CN102867195B - Method for detecting and identifying a plurality of types of objects in remote sensing image - Google Patents

Abstract

The invention relates to a method for detecting and recognizing multiple types of targets in remote sensing images based on sparse representation dictionary learning. The technical features are as follows: first, a dictionary is trained on the preprocessed training data using a training method based on a sparse representation dictionary; then, the sub-image block in the test image is sparsely coded using the dictionary obtained through training, and its sparse representation coefficient is obtained to obtain The reconstruction error of the sub-image block is obtained, and the candidate target area is determined by thresholding the reconstruction error; finally, the accurate detection and recognition of multiple types of targets in the remote sensing image is realized through post-processing. By using the method of the invention, multiple types of targets can be detected and recognized from remote sensing images under complex backgrounds. The invention has higher detection and recognition precision and lower false alarm rate.

Description

A Method for Detection and Recognition of Multiple Types of Targets in Remote Sensing Images

technical field

The invention relates to a method for detecting and identifying multiple types of targets in remote sensing images, which can be applied to detecting and identifying multiple types of targets under complex background remote sensing images.

Background technique

As an application of remote sensing image processing technology, target detection and recognition under complex background remote sensing images is a key technology in the fields of military reconnaissance and precision strikes, and has always been a research hotspot and difficulty in this field, with important military and civilian applications. value has received more and more attention.

At present, there are mainly two methods for object detection in remote sensing images. One is to solve the target detection problem by detecting certain shape and geometric features of the target in the remote sensing image, but due to the complex background of the remote sensing image, there are a large number of similar shape and geometric features to the target, and only rely on these features to detect There will be a large number of missed detections and false detections in the target. The other is based on the idea of classification, the most common of which is the Bag-of-Words (BoW) classification method, which first extracts SIFT features from the image and clusters them, using the cluster center as a set of criteria in the image space base (standard image area), then you can use this set of standard bases to represent the image as a vector, and finally use the SVM classifier to classify and threshold the obtained vector to obtain the detection result; but the BoW method, although the extracted SIFT The feature has scale and rotation invariance, but only uses the statistical characteristics of the feature area, while ignoring the spatial information of the feature area, so the method using BoW has a low detection rate and a high false alarm rate; another classification method Linear Spatial Pyramid Although Matching Using Sparse Coding (ScSPM) takes into account the spatial information of the feature area, the dimension of the obtained vector for classification is too high and the amount of calculation is too large. In addition, most current classification-based target detection methods are limited to single target detection, and cannot detect and recognize multiple targets at the same time.

Contents of the invention

technical problem to be solved

In order to avoid the deficiencies of the prior art, the present invention proposes a method for detecting and recognizing multi-category objects in remote sensing images based on sparse representation dictionary learning. This method can automatically detect and recognize different types of targets from remote sensing images with complex backgrounds, and has high detection accuracy and low false alarm rate.

Technical solutions

A remote sensing image multi-class target detection and recognition method is characterized in that the steps are as follows:

Step 1: Train the dictionary using the method based on sparse representation dictionary learning, the specific steps are as follows:

Step a1 Pre-processing of training images: first unify the objects of the same category in the original image into one main direction, and then unify the direction of the image along 0° to 360°, according to the step size rotate as images in different directions; the original images of different types of targets are processed according to the above method, and the obtained Class training images, where p is the number of different categories of targets to be detected, is the rotation angle, c is the total number of categories of images in different directions for different targets in the obtained training image; where: is rounded down;

Step b1 data preprocessing: use the weighted average method to The weighted average of the RGB three components of the class training image is to obtain a grayscale image, and then the grayscale image is down-sampled to obtain an n×n size image; the energy normalization process is performed on the n×n size image to obtain a normalized Convert the normalized image to n ² ×1-dimensional column vector, and use the column vector as a column in the training data to obtain the preprocessed training data set U=[U ₁ ,U ₂ ,…,U _c ], where U _i is the sub-data set corresponding to the i-th class in the training data set U, i=1,2,...,c;

