CN109002848A - A kind of detection method of small target based on Feature Mapping neural network - Google Patents

Abstract

The invention discloses a kind of detection method of small target based on Feature Mapping neural network, are related to Dim targets detection field；It includes step 1: building, training spindle-type deep neural network；Step 2: by the amplitude figure that the spindle-type deep neural network that the Weak target image input of acquisition has been trained obtains targets improvement and background inhibits；Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.The present invention uses spindle network structure, the powerful expression ability of network is improved, solves the problems, such as that existing Weak target is caused detection accuracy low by noise and interference effect, has reached and has improved the powerful expression ability of network, under high-noise environment, the effect of Dim targets detection precision is improved.

Description

A Weak Target Detection Method Based on Feature Mapping Neural Network

technical field

The invention relates to the field of weak and small target detection, in particular to a weak and small target detection method based on a feature mapping neural network.

Background technique

Passive millimeter-wave PMMW and infrared imaging have excellent characteristics of no radiation and strong penetrating ability, and their application in the military field has attracted increasing attention. Therefore, it is of great significance to study weak and small target detection under millimeter-wave and infrared imaging. Weak and small target detection technology has developed rapidly in recent years. Weak and small targets refer to targets with a diameter of 3-5 pixels. However, high-precision detection of weak and small targets under millimeter-wave and infrared imaging conditions still faces great difficulties: First, the imaging distance of the target Generally far away, the detected target area is small, the signal-to-noise ratio is low, and no texture features can be extracted; second, target imaging is usually interfered by complex backgrounds, a large amount of clutter, noise, and some edge information such as : The existence of cloud edges, sea-sky baselines, building edges, etc., causes the target to be submerged in the background.

For weak and small target detection, a series of detection methods have been proposed in recent years in academia; the background suppression method is the most common method in weak and small target detection. This method estimates the background of the image to be detected and performs target detection on this basis. It is mainly divided into three types of detection methods: the first type is based on filtering methods, which estimate the background through image filtering, and finally enhance the target; the effect of suppressing the background is better when the background is relatively simple, and the effect of suppressing the background is better when the background is more complex. , When the signal-to-noise ratio is low, the false alarm probability increases and the detection accuracy decreases; the second type is a regression-based method, which can be divided into linear regression and nonlinear regression. The classic linear regression method depends on a specific background The clutter model and the parameter estimation of the hypothesis-seeking model; the nonlinear regression method only relies on the data itself to estimate the regression function; the kernel regression algorithm NRRKR is a typical nonlinear regression algorithm; in practical applications, due to the lack of background clutter prior According to empirical knowledge, the nonlinear regression method is more suitable for the detection of weak and small targets under complex background conditions, but it has obvious shortcomings: each local area needs to perform multiple regression iterations, and the overall algorithm efficiency is extremely low; the third type of method is based on local The contrast difference suppresses the background, enhances the target, and completes the detection of the target. This kind of method has better detection effect in the case of simple background, but it is easy to increase the number of false alarm targets and be affected by noise in complex background.

