ratios a _r , and combine the layer's size value _sk to get the width of the default border and height They are:

Among them, for layers 10, 21, and 23, a _r ∈ {1, 2, 1/2}; for layers 15, 17, and 19, a _r ∈ {1, 2, 3, 1/2, 1/3} , and for the above six convolutional layers, there is also a corresponding size

Preferably, in the step S3, when the multi-scale detection network splices the coordinate information and confidence scores extracted from each default frame of the feature map of each layer to generate the detection results, first discard all results whose confidence scores are lower than the confidence threshold; then Using the non-maximum suppression method, while retaining the optimal detection results with higher confidence scores, the sub-optimal results are suppressed; finally, the filtered coordinate information and confidence scores are spliced.

(3) Beneficial effects

The above technical solution of the present invention has the following advantages: the present invention provides a deep learning-based SAR image tower target detection method, using the SAR image to train the SSD model, and detecting the iron tower target in the SAR image through the SSD model. Compared with the prior art, the present invention has the following advantages:

(1) Strong robustness

The invention adopts a multi-layer convolutional neural network model, which can fully extract the high-level features of the target and make full use of the translation invariance of the convolution layer, so the invention has strong robustness to the translation of SAR images. At the same time, the present invention adopts a multi-aspect ratio detection method of multi-scale feature maps, so the present invention also has strong robustness to the deformation of the SAR image.

(2) Fast running speed

Both the traditional CFAR detection method and the Faster R-CNN model need to go through two steps of generating suspicious targets and identifying suspicious targets, and the detection efficiency is low. The invention adopts an end-to-end training and detection method, integrates generation and identification, and improves the speed of training and detection.

(3) High detection performance

The traditional CFAR detection method needs to set parameters according to the prior information in the image to be tested. If the parameter settings are unreasonable, it will seriously affect the detection results. The network model in the present invention fully considers the objectives of multiple aspect ratios of each multi-scale feature map, and at the same time avoids the problem of unreasonable parameter settings in the CFAR detection method, thereby improving the accuracy.

(4) Easy migration

Lower layer parameters in a trained convolutional neural network reflect underlying features common to images, such as edges, shapes, etc. Therefore, when a model for detecting other objects needs to be trained, the training can be completed faster based on transfer learning techniques.

Description of drawings

1 is a schematic diagram of steps of a method for detecting a target in a SAR image iron tower based on deep learning in Embodiment 1 of the present invention;

2 is a schematic diagram of an SSD model in Embodiment 2 of the present invention;

Fig. 3 is a graph of data to be measured in the third embodiment of the present invention;

FIG. 4a and FIG. 4b are partial enlarged views of the detection result of FIG. 3 .

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

Example 1

As shown in FIG. 1 , a deep learning-based SAR image tower target detection method provided by an embodiment of the present invention includes:

S1. Randomly extract multiple SAR images from the SAR data set, obtain sample slices by segmentation, and set up a sample label for each sample slice to construct a training sample set. The sample label includes coordinate information and category information of the target in the sample slice.

Preferably, step S1 includes:

S1-1. Randomly extract 100 SAR images from the MiniSAR dataset;

S1-2. Randomly divide and obtain multiple sample slices from each SAR image;

S1-3. Add a sample label to each sample slice. The sample label includes the absolute path of the sample slice, the number of targets, the coordinates of the target frame (ie coordinate information) and the target category (ie category information); specifically, the sample label can be composed of:

<DIR>nx _t1 y _t1 x _b1 y _b1 c ₁ x _t2 y _t2 x _b2 y _b2 c ₂ …x _tn y _tn x _bn y _bn c _n

Among them, <DIR> represents the absolute path of the corresponding sample slice, n represents the number of tower targets in the slice; followed by n groups of target information, each group of target information includes (x _ti , y _ti ), (x _bi , y _bi ) ) and c _i : x _ti and y _ti represent the coordinates of the upper left corner of the ith target frame, x _bi and y _bi represent the coordinates of the lower right corner of the _ith target frame, and ci represent the category of the ith target frame.

S1-4. Each of the sample slices with corresponding sample labels is formed into a training sample set.

S2. Process each sample slice in the training sample set to generate a plurality of different artificial sample slices, and add the training sample set after adding sample labels to achieve expansion.