Step c1 training dictionary: use the FDDL software package released by Fisher Discrimination Dictionary Learning for SparseRepresentation to train the known training data set U=[U ₁ , U ₂ ,...,U _c ] to obtain the dictionary D=[D ₁ , D ₂ ,...,D _c ], where D _i is the sub-dictionary corresponding to the i-th class;

Step 2 Sparse coding: according to the dictionary D=[D ₁ , D ₂ , ..., D _c ] obtained from training, perform sparse coding on each sub-image block in the test image, and find the corresponding sparseness of each sub-image block The specific processing steps are as follows:

Step a2 Test image preprocessing: first use the weighted average method described in step b1 to convert the test image into a test grayscale image, and then use a sliding window of size S×S to slide along the test grayscale image with an interval step b Obtain the sub-image block; downsample the sub-image block to an image of size n×n, then perform energy normalization processing, and then convert the image after energy normalization processing into an n ² ×1-dimensional column vector β, using the column vector β to represent the pixel gray value information of the sub-image block obtained through the sliding window;

Step b2 sparse coding: pass the optimization model for each sub-image block

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D\&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}$

Get the sparse coding coefficient corresponding to each sub-image block in is the coefficient vector corresponding to the sub-dictionary D _i , ε>0 is the allowable error, ||·|| ₁ is the l ₁ norm, and ||·|| ₂ is the l ₂ norm;

Step c2 calculates the reconstruction error: according to the sparse coding coefficient Calculate the reconstruction error e _i of each sub-image block and each category, take e=min{e _i } as the reconstruction error of this sub-image block, and record its corresponding category Then, according to the size relationship between the reconstruction error e and the preset threshold τ, it is determined whether the sub-image block contains the target: if e<τ, it means that the target is contained, otherwise, it means that the sub-image block is the background;

Step 3 target detection and recognition:

Step a3: Combine the reconstruction error e corresponding to each sub-image block determined to contain the target in step c2 to form a reconstruction error matrix E=( _est ) _P that is consistent with the size of the test grayscale image and represents the candidate target area _{× Q} ; where, e _st is the value of the reconstruction error matrix at the coordinate point (s, t), $e_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ e& e<\mathrm{\&tau;}\end{array}\right.,$ P×Q is the size of the test image, s=1,2,...P, t=1,2,...Q;

In step c2, the category C corresponding to each sub-image block that contains the target is determined to form a category matrix L=(C _st ) _P×Q that is consistent with the size of the test grayscale image and represents the candidate target category; where C _st is The value of the category matrix at the coordinate point (s, t), $C_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ C& e<\mathrm{\&tau;}\end{array}\right.;$

Step b3: Change the size of the sliding window S×S G times, repeat steps 2 to 3G times to obtain G reconstruction error matrices and G category matrices, and the value range of G is 5~10; the obtained G Reconstruction error matrices form a multi-scale reconstruction error matrix MAP=( _estg ) _P×Q×G ; among them, e _stg is an element in the matrix MAP, and its value is the reconstruction obtained by changing the size of the sliding window for the gth time The _est corresponding to the error matrix, P×Q×G is the size of the multi-scale reconstruction error matrix, g=1,2,...G;

The obtained G category matrices constitute a multi-scale category matrix CLASS=(C _stg ) _P×Q×G ; where, C _stg is an element in the matrix CLASS, and its value is the category matrix obtained by changing the size of the sliding window for the gth time The corresponding C _st ; According to the multi-scale reconstruction error matrix MAP, a minimum reconstruction error matrix (map(s,t)) _P×Q is obtained, where map(s,t) is the corresponding minimum reconstruction error matrix at the coordinate point the value at (s,t),

Then find the minimum class matrix (class(s, t)) _P×Q corresponding to the minimum reconstruction error matrix, where class(s, t) is the value of the minimum class matrix at the coordinate point (s, t),