In addition to the background suppression method, there is also a detection method based on machine learning. This type of method uses the idea of pattern classification to solve the problem of target detection. Whether the image block contains the target, such as NLPCA, SPCA, FLD, etc. Later, with the emergence of sparse representation theory, a new method was brought to solve the problem of weak and small target detection. Infrared weak and small target detection algorithm SR based on image sparse representation, this method uses a binary Gaussian model to generate a target dictionary, and then judges the position of the target by the difference between the sparse coefficients of the background sub-block and the target sub-block in the target dictionary. As a typical structured over-complete dictionary, Gaussian dictionary is only suitable for weak targets with Gaussian distribution, while for unstructured targets, its sparse representation coefficients are not enough to distinguish targets from background clutter. Later, Wang et al. proposed a sparse dictionary with multi-scale adaptive morphology to detect weak and small targets. By using atoms of different sizes to describe different components of the image, the more subtle local features of the image were captured and the detection accuracy was improved; then, Chen et al. People proposed a method based on sparsity, which manually constructs a discriminative double dictionary in an offline learning manner to improve the difference in sparse representation; later, a new method was proposed: weak and small moving target detection based on space-time joint sparse reconstruction Algorithm STCSR, this method first constructs an adaptive type over-complete space-time dictionary by learning the content of sequence images, and then uses the multivariate Gaussian model to extract the target space-time dictionary and background space-time dictionary from the over-complete dictionary, and divides the multi-frame images into Sparse reconstruction is performed on the target space-time dictionary and the background space-time dictionary, and the reconstruction difference is used to distinguish the target from the background. Aiming at the shortcomings of this method in dictionary learning, an improved ISR method is proposed, which proposes a salient background and target double dictionary construction method, which has better target and background modeling capabilities. The above methods have improved the detection accuracy to a certain extent, but the above methods have disadvantages: on the one hand, they are easily disturbed by noise, on the other hand, the difference between the sparse features of the target and the background is not obvious, and they are easy to mix together, which increases the difficulty of detection.

Therefore, there is a need for a weak and small target detection method that can overcome the noise factor of weak and small targets and improve the detection accuracy at the same time.

Contents of the invention

The object of the present invention is: the present invention provides a method for detecting weak and small targets based on a feature mapping neural network, which solves the problem of low detection accuracy caused by existing weak and small targets being affected by noise and interference.

The technical scheme that the present invention adopts is as follows:

A weak and small target detection method based on a feature map neural network, comprising the steps of:

Step 1: Construct and train the spindle deep neural network;

Step 2: Input the collected weak target image into the trained spindle deep neural network to obtain the amplitude map of target enhancement and background suppression;

Step 3: The amplitude map uses the constant false alarm rate method to complete weak and small target detection.

Preferably, said step 1 includes the following steps:

Step 1.1: Construct the structure of a spindle-type deep neural network including an input layer, a decoding layer, an encoding layer and a softmax output layer;

Step 1.2: using the cross-validation method to determine the hyperparameters of the spindle-shaped deep neural network to obtain the spindle-shaped deep neural network;

Step 1.3: Construct the training data set;

Step 1.4: Input the training data set into the spindle-type deep neural network and train in an unsupervised manner to obtain initialized network weights to complete the training.

Preferably, the training calculation of the decoding layer and encoding layer is as follows:

h ^k ＝σ(W ^k X+b ^k )

Among them, W ^k represents the weight matrix, b ^k represents the bias vector, σ represents the activation function, X={x ₁ , x ₂ ,...,x _m } represents the input of the current layer, and h ^k represents the output of the current layer.

Preferably, said step 2 includes the following steps:

Step 2.1: After inputting the collected weak target image into the trained spindle deep neural network, set the weak target sample label to 1 and the background sample label to 0;

Step 2.2: Use the sliding window method to discriminate the weak and small target image to obtain the probability value of the weak and small target in the image;

Step 2.3: Use the output layer logistic regression results, that is, multiple probability values as the response amplitude of the window coordinate point to obtain the target enhancement and background suppression amplitude maps corresponding to multiple weak targets.

Preferably, said step 3 includes the following steps:

Step 3.1: Each value in the magnitude map uses a sliding window to extract sub-blocks For detection, input the spindle-type deep neural network to obtain the amplitude;

Step 3.2: Use the constant false alarm rate to perform statistics on each detected amplitude to obtain the false alarm probability;

Among them, T represents the threshold of likelihood ratio detection, Represents the mean value of the sliding window sub-block, p represents the number of points in the window sub-block, P _fa represents the false alarm probability set by the constant false alarm rate detection, τ _CFAR represents the detection threshold, F _1,p-1 (τ _CFAR ) Represents the cumulative distribution function of the central F random variable;

Step 3.3: Calculate the total number of candidate targets according to the false alarm probability, sort the false alarm probability from high to low, and detect and locate the target from multiple amplitudes.