Preferably, in step S2, each sample slice in the training sample set is processed, and when multiple different artificial sample slices are generated, one or more methods of translation, flip, rotation, and noise addition are randomly used for processing, From each sample slice transformation, 100 different artificial sample slices are generated, and the sample labels corresponding to the artificial sample slices are added according to the specific operation, and each artificial sample slice with the corresponding sample label is added to the training sample set to obtain the expansion After the training sample set.

S3. Build the SSD model, including:

A 15-layer convolutional neural network is used to initially extract the image features of the input image; the construction can be based on the VGG-16 model network;

The 8-layer convolutional neural network is connected behind the 15-layer convolutional neural network to further extract image features of different scales;

The multi-scale detection network is connected behind the 8-layer convolutional neural network to detect the extracted image features of different scales.

S4. Input the expanded training sample set in the step S2 as an input image to the SSD model constructed in the step S3, and use the gradient descent method to train the SSD model.

Preferably, step S4 includes:

S4-1. Input the expanded training sample set into the SSD model;

S4-2, use the forward propagation method to calculate the cost function of the input image to the current SSD model;

S4-3. Using the back-propagation method, respectively calculate the gradient values of the cost function for the parameters in each convolutional layer in the SSD model;

S4-4. Using the gradient descent method, update the parameters in each convolution layer according to the gradient value of the cost function for the parameters in each convolution layer;

S4-5. Perform the steps S4-2 to S4-4 in a loop to update the SSD model. When the number of cycles reaches the set value, or the parameters of each convolutional layer in the SSD model are no longer updated, the training is completed and the current state is saved. SSD model.

S5. Cut the data graph to be tested into a plurality of slices to be detected with the same or similar size as the sample slice, input the SSD model trained in step S4, and obtain the target detection result of the data graph to be tested. Preferably, the sample slice should be a square or a rectangle similar to a square, such as a rectangle with an aspect ratio not greater than 3/2, and the size of the slice to be detected should be similar to the size of the sample slice during training, such as two slices (sample slice and The ratio between the long sides of the slice to be detected) is 2/3 to 3/2.

Preferably, step S5 includes:

S5-1, cutting the data graph to be tested into a plurality of slices with a size similar to the sample slice in a grid shape, as the slices to be detected;

S5-2. Input the slice to be detected into the SSD model trained in step S4, and use the SSD model to perform target detection on each slice to be detected, and obtain the target detection result of each slice to be detected;

S5-3. Merge the target detection results of each slice to be detected at the corresponding positions of the original data map to be tested to obtain the target detection results of the data map to be tested.

The method provided by the present invention adopts the method of extracting target features through deep learning of a multi-layer convolutional neural network model to realize the detection of the iron tower target in the SAR image, without the need for a high contrast between the target and the background, and uses the multi-scale feature map for multi-aspect ratio detection. The method extracts features at multiple scales, has strong robustness to the deformation of SAR images, and can detect complex scenes. And the generation of suspicious targets and the identification of suspicious targets are integrated, and the speed of training and detection is fast.

Embodiment 2

As shown in Figure 2, the second embodiment is basically the same as the first embodiment, the similarities will not be repeated, and the differences are:

The input image size of the SSD model constructed in step S3 is 300 × 300 pixels.

FIG. 2 is a schematic diagram of the SSD model in this embodiment, and the key convolution layers are marked in the figure. Among them, the convolutional layer 4_3 represents the third convolutional layer in the fourth group of convolutional layers, that is, the tenth convolutional layer, the same below.

As shown in Figure 2, the specific structure of the 15-layer convolutional neural network constructed in step S3 includes:

The first group of convolutional layers: 1-2 layers of convolutional layers, each layer uses 64 convolution kernels with a window size of 3 × 3, the window movement step size is 1, the edge is filled with 1 pixel, and the output size is 64 300×300 feature map;

The second group of convolutional layers: 3-4 layers of convolutional layers, each layer uses 128 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 128 150×150 feature map;

The third group of convolutional layers: 5-7 layers of convolutional layers, each layer uses 256 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 256 75×75 feature map;

The fourth group of convolutional layers: 8-10 layers of convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 38×38 feature map;