Calculate the scale matrix scale=(scale(s,t)) _P×Q according to the multi-scale reconstruction error matrix MAP, scale(s,t) is the value of the corresponding scale matrix at the coordinate point (r,t), $\mathrm{scale}(\mathrm{the\; s},t)=\left\{\begin{array}{cc}0& e_{\mathrm{st}}=0\\ \underset{g}{\arg \min }\left\{e_{\mathrm{stg}}\right\}& e_{\mathrm{st}}\&NotEqual;0\end{array}\right.;$

Step c3: Calculate the local neighborhood minima of the minimum reconstruction error matrix (map(s,t)) _P×Q as the detected target response value, and the local neighborhood minima are in the minimum reconstruction error matrix (map (s,t)) _P×Q corresponds to the center position of the target, according to the center position in (class(s,t)) _P×Q and (scale(s,t)) _P×Q corresponding Find the category and scale size corresponding to the target.

The calculation formula of the weighted average method is f(x, y)=0.3R(x, y)+0.59G(x, y)+0.11B(x, y), in the formula, f(x, y) is weighted The gray value of the grayscale image obtained by the averaging method at the pixel point (x, y), R(x, y), G(x, y) and B(x, y) are the input training image at the pixel point ( x, y) RGB three-component value.

The energy normalization calculation formula is In the formula, f _norm (x, y) is the gray value of f(x, y) after energy normalization, and u and v are the row and column sizes of the gray image, respectively.

The calculation formula of the _l1 norm is

${<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; |</span>$

In the formula, z is a vector with size M×1, ξ _k is the element of vector z, k=1,2,...,M.

The calculation formula of the ₁₂ norm is

${<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

In the formula, z is a vector with size M×1, ξ _k is the element of vector z, k=1,2,...,M.

The calculation formula of the reconstruction error e _i is

${\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&gamma;</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

In the formula, γ is a preset weight, and the value range of γ is 0~1, m _i is the mean value vector obtained by averaging the elements of each row in Y _i ; Y _i is that U _i is sparse through dictionary D The optimal encoding coefficient obtained by encoding.

The rotation angle The value range is from 0° to 90°.

The range of the FDDL software package parameter _λ1 is 0.001-0.01, and the range of _λ2 is 0.01-0.1.

The value range of S is an integer between 40 and 90, and the value range of b is an integer between 1 and 15.

The value range of the threshold τ is 0~1.

Beneficial effect

The present invention proposes a method for detecting and recognizing multi-category targets in remote sensing images based on sparse representation dictionary learning. First, the preprocessed training data is used to train a redundant dictionary, and then the sub-image blocks in the test image are obtained by training. The dictionary is sparsely encoded to obtain its sparse representation coefficient, and then obtain the reconstruction error of the sub-image block through the sparse representation coefficient, and perform thresholding processing on it to determine the candidate target area; finally, after some post-processing, the remote sensing image Accurate detection and recognition of multiple types of targets.

The invention can automatically detect and recognize multiple types of targets from remote sensing images under complex backgrounds. Practice has proved that this method has high detection and recognition accuracy and low false alarm rate.

Description of drawings

Fig. 1: basic flowchart of the inventive method

Figure 2: Training data in the inventive method

Figure 3: Partial test results of the inventive method

(a) Aircraft target detection results (red boxes represent aircraft targets, yellow boxes are false alarms)

(b) Ship target detection results (the white box represents the ship target)

(c) Oil depot target detection results (the blue box represents the oil depot target)

(d) Aircraft and ship target detection results

(e) Detection results of aircraft and oil depot targets

(f) Target detection results of ships and oil depots

Detailed ways

Now in conjunction with embodiment, accompanying drawing, the present invention will be further described:

The hardware environment used for implementation is: Intel Pentium 2.93GHz CPU computer, 2.0GB memory, and the running software environment is: Matlab R2011a and Windows XP. 100 remote sensing images obtained from Google Earth were selected for multi-type target detection experiments, mainly including three types of targets: aircraft, ships, and oil depots. Among them, there are 200 aircraft targets, 120 ship targets, and oil depot targets. 420 in total.