Preferably, the construction training data set in said step 1.3 includes the following steps:

Step 1.3.1: Randomly generate coordinate points in the image that does not contain the weak target as the simulation target, and extract the N*N window area as the background sample;

Step 1.3.2: Use a two-dimensional Gaussian intensity model plus a simulation target as the target sample in the background sample. The two-dimensional Gaussian model is as follows:

Among them, (x ₀ , y ₀ ) represents the center position of the target image, s(i,j) represents the pixel value of the target image at position (i, j), s _E represents the intensity of the generated target, and its value is (0, 1], σ _x and σ _y represent the horizontal and vertical spread parameters respectively, and their values are between [0, 2];

Step 1.3.3: Adjust different parameters of the target sample to generate weak targets with different signal-to-noise ratios to complete the construction of the training data set.

In summary, owing to adopting above-mentioned technical scheme, the beneficial effect of the present invention is:

1. The present invention adopts a spindle-shaped network structure, first maps low-dimensional weak target block features to high-dimensional space, then extracts high-resolution features through an encoding neural network, completes background and target discrimination, and obtains background according to the intensity of network discrimination output Suppress the enhanced image of the target, and finally use the detection method based on the constant false alarm rate to complete the detection of weak and small targets, which solves the problem of low detection accuracy caused by the influence of noise and interference on existing weak and small targets, and achieves the improvement of the powerful representation ability of the network. In a high-noise environment, the effect of improving the detection accuracy of weak and small targets;

2. The present invention detects millimeter-wave and infrared images in different scenarios, and realizes the decoding operation at the front end of the network, so that the network has more powerful representation capabilities. First, the pixel features of weak and small target areas are mapped to a high-dimensional feature space, and then High-dimensional features are encoded and mapped to low-dimensional feature spaces that are easy to distinguish, achieving lower false alarm rates, higher detection accuracy, and stronger robustness;

3. The unsupervised learning method pre-training adopted by the network of the present invention builds a deeper hierarchical structure, and the initialization network weight obtained through unsupervised learning can improve network stability, so as to avoid the local minimum problem in the process of directly training the deep neural network, and at the same time Unsupervised learning acquires the internal features of a series of related data sets and removes redundant components of the input data, which is conducive to obtaining highly discriminable features and further improving the discriminative accuracy of the network;

4. adopt slide block to detect in the constant false alarm rate method of the present invention, adopt slide block to carry out constant false rate detection when assuming that two target distances have certain distance according to actual situation, can accelerate detection speed on the one hand, be beneficial to on the other hand Improve detection accuracy.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and thus It should be regarded as a limitation on the scope, and those skilled in the art can also obtain other related drawings based on these drawings without creative work.

Fig. 1 is the schematic diagram of spindle type deep neural network structure of the present invention;

Fig. 2 is the specific implementation schematic diagram of the spindle type deep neural network of the present invention;

Fig. 3 is the two-dimensional diagram of training data of the present invention;

Fig. 4 is the training data three-dimensional map of the present invention;

Fig. 5 is the training data two-dimensional map of the present invention;

Fig. 6 is the training data schematic diagram of the present invention;

Fig. 7 is the data schematic diagram of simulation test data set of the present invention;

Fig. 8 is a schematic diagram of detection results after the background suppression of the simulation diagram of the present invention;

Fig. 9 is the relationship figure between detection probability P _d and false alarm rate P _fa when SNR<10 of the present invention;

Fig. 10 is the relationship diagram between detection probability P _d and false alarm rate P _fa when 10<SNR<20 of the present invention;

Fig. 11 is the relationship diagram between the detection probability _{Pd and the false alarm rate Pfa when} 20<SNR<30 of the present invention;

Fig. 12 is the relationship diagram between the detection probability P _d and the false alarm rate P _fa when 30<SNR<40 of the present invention;

Fig. 13 is a detection diagram of the variation effect of _Pd with SNR under P _fa = 1e-4 of the present invention;