The fifth group of convolutional layers: the 11-13th convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 19×19 feature map;

The sixth group of convolutional layers: the 14th convolutional layer, using 1024 convolution kernels with a window size of 3 × 3, a window shift step size of 1, and edge padding of 1 pixel, outputting 1024 features of size 19 × 19 picture;

The seventh group of convolutional layers: the 15th convolutional layer, using 1024 convolution kernels with a window size of 1×1, a window movement step size of 1, no edge padding, and outputting 1024 feature maps with a size of 19×19 ;

Among them, the feature maps output by the 2nd, 4th, 7th, 10th, and 13th convolutional layers also need to go through the largest downsampling layer for dimensionality reduction, using a downsampling kernel with a window size of 2 × 2, and a window moving step size of 2.

The specific structure of the 8-layer convolutional neural network in step S3 includes:

The eighth group of convolutional layers: the 16-17th convolutional layer, the 16th convolutional layer uses 256 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer Use 512 convolution kernels with a window size of 3 × 3, a window movement step size of 2, and edge padding of 1 pixel; the eighth group overall outputs 512 feature maps with a size of 10 × 10;

The ninth group of convolutional layers: 18-19 layers of convolutional layers, the 18th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer uses 256 A convolution kernel with a window size of 3×3, a window movement step size of 2, and edge padding of 1 pixel; the ninth group outputs 256 feature maps with a size of 5×5 as a whole;

The tenth group of convolutional layers: 20-21 layers of convolutional layers, the 20th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 21st layer uses 256 The convolution kernel with a window size of 3×3, the window moving step size is 1, and the edges are not filled; the tenth group outputs 256 feature maps with a size of 3×3 as a whole;

The eleventh group of convolutional layers: 22-23 layers of convolutional layers, the 22nd layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 23rd layer uses 256 A convolution kernel with a window size of 3 × 3, a window movement step size of 1, and no edge padding; the eleventh group outputs 256 feature maps with a size of 1 × 1 as a whole.

The multi-scale detection network in step S3 performs the following operations on the feature maps output by the 10th, 15th, 17th, 19th, 21st, and 23rd convolutional layers respectively:

a) Normalize the feature map.

b) Use a convolution kernel with 4k windows of size 3×3, the window movement step size is 1, and the edge is filled with a convolution layer of 1 pixel, and the coordinate information of k default borders is extracted; among them, k=1, 2 , 3, 4, 5, and 6 correspond to the 10th, 15th, 17th, 19th, 21st, and 23rd floors respectively.

c) Use a convolution kernel with (classes+1)×k window size of 3×3, the window shift step size is 1, and the edge padding is 1 pixel convolution layer, extracting k default borders for each class The confidence score of , where classes is the number of categories of the target to be detected, that is, the total number of categories of each target in the training sample set.

The default border size _sk and aspect ratio a _r are hyperparameters. For the 10th, 15th, 17th, 19th, 21st, and 23rd layers, the default border size values are in the following formulas s ₁ to s ₆ :

In the above formula, s _k is the size value of the default frame, s _min = 0.2, s _max = 0.9, that is, the size of the default frame in the feature map of the 10th layer of the lowest layer is 0.2, and the value of the default frame of the feature map of the 23rd layer of the highest layer is 0.2. The size value of the default border is 0.9, and the size values of the default borders in the feature maps of each layer are evenly spaced.

For each layer of feature maps, set different aspect ratios a _r , and combine the layer's size value _sk to get the width of the default border and height They are:

Among them, for layers 10, 21, and 23, a _r ∈ {1, 2, 1/2}; for layers 15, 17, and 19, a _r ∈ {1, 2, 3, 1/2, 1/3} , and for the above-mentioned six convolutional layers of the 10th, 15th, 17th, 19th, 21st, and 23rd layers, consider one more size Therefore, for the feature maps of the 10th, 21st, and 23rd layers, a total of 4 different default borders can be generated for each position; for the feature maps of the 15th, 17th, and 19th layers, a total of 6 different default borders can be generated for each position. the default border. This setup basically covers tower targets of various shapes and sizes in the input image.

d) Concatenate the coordinate information and confidence scores extracted from each default frame of the feature map of each layer to generate a detection result. Preferably, four coordinate information and one confidence score are grouped together to represent one detection result, and the corresponding relationship between the coordinate information and the confidence score is maintained during splicing, and multiple detection results are spliced.