The present invention is specifically implemented as follows:

1. Training redundant dictionaries: Training dictionaries based on sparse representation dictionary learning methods, the specific process is as follows:

(1.1) Pre-processing of the training image: the specific processing process is: first unify the objects of the same category in the original image into one main direction, and then rotate the image after the unified direction along 0° to 360°, every 10°, 36 types of training data are obtained, and the original images of different types of targets are processed according to the above method, and finally 55 types of training images are obtained, that is, c=55, of which, there are 36 types of aircraft, 18 types of ships, and 1 type of oil depot;

(1.2) Data preprocessing: weighted average of the RGB three components of 55 types of training images by weighted average method to obtain a grayscale image, and then downsampled the grayscale image to obtain an image of 15×15 size, and for 15×15 Perform energy normalization processing on images with a size of 15 to obtain a normalized image, and then convert the normalized image into a 255×1-dimensional column vector, and use this column vector as a column in the training data to obtain the preprocessed training data Set U=[U ₁ , U ₂ ,..., U _c ], where U _i is the sub-data set corresponding to the i-th class in the training data set U, i=1, 2,..., c;

(1.3) Train the known training data set U=[U ₁ ,U ₂ ,...,U _c ] through the FDDL software package released by Lei Zhang, and get the dictionary D=[D ₁ ,D ₂ ,...,D _c ] , where D _i is a sub-dictionary corresponding to the i-th category; software package parameters λ ₁ =0.005, λ ₂ =0.05;

The FDDL software package of Lei Zhang can be found in the paper: Meng Yang, Lei Zhang, Xiangchu Feng, David Zhang. Fisher Discrimination Dictionary Learning for Sparse Representation[C].ICCV, 2011

2. Sparse coding: According to the dictionary D=[D ₁ ,D ₂ ,...,D _c ] obtained from training, perform sparse coding on each sub-image block in the test image, and find the corresponding sparse coefficient of each sub-image block , the specific processing steps are as follows:

(2.1) Test image preprocessing: first use the weighted average method described in (1.1) to convert the test image into a test grayscale image, and then use a sliding window of size S×S to step along the test grayscale image at intervals The sub-image block is obtained by sliding 5 pixels, and the initial value of S is 90; for each sub-image block obtained by using the sliding window, it is down-sampled to an image with a size of 15×15, and then energy normalization is performed. Then convert the energy-normalized image into a 225×1-dimensional column vector β, and use the column vector β to represent the pixel gray value information of the sub-image block obtained through the sliding window;

(2.2) Sparse coding: pass through the optimization model for each sub-image block:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D\&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}$

Get the sparse coding coefficient vector corresponding to each sub-image block in is the subdictionary D _i

Corresponding coefficient vector, allowable error ε=0.15, ||·|| ₁ is the l ₁ norm, ||·|| ₂ is the l ₂ norm;

(2.3) Find the reconstruction error: according to the sparse coding coefficient Calculate the reconstruction error of the sub-image block and each class Weight γ=0.5, take e=min{e _i } as the reconstruction error of this sub-image block, and record its corresponding category Then, according to the size relationship between the reconstruction error e and the preset threshold τ=0.3, it is determined whether the sub-image block contains the target: if e<τ, it means that the target is contained, otherwise, it means that the sub-image block is the background;

3. Target detection and recognition:

(3.1) Combine the reconstruction error e corresponding to each sub-image block that contains the target in (2.3) to form a reconstruction error matrix E=( _est ) that is consistent with the size of the test grayscale image and represents the candidate target area _P×Q ; where e _st is the value of the reconstruction error matrix at the coordinate point (s, t), $e_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ e& e<\mathrm{\&tau;}\end{array}\right.,$ P×Q is the size of the test image, s=1,2,...P, t=1,2,...Q; the category corresponding to each sub-image block determined to contain the target in (2.3) is composed of a test gray The category matrix L=(C _st ) _P×Q representing the candidate target category with the same size of the degree image; where C _st is the value of the category matrix at the coordinate point (s, t), $C_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ C& e<\mathrm{\&tau;}\end{array}\right.;$