Fig. 14 is P _fa =10e-4 of the present invention, P _d changes the effect detection diagram with SNR;

Fig. 15 is P _fa =20e-4 of the present invention, P _d changes effect detection figure with SNR;

Fig. 16 is P _fa =30e-4 of the present invention, P _d changes effect detection figure with SNR;

FIG. 17 is a schematic enlarged view of the effect of the DL algorithm of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, and are not intended to limit the present invention, that is, the described embodiments are only some of the embodiments of the present invention, but not all of the embodiments. The components of the embodiments of the invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

Accordingly, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts belong to the protection scope of the present invention.

It should be noted that relative terms such as the terms "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. There is no such actual relationship or order between them. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

Technical problem: Solve the problem of low detection accuracy caused by existing weak and small targets affected by noise and interference

Technical means:

A weak and small target detection method based on a feature map neural network, comprising the steps of:

Step 1: Construct and train the spindle deep neural network;

Step 2: Input the collected weak target image into the trained spindle deep neural network to obtain the amplitude map of target enhancement and background suppression;

Step 3: The amplitude map uses the constant false alarm rate method to complete weak and small target detection.

Step 1 includes the following steps:

Step 1.1: Construct the structure of a spindle-type deep neural network including an input layer, a decoding layer, an encoding layer and a softmax output layer;

Step 1.2: using the cross-validation method to determine the hyperparameters of the spindle-shaped deep neural network to obtain the spindle-shaped deep neural network;

Step 1.3: Construct the training data set;

Step 1.4: Input the training data set into the spindle-type deep neural network and train in an unsupervised manner to obtain initialized network weights to complete the training.

The decoding layer and encoding layer training calculation are as follows:

h ^k ＝σ(W ^k X+b ^k )

Among them, W ^k represents the weight matrix, b ^k represents the bias vector, σ represents the activation function, X={x ₁ , x ₂ ,...,x _m } represents the input of the current layer, and h ^k represents the output of the current layer.

Step 2 includes the following steps:

Step 2.1: After inputting the collected weak target image into the trained spindle deep neural network, set the weak target sample label to 1 and the background sample label to 0;

Step 2.2: Use the sliding window method to discriminate the weak and small target image to obtain the probability value of the weak and small target in the image;

Step 2.3: Use the output layer logistic regression results, that is, multiple probability values as the response amplitude of the window coordinate point to obtain the target enhancement and background suppression amplitude maps corresponding to multiple weak targets.

Step 3 includes the following steps:

Step 3.1: Each value in the magnitude map uses a sliding window to extract sub-blocks For detection, input the spindle-type deep neural network to obtain the amplitude;

Step 3.2: Use the constant false alarm rate to perform statistics on each detected amplitude to obtain the false alarm probability;

Among them, T represents the threshold of likelihood ratio detection, Represents the mean value of the sliding window sub-block, p represents the number of points in the window sub-block, P _fa represents the false alarm probability set by the constant false alarm rate detection, τ _CFAR represents the detection threshold, F _1,p-1 (τ _CFAR ) Represents the cumulative distribution function of the central F random variable;

Step 3.3: Calculate the total number of candidate targets according to the false alarm probability, sort the false alarm probability from high to low, and detect and locate the target from multiple amplitudes.

The construction of the training data set in step 1.3 includes the following steps:

Step 1.3.1: Randomly generate coordinate points in the image that does not contain the weak target as the simulation target, and extract the N*N window area as the background sample;

Step 1.3.2: Use a two-dimensional Gaussian intensity model plus a simulation target as the target sample in the background sample. The two-dimensional Gaussian model is as follows:

Among them, (x ₀ , y ₀ ) represents the center position of the target image, s(i,j) represents the pixel value of the target image at position (i, j), s _E represents the intensity of the generated target, and its value is (0, 1], σ _x and σ _y represent the horizontal and vertical spread parameters respectively, and their values are between [0, 2];

Step 1.3.3: Adjust different parameters of the target sample to generate weak targets with different signal-to-noise ratios to complete the construction of the training data set.