Preferably, all results whose confidence scores are lower than the confidence threshold (0.1 in the present invention) are first discarded; then the non-maximum suppression method (NMS) is used to retain the optimal detection results with higher confidence scores while suppressing the suboptimal detection results. Results; coordinate information and confidence scores after final stitching screening. The non-maximum value suppression method can prevent the situation that the same target is detected multiple times. Preferably, in the present invention, the intersection and union ratio IoU(AB) is used as the basis to judge whether the overlap rate of the two regions is too high, if:

Then it is considered that the overlap rate of area A and area B is too high.

Further preferably, using the forward propagation method in step S4-2, calculating the cost function of the input image to the current SSD model includes:

Match target borders and default borders based on the intersection ratio: first match each target border with the default border with the largest intersection ratio with it, and secondly match the default border with any intersection with it above a threshold (the present invention 0.5), which simplifies the learning process and better handles predictions between overlapping objects.

Assume Indicates whether there is a match between the ith default border and the jth target border for category p ( means that the ith default bounding box matches the jth target bounding box of category p, Indicates that the i-th default frame does not match the j-th target frame of category p), as can be seen from the above matching strategy, That is, one target border may match more than one default border.

Preferably, the cost function, that is, the overall loss function can be expressed by the following formula:

where x is c is the set of classification results predicted by the SSD model network; l is the predicted bounding box, that is, the set of positioning results predicted by the SSD model network; g is the information set of the target bounding box. N is the number of matching default borders. If N=0, the loss function is set to 0, and the weight parameter α is set to 1 in this method. L _conf (x, c) is the classification loss sub-item in the overall loss function, and L _loc (x, l, g) is the localization loss sub-item in the overall loss function.

The classification loss function L _conf (x,c) is the softmax loss considering all classifications, which can be expressed by the following formula:

in, Represents the output classification result of the ith default border for category p, Represents the confidence probability obtained by softmaxing the i-th default frame on the output classification results of different categories p.

The localization loss function L _loc (x,l,g) is the Smooth L1 loss between the predicted bounding box and the target bounding box, which can be expressed by the following formula:

in, represents the x-coordinate of the ith predicted bounding box, represents the y-coordinate of the ith predicted bounding box, represents the width of the ith predicted bounding box, Represents the height of the i-th predicted frame; Represents the x-coordinate of the j-th target frame after degradation, represents the y-coordinate of the degenerate j-th target frame, represents the width of the j-th target border after degradation, Represents the height of the j-th target frame after degradation; the Smooth L1 loss can be expressed by the following formula:

The coordinates of the target bounding box are degenerated using the coordinates of the default bounding box to:

in, represents the x-coordinate of the jth target border, represents the y-coordinate of the jth target border, represents the width of the jth target border, represents the height of the ith target border, represents the x-coordinate of the ith default border, represents the y-coordinate of the ith default border, represents the width of the ith default border, Indicates the height of the ith default border.

In the method provided by the present invention, the multi-aspect ratio target of each multi-scale feature map is fully considered when constructing the network model, and different aspect ratios are set for lower layer parameters and higher layer parameters, while avoiding the CFAR detection method. The problem of unreasonable parameter settings in the medium is improved, thereby improving the accuracy. Also, the lower layer parameters in the trained convolutional neural network reflect the underlying features common to the images, such as edges, shapes, etc. Therefore, when a model for detecting other objects needs to be trained, the training can be completed faster based on transfer learning techniques.

Embodiment 3

As shown in Figure 3, Figure 4a and Figure 4b, the third embodiment is basically the same as the second embodiment, the similarities will not be repeated, and the differences are:

The data image to be measured used in this embodiment is a scene SAR image, as shown in FIG. 3 . The image size of the data map to be tested is 16384×8192 pixels, which includes various artificial targets such as towers and buildings. The purpose is to detect and locate all types of tower targets in the data map to be tested.

In order to better observe the detection results, as shown in Figures 4a and 4b, the part of the target detection result is enlarged, and the part marked with a white frame in the figure is the part that is detected as an iron tower.