(3.2) Change the size of the sliding window S×S, change S=90-10×j, j=1,2,...G as the number of changes, repeat 2, step (3.1) for a total of G times, and get G reconstructions Error matrix and G category matrices; the obtained G reconstruction error matrices form a multi-scale reconstruction error matrix MAP=(e _stg ) _P×Q×G ; where, e _stg is an element in the matrix MAP, and its value is the _est corresponding to the reconstruction error matrix obtained by changing the size of the sliding window for the gth time, P×Q×G is the size of the multi-scale reconstruction error matrix, g=1,2,...G; the G categories that will be obtained The matrix constitutes a multi-scale category matrix CLASS=(C _stg ) _P×Q×G ; wherein, C _stg is an element in the matrix CLASS, and its value is C _st corresponding to the category matrix obtained by changing the size of the sliding window for the gth time; According to the multi-scale reconstruction error matrix MAP, a minimum reconstruction error matrix (map(s,t)) _P×Q is obtained, where map(s,t) is the corresponding minimum reconstruction error matrix at the coordinate point (s,t) the value of Then find the minimum class matrix (class(s, t)) _P×Q corresponding to the minimum reconstruction error matrix, where class(s, t) is the value of the minimum class matrix at the coordinate point (s, t), Calculate the scale matrix according to the multi-scale reconstruction error matrix MAP scale(s,t) is the value of the corresponding scale matrix at the coordinate point (r,t),

$\mathrm{<span\; class="notranslate">\; scale</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}& {\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \underset{\mathrm{<span\; class="notranslate">\; g</span>}}{\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; \{</span>{\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; stg</span>}}<span\; class="notranslate">\; \}</span>& {\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; st</span>}}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

(3.3): Find the minimum reconstruction error matrix (map(s,t)) _P×Q local neighborhood minimum value as the detected target response value, the local neighborhood minimum value is in the minimum reconstruction error matrix ( The corresponding coordinates in map(s,t)) _P×Q are the center position of the target. According to the center position, the coordinates in (class(s,t)) _P×Q and (scale(s,t)) _P× The corresponding position in _Q finds the category and scale size corresponding to the target.

The formula for calculating the weighted average method is

f(x,y)=0.3R(x,y)+0.59G(x,y)+0.11B(x,y)

In the formula, f(x,y) is the grayscale value of the grayscale image at the pixel point (x,y) obtained by the weighted average method, and R(x,y), G(x,y) and B(x,y) ) are the RGB three component values of the input training image at the pixel point (x, y).

The energy normalization calculation formula is

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; the\; norm</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{\sqrt{\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; v</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}$

In the formula, f _norm (x, y) is the gray value of f(x, y) after energy normalization, u and v are the row and column sizes of the gray image, u=15, v=15.

The calculation formula of the _l1 norm is

${<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; |</span>$

In the formula, z is a vector with size M×1, ξ _k is the element of vector z, k=1,2,...,M.

The calculation formula of the ₁₂ norm is

${<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The calculation formula of the reconstruction error e _i is

${\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&gamma;</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

In the formula, γ is the preset weight, γ=0.5, m _i is the mean value vector obtained by averaging the elements of each row in Y _i ; Y _i is the optimal code obtained by sparse coding of U _i through dictionary D coefficient;

The effectiveness of the present invention is evaluated by selecting correct detection rate and false alarm rate. Among them, the correct detection rate is defined as the ratio of the number of correctly detected targets to the total number of targets, and the false alarm rate is defined as the ratio of the number of false alarms to the sum of the number of correctly detected targets and the number of false alarms. At the same time, the detection results obtained by the present invention are compared with the multi-class target detection algorithm based on BoW, and the comparison results are shown in Table 1. Both the correct detection rate and the false alarm rate show the validity of the method of the present invention.