Technical effect:

The present invention adopts a spindle network structure, first maps low-dimensional weak target block features to high-dimensional space, then extracts high-resolution features through a coding neural network, completes background and target discrimination, and obtains background suppression targets according to the intensity of network discrimination output In the enhanced image, the detection method based on the constant false alarm rate is finally used to complete the detection of weak and small targets, which solves the problem of low detection accuracy caused by the influence of noise and interference on existing weak and small targets, and achieves the improvement of the powerful representation ability of the network. In the environment, the effect of improving the detection accuracy of weak and small targets.

The characteristics and performance of the present invention will be described in further detail below in conjunction with the examples.

Example 1

The training data of the network is shown in Figure 3. The small circles represent one type of project, and the x point represents another type of sample. The training samples are linear and inseparable, which is difficult to distinguish; after training the deep network through the training samples, the output The second and third layer features are shown in Figure 4. After the linearly inseparable two-dimensional features are mapped to three-dimensional through the neural network, they have linear discriminability in the three-dimensional feature space. After the three-dimensional features are re-encoded to two-dimensional, as shown in Figure 5 As shown, the original non-linear and non-separable low-dimensional features are mapped to the linear discriminable feature space, and the blue x-point samples are distributed in a compact area. As shown in Figure 2, the entire network contains a total of 6 layers, which can be divided into 2 parts. The input layer to the second layer is the decoding layer, which completes the feature upgrade and the low-dimensional to high-dimensional mapping; the 2nd to 5th layers are A typical sparse autoencoder realizes the encoding and extraction of abstract advanced features, and the last layer is the softmax output layer; the main difference between the network model proposed in this paper and the typical deep neural network model is the decoding layer from the first layer to the second layer, traditional The deep neural network mainly processes high-dimensional data and extracts abstract dimensionality reduction features; based on the small target detection window and low dimensionality, the decoding operation is first implemented at the front end of the network, so that the network has more powerful representation capabilities; In a high-noise environment, the detection accuracy of weak and small objects is higher; construct and train a spindle-shaped deep neural network: Step 1.1: Construct the structure of a spindle-shaped deep neural network including an input layer, a decoding layer, an encoding layer and a softmax output layer; Step 1.2 : Using cross-validation method to determine the hyperparameters of the spindle-type deep neural network; Step 1.3: Construct the training data set; Step 1.4: Input the training data set into the spindle-type deep neural network and use unsupervised training to obtain the initialized network weights to complete the training. The size of each layer of the trained network model is [81, 512, 256, 121, 81, 1], where the input layer {I1, I2, ..., IN} is a linear arrangement of pixels in the weak and small target detection window in the image; feature conversion , we use a sparse autoencoder, and the calculation formula is as follows:

Among them, x ⁽ⁱ⁾ represents the activation degree of the hidden neuron j of the self-encoder neural network when the given input is x, represents the average activity of hidden neuron j; ρ represents the sparsity parameter, and β represents the hyperparameter.

ρ is set to 0.5, β is set to 3, the gradient descent method is used to learn network parameters, and the learning rate is set to 0.01; during training, the label of weak and small target samples is set to 1, the label of background samples is set to 0, and the method of sliding window is used to All areas of the image are input into the network for discrimination, and the result of the logistic regression of the output layer is used as the response amplitude of the window coordinate point. The larger the value, the higher the probability that the detection window contains a weak target. The target enhancement and The amplitude map of background suppression is I _o ; the amplitude map uses the constant false alarm rate method to complete the detection of weak and small targets; the training data set is obtained by simulation method, and artificially add weak and small targets to 220 images that do not contain weak and small targets. Construct a training set; randomly generate coordinate points in each image, extract a 9*9 area as a background sample, and use a two-dimensional Gaussian intensity model plus a simulation target as a target sample in the background sample. The two-dimensional Gaussian model is as follows:

Among them, (x ₀ , y ₀ ) represents the center position of the target image, s(i,j) represents the pixel value of the target image at position (i, j), s _E represents the intensity of the generated target, and its value is (0, 1], σ _x and σ _y represent the horizontal and vertical scatter parameters respectively, and their values are between [0, 2].