In this embodiment, by calculating the detection results of the entire data map to be tested, it is obtained that the method provided by the present invention has an accuracy rate of 82.4%, a recall rate (recall rate) of 97.6%, and a comprehensive evaluation index F1 value (precise value and The harmonic mean of recall) is 89.4%, which achieves better test results.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. a SAR image tower target detection method based on deep learning, is characterized in that, comprises: S1. Randomly extract multiple SAR images from the SAR data set, segment to obtain sample slices, set a sample label for each sample slice to construct a training sample set, and the sample label includes the coordinate information and category information of the target in the sample slice; S2, processing each sample slice in the training sample set to generate a plurality of different artificial sample slices, and adding the training sample set to realize expansion after adding the sample label; S3. Build the SSD model, including: A 15-layer convolutional neural network is used to initially extract the image features of the input image; 8-layer convolutional neural network for further extracting image features of different scales; The multi-scale detection network detects the extracted image features of different scales; S4, using the expanded training sample set in the step S2 as an input image and inputting the SSD model constructed in the step S3, and using the gradient descent method to train the SSD model; S5. Cut the data graph to be tested into a plurality of slices to be detected with the same size as the sample slice, input the SSD model trained in step S4, and obtain the target detection result of the data graph to be tested. 2. The SAR image iron tower target detection method based on deep learning according to claim 1, is characterized in that, described step S1 comprises: S1-1. Randomly extract 100 SAR images from the MiniSAR dataset; S1-2, randomly dividing each SAR image to obtain multiple sample slices; S1-3. Add a sample label to each sample slice, and the sample label includes the absolute path of the sample slice, the number of targets, the coordinates of the target frame, and the category of the target; S1-4. Each of the sample slices with corresponding sample labels is formed into a training sample set. 3. the SAR image iron tower target detection method based on deep learning according to claim 1, is characterized in that: In the step S2, each sample slice in the training sample set is processed, and when multiple different artificial sample slices are generated, one or more methods of translation, flipping, rotation, and noise addition are randomly used for processing, from Each sample slice transformation generates 100 different artificial sample slices. 4. the SAR image tower target detection method based on deep learning according to claim 1, is characterized in that, described step S4 comprises: S4-1. Input the expanded training sample set into the SSD model; S4-2, use the forward propagation method to calculate the cost function of the input image to the current SSD model; S4-3. Using the back-propagation method, respectively calculate the gradient values of the cost function for the parameters in each convolutional layer in the SSD model; S4-4. Using the gradient descent method, update the parameters in each convolution layer according to the gradient value of the cost function for the parameters in each convolution layer; S4-5. Repeat steps S4-2 to S4-4 to update the SSD model. When the number of cycles reaches the set value, or the parameters of each convolutional layer in the SSD model are no longer updated, the training is completed and the current SSD model is saved . 5. SAR image tower target detection method based on deep learning according to claim 1, is characterized in that, described step S5 comprises: S5-1, cutting the data graph to be tested into a plurality of slices with a size similar to the sample slice in a grid shape, as the slices to be detected; S5-2. Input the slice to be detected into the SSD model trained in step S4, and use the SSD model to perform target detection on each slice to be detected, and obtain the target detection result of each slice to be detected; S5-3. Merge the target detection results of each slice to be detected at the corresponding positions of the original data map to be tested to obtain the target detection results of the data map to be tested. 