Table 1 Test result evaluation

Claims (5)

1. A remote sensing image multi-category target detection and recognition method is characterized in that the steps are as follows: Step 1: Train the dictionary using the method based on sparse representation dictionary learning, the specific steps are as follows: Step a1 Pre-processing of training images: first unify the objects of the same category in the original image into one main direction, and then unify the direction of the image along 0° to 360°, according to the step size rotate as images in different directions; the original images of different types of targets are processed according to the above method, and the obtained Class training images, where p is the number of different categories of targets to be detected, is the rotation angle, c is the total number of categories of images in different directions for different targets in the obtained training image; where: is rounded down; Step b1 data preprocessing: use the weighted average method to The RGB three components of class training image carry out weighted average and obtain gray scale image, and described weighted average method calculation formula is f (x, y)=0.3R (x, y)+0.59G (x, y)+0.11B ( x, y), where f(x, y) is the gray value of the gray image at the pixel point (x, y) obtained by the weighted average method, R(x, y), G(x, y) and B(x, y) are the RGB three component values of the input training image at the pixel point (x, y); then the grayscale image is down-sampled to obtain an n×n size image; for n×n size The image of the energy normalization processing, the energy normalization calculation formula is In the formula, f _norm (x, y) is the gray value of f(x, y) after energy normalization, u and v are the row and column sizes of the gray image respectively; after energy normalization processing, we get After obtaining the normalized image, the normalized image is converted into n ² ×1-dimensional column vector, and the column vector is used as a column in the training data to obtain the preprocessed training data set U=[U ₁ , U ₂ , ..., U _c ], where U _i is the sub-data set corresponding to the i-th class in the training data set U, i=1,2,...,c; Step c1 training dictionary: use the FDDL software package released by Fisher Discrimination Dictionary Learning for SparseRepresentation to train the known training data set U=[U ₁ , U ₂ ,...,U _c ] to obtain the dictionary D=[D ₁ , D ₂ ,...,D _c ], where D _i is the sub-dictionary corresponding to the i-th class; Step 2 sparse coding: according to the dictionary D=[D ₁ ,D ₂ ,…,D _c ] obtained from training, perform sparse coding on each sub-image block in the test image, and obtain the corresponding sparse coefficient of each sub-image block, The specific processing steps are as follows: Step a2 Test image preprocessing: first use the weighted average method described in step b1 to convert the test image into a test grayscale image, and then use a sliding window of size S×S to slide along the test grayscale image with an interval step b Obtain the sub-image block; downsample the sub-image block to an image of size n×n, then perform energy normalization processing, and then convert the image after energy normalization processing into an n ² ×1-dimensional column vector β, using the column vector β to represent the pixel gray value information of the sub-image block obtained through the sliding window; Step b2 sparse coding: pass the optimization model for each sub-image block $\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; min</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; \&beta;</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; D\&alpha;</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&epsiv;</span>}$ Get the sparse coding coefficient corresponding to each sub-image block in is the coefficient vector corresponding to the sub-dictionary D _i , ε>0 is the allowable error, ||·|| ₁ is the l ₁ norm, and ||·|| ₂ is the l ₂ norm; the l _l norm The calculation formula is The calculation formula of the ₁₂ norm is In the two formulas, z is a vector of size M×1, ξ _k is the element of vector z, k=1,2,...,M Step c2 calculates the reconstruction error: according to the sparse coding coefficient Calculate the reconstruction error e _i of each sub-image block and each class, the calculation formula of the reconstruction error e _i is In the formula, γ is the preset weight, m _i is the mean vector obtained by averaging the elements of each row in Y _i ; Y _i is the optimal