Weak targets with different SNRs are generated by adjusting different parameters. The SNRs of the weak targets generated in this paper are between [0-120]. Samples and 26400 negative samples, part of the generated training sample images are shown in Figure 6, and part of the simulation test data set data is shown in Figure 7.

The test set includes a simulation test set and a real data test set. The image data includes millimeter-wave images and infrared images. The infrared images come from multiple data sets. The simulation test data set uses the same weak target simulation method as the training data. Add weak and small targets with different signal-to-noise ratios. In this application, a total of 1920 weak and small targets are added to 32 images, and their signal-to-noise ratios are similarly evenly distributed between [0-120]Db; input the test set data into the trained The spindle-shaped neural network completes the test. The low-dimensional weak target block features are first mapped to the high-dimensional space through the spindle-shaped neural network, and then the high-resolution features are extracted through the encoding neural network to complete the background and target discrimination. According to the strength of the output of the network discrimination The enhanced image of the background suppressed target is obtained, and finally the detection method based on constant false alarm rate is used to complete the detection of weak and small targets, which effectively improves the detection accuracy of weak and small targets.

Example 2

Step 1 includes the following steps:

Step 1.1: Construct the structure of a spindle-type deep neural network including an input layer, a decoding layer, an encoding layer and a softmax output layer;

Step 1.2: using the cross-validation method to determine the hyperparameters of the spindle-shaped deep neural network to obtain the spindle-shaped deep neural network;

Step 1.3: Construct the training data set;

Step 1.4: Input the training data set into the spindle-type deep neural network and train in an unsupervised manner to obtain initialized network weights to complete the training.

The decoding layer and encoding layer training calculation are as follows:

h ^k ＝σ(W ^k X+b ^k )

Among them, W ^k represents the weight matrix, b ^k represents the bias vector, σ represents the activation function, X={x ₁ , x ₂ ,...,x _m } represents the input of the current layer, and h ^k represents the output of the current layer.

Step 2 includes the following steps:

Step 2.1: After inputting the collected weak target image into the trained spindle deep neural network, set the weak target sample label to 1, and the background sample label to 0. The details of the image are as follows:

Among them, s represents the image collected by the sensor, s _t represents the target signal, s _b represents the background signal, and n represents the noise.

Step 2.2: Use the sliding window method to discriminate the weak and small target image to obtain the probability value of the weak and small target in the image;

Step 2.3: Use the output layer logistic regression results, that is, multiple probability values as the response amplitude of the window coordinate point to obtain the target enhancement and background suppression amplitude maps corresponding to multiple weak targets.

Step 3 includes the following steps:

Step 3.1: Each value in the magnitude map uses a sliding window to extract sub-blocks Perform detection and input the network to obtain the amplitude;

Step 3.2: Use the constant false alarm rate to perform statistics on each detected amplitude to obtain the false alarm probability;

Among them, T represents the threshold of likelihood ratio detection, Represents the mean value of the sliding window sub-block, p represents the number of points in the window sub-block, P _fa represents the false alarm probability set by the constant false alarm rate detection, τ _CFAR represents the detection threshold, F _1,p-1 (τ _CFAR ) Represents the cumulative distribution function of the central F random variable;

Step 3.3: Calculate the total number of candidate targets according to the false alarm probability, sort the false alarm probability from high to low, and detect and locate the target from multiple amplitudes.