6. The deep learning-based SAR image tower target detection method according to any one of claims 1-5, wherein the SSD model input image size constructed in the step S3 is 300×300 pixels; The 15-layer convolutional neural network includes: The first group of convolutional layers: 1-2 layers of convolutional layers, each layer uses 64 convolution kernels with a window size of 3 × 3, the window movement step size is 1, the edge is filled with 1 pixel, and the output size is 64 300×300 feature map; The second group of convolutional layers: 3-4 layers of convolutional layers, each layer uses 128 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 128 150×150 feature map; The third group of convolutional layers: 5-7 layers of convolutional layers, each layer uses 256 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 256 75×75 feature map; The fourth group of convolutional layers: 8-10 layers of convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 38×38 feature map; The fifth group of convolutional layers: the 11-13th convolutional layers, each layer uses 512 convolution kernels with a window size of 3 × 3, the window moving step is 1, the edge is filled with 1 pixel, and the output size is 512 19×19 feature map; The sixth group of convolutional layers: the 14th convolutional layer, using 1024 convolution kernels with a window size of 3 × 3, a window shift step size of 1, and edge padding of 1 pixel, outputting 1024 features of size 19 × 19 picture; The seventh group of convolutional layers: the 15th convolutional layer, using 1024 convolution kernels with a window size of 1×1, a window movement step size of 1, no edge padding, and outputting 1024 feature maps with a size of 19×19 ; Among them, the feature maps output by the 2nd, 4th, 7th, 10th, and 13th convolutional layers also need to go through the largest downsampling layer for dimensionality reduction, using a downsampling kernel with a window size of 2 × 2, and a window moving step size of 2. 7. the SAR image iron tower target detection method based on deep learning according to claim 6, is characterized in that, the 8-layer convolutional neural network in described step S3 comprises: The eighth group of convolutional layers: the 16-17th convolutional layer, the 16th convolutional layer uses 256 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer Use 512 convolution kernels with a window size of 3 × 3, a window movement step size of 2, and edge padding of 1 pixel; output 512 feature maps with a size of 10 × 10; The ninth group of convolutional layers: 18-19 layers of convolutional layers, the 18th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 17th layer uses 256 A convolution kernel with a window size of 3×3, a window movement step size of 2, and edge padding of 1 pixel; output 256 feature maps with a size of 5×5; The tenth group of convolutional layers: 20-21 layers of convolutional layers, the 20th layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 21st layer uses 256 A convolution kernel with a window size of 3×3, a window moving step size of 1, and no edge padding; output 256 feature maps with a size of 3×3; The eleventh group of convolutional layers: 22-23 layers of convolutional layers, the 22nd layer uses 128 convolution kernels with a window size of 1 × 1, the window movement step size is 1, and the edges are not filled; the 23rd layer uses 256 A convolution kernel with a window size of 3 × 3, the window movement step size is 1, and the edges are not filled; 256 feature maps with a size of 1 × 1 are output. 8 . The deep learning-based SAR image tower target detection method according to claim 7 , wherein the multi-scale detection network in the step S3 is respectively used for the 10th, 15th, 17th, 19th, 21st, and 23rd layer volumes. 9 . The feature map output by the layered layer performs the following operations: Normalize the feature map; Use a convolution kernel with 4k windows of size 3 × 3, the window movement step size is 1, and the edge is filled with a convolution layer of 1 pixel, and the coordinate information of k default borders is extracted; among them, k=1, 2, 3 , 4, 5, and 6 correspond to the 10th, 15th, 17th, 19th, 21st, and 23rd floors respectively; Use a convolution kernel with (classes+1)×k window size of 3×3, window shift step size 1, edge padding 1 pixel convolution layer, and extract the confidence for each class of k default borders Score, where classes is the number of categories of the target to be detected; The coordinate information and confidence scores extracted from the default borders of the feature maps of each layer are spliced together to generate detection results. 9. The deep learning-based SAR image tower target detection method according to claim 8, wherein for the 10th, 15th, 17th, 19th, 21st, and 23rd layers, the size value of the default frame is in the following formula. The following formulas s ₁ to s ₆ : In the above formula, s _k is the size value of the default frame, s _min =0.2, s _max =0.9; For each layer of feature maps, set different aspect ratios a _r , and combine the layer's size value _sk to get the width of the default border and height They are: Among them, for layers 10, 21, and 23, a _r ∈ {1, 2, 1/2}; for layers 15, 17, and 19, a _r ∈ {1, 2, 3, 1/2, 1/3} , and for the above six convolutional layers, there is also a corresponding size 10. SAR image iron tower target detection method based on deep learning according to claim 9, is characterized in that: In the step S3, when the multi-scale detection network splices the coordinate information and confidence scores extracted from the default frames of the feature maps of each layer to generate the detection results, first discard all results whose confidence scores are lower than the confidence threshold; then use non-polarity The large-value suppression method retains the optimal detection results with higher confidence scores while suppressing the sub-optimal results; finally, the coordinate information and confidence scores after screening are spliced.