coding coefficient obtained by sparse coding of U _i through dictionary D; take e = min{e _i } as the reconstruction error of this sub-image block, and record its corresponding category Then, according to the size relationship between the reconstruction error e and the preset threshold τ, it is determined whether the sub-image block contains the target: if e<τ, it means that the target is contained, otherwise, it means that the sub-image block is the background; Step 3 target detection and recognition: Step a3: Combine the reconstruction error e corresponding to each sub-image block determined to contain the target in step c2 to form a reconstruction error matrix E=( _est ) _P that is consistent with the size of the test grayscale image and represents the candidate target area _{× Q} ; where, e _st is the value of the reconstruction error matrix at the coordinate point (s, t), $e_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ e& e<\mathrm{\&tau;}\end{array}\right.,$ P×Q is the size of the test image, s=1,2,...P, t=1,2,...Q; In step c2, the category C corresponding to each sub-image block that contains the target is determined to form a category matrix L=(C _st ) _P×Q that is consistent with the size of the test grayscale image and represents the candidate target category; where C _st is The value of the category matrix at the coordinate point (s, t), $C_{\mathrm{st}}=\left\{\begin{array}{cc}0& e\&Greater\; Equal;\mathrm{\&tau;}\\ C& e<\mathrm{\&tau;}\end{array}\right.,$ Step b3: Change the size of the sliding window S×S G times, repeat steps 2 to 3G times to obtain G reconstruction error matrices and G category matrices, and the value range of G is 5 to 10; the obtained G Reconstruction error matrices form a multi-scale reconstruction error matrix MAP=( _estg ) _P×Q×G ; among them, e _stg is an element in the matrix MAP, and its value is the reconstruction obtained by changing the size of the sliding window for the gth time The _est corresponding to the error matrix, P×Q×G is the size of the multi-scale reconstruction error matrix, g=1,2,...G; The obtained G category matrices constitute a multi-scale category matrix CLASS=(C _stg ) _P×Q×G ; where, C _stg is an element in the matrix CLASS, and its value is the category matrix obtained by changing the size of the sliding window for the gth time The corresponding C _st ; According to the multi-scale reconstruction error matrix MAP, a minimum reconstruction error matrix (map(s,t)) _P×Q is obtained, where map(s,t) is the corresponding minimum reconstruction error matrix at the coordinate point the value at (s,t), Then find the minimum class matrix (class(s, t)) _P×Q corresponding to the minimum reconstruction error matrix, where class(s, t) is the value of the minimum class matrix at the coordinate point (s, t), Calculate the scale matrix scale=(scale(s,t)) _P×Q according to the multi-scale reconstruction error matrix MAP, scale(s,t) is the value of the corresponding scale matrix at the coordinate point (r,t), $\mathrm{scale}(\mathrm{the\; s},t)=\left\{\begin{array}{cc}0& e_{\mathrm{st}}=0\\ \underset{g}{\arg \min }\left\{e_{\mathrm{stg}}\right\}& e_{\mathrm{st}}\&NotEqual;0\end{array}\right.;$ Step c3: Calculate the local neighborhood minima of the minimum reconstruction error matrix (map(s,t)) _P×Q as the detected target response value, and the local neighborhood minima are in the minimum reconstruction error matrix (map (s,t)) _P×Q corresponds to the center position of the target, according to the center position in (class(s,t)) _P×Q and (scale(s,t)) _P×Q corresponding Find the category and scale size corresponding to the target. 2. according to claim 1 described remote sensing image multi-class target detection and recognition method, it is characterized in that: described rotation angle The value range is from 0° to 90°. 3. according to the described remote sensing image multiclass target detection and identification method of claim 1, it is characterized in that: the scope of described FDDL software package parameter λ ₁ is 0.001～0.01, and the scope of λ ₂ is 0.01～0.1. 4. The multi-class target detection and recognition method for remote sensing images according to claim 1, characterized in that: the value range of S is an integer between 40 and 90, and the value range of b is an integer between 1 and 15. integer. 5 . The method for detecting and identifying multiple types of targets in remote sensing images according to claim 1 , wherein the value range of the threshold τ is 0-1. 6 .