The construction of the training data set in step 1.3 includes the following steps:

Step 1.3.1: Randomly generate coordinate points in the image that does not contain the weak target as the simulation target, and extract the N*N window area as the background sample;

Step 1.3.2: Use a two-dimensional Gaussian intensity model plus a simulation target as the target sample in the background sample. The two-dimensional Gaussian model is as follows:

Among them, (x ₀ , y ₀ ) represents the center position of the target image, s(i,j) represents the pixel value of the target image at position (i, j), s _E represents the intensity of the generated target, and its value is (0, 1], σ _x and σ _y represent the horizontal and vertical spread parameters respectively, and their values are between [0, 2];

Step 1.3.3: Adjust different parameters of the target sample to generate weak targets with different signal-to-noise ratios to complete the construction of the training data set.

By comparing and analyzing the detection effects of several other mainstream algorithms and the detection effect of this application, the accuracy of this application is reflected. Two types of curves are used as evaluation indicators. The first type of curve is the ROC curve. In target detection, it reflects the relationship between the detection probability P _d and the false alarm rate P _fa . The larger the area under the ROC curve, the better the detection performance. The better, the calculation formula of P _d and P _fa is as follows:

Among them, N _t represents the number of correctly detected targets, N _a represents the total number of targets, N _f represents the false number of detected targets, and N represents the number of all pixels in the image.

The second type of curve is the change relationship between the detection probability P _d and the signal-to-noise ratio SNR. As the SNR value increases, P _d will gradually increase, and finally approach 1. The SNR calculation formula used is:

Among them, g _t represents the average value of the pixels in the target local area, and g _b and σ _b represent the average value and standard deviation of the pixels in the background local area, respectively.

The detection effects of several mainstream algorithms include: ACSDM, CSCD, SR, ISTCR, and ISTCSR-CSCD. This application represents the DL algorithm; CSCD and DL algorithms; among them, the green solid line box represents the real position of the target, and the red box shows the output of the detector. If the detector output position coincides with the real position of the target, the red box will cover the green box, and the target point is only real red Frame, there are only green frame targets, which means that the detector has missed detection under the specified false alarm rate. It can be seen from the figure that the DL algorithm has the largest number of detected real targets and the least number of missed targets compared with other algorithms. ;

The quantitative analysis results of the six algorithms are shown in Figure 9-12. Figure 9-12 describes the different SNRs (SNR<10, 10<SNR<20, 20<SNR<30, 30<SNR<40 ) between the detection probability P _d and the false alarm rate P _fa , the solid line marked star is the result of the DL algorithm proposed in this paper, it can be seen that the algorithm of this application is generally better than The other 5 algorithms; when SNR<10, when P _fa =1×10 ^-4 , the ISTCR algorithm is better than the method in this paper, and has the best result, and our method Dl has better detection accuracy in other cases; in When 10<SNR<20, the detection rate of the algorithm in this paper is much higher than other methods in each false alarm rate, compared with the second method CSCD, DL is about 20% higher; when 20<SNR<30, higher than The best method of the same kind is about 12%; when 30<SNR<40, the method in this paper is also significantly better than similar methods, about 8% higher;

Figure 13-16 shows the change of P _d with SNR under the same P _fa . The solid line marked with star represents the DL algorithm proposed by this application. It can be seen that in the case of four different constant false alarm rates, DL algorithms have achieved the best results, especially when P _fa >10×10 ^-4 , the method in this paper is 20% higher than the best method of the same kind on average; when P _fa =1×10 ^-4 , DL, ISTCR and The performance of ISTCR-CSCD three algorithm detection methods is relatively close, but obviously better than the other three algorithms. Figure 17 is an enlarged schematic diagram of the detection effect of the DL algorithm. Because this detection uses color to distinguish the real position frame and the detection output frame, the effect may not be obvious after black and white according to the patent law. If necessary, a color map can be provided.

By constructing the spindle neural network structure, the characteristics of weak and small targets in complex backgrounds are learned, and the image blocks are discriminated through the deep neural network to output the target probability. The target area has a higher probability and the background area has a lower probability, and the target intensity is constructed based on this probability. map, and then use the constant false alarm rate (CFAR) method for target detection and positioning. Compared with the mainstream algorithm, the detection accuracy of the DL algorithm is increased by about 15% on average. Especially on the real test image, the DL performance is far superior to similar methods. The learning method has better detection ability in complex backgrounds, which shows that the DL method has a generalization ability far exceeding similar methods when the detection accuracy is high.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention should be included in the protection of the present invention. within range.

Claims (6)

1. a kind of detection method of small target based on Feature Mapping neural network, characterized by the following steps:

Step 1: building, training spindle-type deep neural network；

Step 2: the spindle-type deep neural network that the Weak target image input of acquisition has been trained is obtained into targets improvement and back The amplitude figure that scape inhibits；

Step 3: amplitude figure completes Dim targets detection using constant false alarm rate method.

2. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist In: the step 1 includes the following steps:

Step 1.1: building includes the spindle-type deep neural network of input layer, decoding layer, coding layer and softmax output layer Structure；

Step 1.2: determining that the hyper parameter of spindle-type deep neural network obtains spindle-type depth nerve using cross validation method Network；

Step 1.3: building training dataset；

Step 1.4: training dataset input spindle-type deep neural network is trained acquisition initially using unsupervised mode Change network weight and completes training.

3. a kind of detection method of small target based on Feature Mapping neural network according to claim 2, feature exist In: the decoding layer, coding layer training calculate as follows:

h^k=σ (W^kX+b^k)

Wherein, W^kIndicate weight matrix, b^kIndicate that bias vector, σ indicate activation primitive, X={ x₁,x₂,...,x_mIndicate current The input of layer, h^kIndicate the output of current layer.

4. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist In: the step 2 includes the following steps:

Step 2.1: after the spindle-type deep neural network that the Weak target image input of acquisition has been trained, by Weak target sample This label is set as 1, and background sample label is set as 0；

Step 2.2: Weak target image being carried out to differentiate the probability for obtaining Weak target in image using the method for sliding window Value；

Step 2.3: being obtained using result, that is, multiple probability values that output layer logistic is returned as window coordinates point response amplitude The amplitude figure for taking the corresponding targets improvement of multiple Weak targets and background to inhibit.

5. a kind of detection method of small target based on Feature Mapping neural network according to claim 1, feature exist In: the step 3 includes the following steps:

Step 3.1: each value in amplitude figure extracts sub-block using sliding windowIt is detected, inputs spindle-type depth Neural network obtains amplitude；

Step 3.2: statistics being carried out to each amplitude after detection using constant false alarm rate and obtains false-alarm probability；

In, T indicates the threshold value of Likelihood ration test,Indicate the mean value of sliding window sub-block, p indicates put in window sub-block Number, P_faIndicate the false-alarm probability of constant false alarm rate detection setting, τ_CFARIndicate detection threshold value, F_1,p-1(τ_CFAR) indicate that center F is random The cumulative distribution function of variable；

Step 3.3: candidate target sum being calculated according to false-alarm probability, false-alarm probability is sorted from high to low, from multiple amplitudes Detection positioning target.

6. a kind of detection method of small target based on Feature Mapping neural network according to claim 2, feature exist In: the building training dataset in the step 1.3 includes the following steps:

Step 1.3.1: coordinate points are randomly generated in not including Weak target image as simulation objectives, extract N*N window region Domain is as background sample；

Step 1.3.2: in background sample using dimensional Gaussian strength model plus simulation objectives as target sample, two It is as follows to tie up Gauss model:

Wherein, (x₀,y₀) indicate target image center, s (i, j) indicate target image position (i, j) pixel value, s_EIndicate generate target intensity, value be (0,1] between random number, σ_xAnd σ_yHorizontal and vertical distribution parameter is respectively indicated, Its value is between [0,2]；

Step 1.3.3: the different parameters for adjusting target sample generate the Weak targets of different signal-to-noise ratio and complete training dataset Building.