An Image Flying Target Recognition Method Based on Grid Classification and Geometric Reconstruction

Abstract

The present invention relates to an image flight target recognition method based on grid classification and geometric reconstruction. First, a target pre-classification method based on grid-shaped classification units is proposed to achieve target pre-classification, and then the target attitude is reconstructed, including flight attitude normalization and image preprocessing. At the same time, a multi-scale analysis method is used to extract the low-frequency and high-frequency feature vectors of the image. Finally, an identification model based on hidden semi Markov models (HSMM) is constructed. After the extracted feature vectors are selected by attitude, the CHSMM model is used to identify the feature vectors to obtain the identified target result. The accuracy of this recognition method is tested by using non-overlapping picture materials of different aircraft models in the picture library. The results show that the recognition rate is high and the recognition time is short. It is proved that the invention has high engineering applicability and broad application prospects in the field of target detection and perception.

Description

Technical Field

[0001] The invention belongs to the field of mobile target image recognition in radar data processing. Background Art

[0002] The main difficulties in identifying image targets are: 1. The images of flying targets are usually captured, with diverse postures and different geometric features, which increases the difficulty of identification. How to reconstruct the flight posture and preprocess the image; 2. The method of extracting feature vectors to enable it to have good target representation capabilities; 3. The pattern recognition method to enable high recognition accuracy and short time.

[0003] Currently in the field of target recognition, there are many methods for extracting image feature vectors, such as using wavelet transform. However, wavelet transform reflects the point singularity of the signal and cannot accurately describe the direction of the image edge, nor can it achieve sparse representation of the image, thus affecting recognition accuracy. There are also many recognition methods, some of which use corner features and kernel clustering algorithms, some use wavelet transform for features, some are based on closed contour features, and some use template matching and so on. Summary of the invention

[0004] In view of the problems that the existing target recognition methods have long operation time, low recognition efficiency and cannot be applied in projects with high real-time requirements, the present invention provides an image flying target recognition method based on grid classification and geometric reconstruction. Firstly, a target pre-classification method based on grid-like classification units is designed to achieve target pre-classification, and then the flight attitude is geometrically reconstructed, and then a multi-scale analysis method is used to extract the low-frequency feature vector and high-frequency feature vector of the image. Finally, a model based on hidden semi-Markov is used to quickly and effectively realize the recognition of image flying targets of different models, multiple flight attitudes and different lighting changes.

[0005] The technical solution to realize the present invention includes:

[0006] Step 1: Based on the grid-like classification unit pre-classification method, a chain search is performed within the target contour area to complete the target pre-classification;

[0007] Step 2: geometrically reconstruct the flight attitude and perform image preprocessing on the flight target images in the library;

[0008] Step 3: Use a two-layer Contourlet transform on the target image in the model library to extract the low-frequency features that express the contour and the high-frequency feature vectors that express the detail information as the training set;

[0009] Step 4: Modeling of the recognition model based on the hidden semi-Markov model: Classify the feature data postures and set the initial values through different flight postures and the extracted information, generate training samples of multiple groups of observation sequences, and extract l groups of observation sequence vectors for k postures of the training samples: o k =[o(1), o(2), …, o(l)], k = 1, 2…k. Take the observation sequence vector o k =[o(1), o(2), …, o(l)], k = 1, 2…k of each posture as the training data to train the CHSMM i model, so as to obtain classifiers for k postures;

[0010] Step 5: Image flight target recognition: The method of target recognition is to recognize the feature vector through the CHSMM model. Input the test sample into k classifiers to obtain k output probability values, and select the one with the largest output probability value as the recognized posture by comparison.

[0011] Compared with the existing methods, the significant advantages of the present invention are as follows:

[0012] 1. While preprocessing the target image, the present invention designs a matching model according to the respective characteristics of different aircraft types in the image. There is a very bright tail flame at the tail of a fighter jet, a propeller at the top of a helicopter, and the pixels within the overall contour of a civil aircraft are very uniform. Therefore, different classification models are designed according to the characteristics of different aircraft types. As Figure 4 shown, the dotted line box represents the target contour, and the 3*3 grid represents the basic matching grid unit. The length and width of the target area contour are h and w. The grid unit A is the contour center. The matching grid unit can be selected in real time according to the target area to be classified. The elements within the target connected area are p i , the pixel mean is The pixel maximum value is p max , and the pixel value within the grid is The covariance within the grid is:

[0013]

[0014] Establish a discriminant function as shown in Equation (5) according to the grid covariance and the target area maximum value,

[0015]

[0016] According to this discriminant function, set the discrimination strategy to achieve target pre-classification.

[0017] 2. Through geometric reconstruction, adjust the target posture of the captured image, providing an important basis for multi-scale analysis of the image.

[0018] 3. The Contourlet transform is used to extract the low-frequency features and high-frequency energy features respectively, which has better approximation energy than the wavelet transform, has a more "sparse" representation of curves and straight lines in images, and is more conducive to preprocessing the images to extract the target feature information.

[0019] 4. The modeling method based on hidden semi Markov models (HSMM) is a kind of HMM that considers the pose residence probability distribution as explicit. It adds a time component to the defined HMM structure, allowing each state to have a variable duration, thus overcoming the limitations of HMM modeling caused by the assumptions of Markov chains, and improving both the pattern classification ability and accuracy.

[0020] The present invention will be further described in detail below with reference to the accompanying drawings. Description of the Drawings

[0021] Figure 1 It is a comparison diagram for image geometric reconstruction and gray normalization.

[0022] Figure 2 It is a 16-direction distribution diagram of the high-frequency features of a certain flight target image.

[0023] Figure 3 It is the overall flow chart of the recognition method.

[0024] Figure 4 It is a schematic diagram of the classification model. Specific Embodiments

[0025] The preferred embodiments of the present invention are as follows:

[0026] 1. Pre-classification method based on grid-like classification units

[0027] According to different aircraft models to be recognized, set the search interval. As Figure 4 shown, according to the discriminant function formula (5), calculate the means of the matching grids A, B, and C. As shown in formula (6):

[0028]

[0029] Take f of matrix A A as the discriminant threshold of this target. Because of the unique propeller of the helicopter, it is necessary to use the matching grids D, E, and F for joint discrimination, and the strategy is as shown in formula (7):

[0030]

[0031] Set the pre-classification strategy as shown in Equation (8). When F = 1, the pre-classification is a fighter plane; when F = 2, the pre-classification is a civil aviation plane; when F = 3, the pre-classification is a helicopter.

[0032]

[0033] 2. Geometric reconstruction of the flight target attitude

[0034] The flight targets captured by the optical camera have diverse attitudes, which cause certain interference to recognition. Now, the flight attitude is reconstructed to adjust it to the nose-up position. It is necessary to extract the center point of the flight target and perform equal-scale translation and rotation along its fuselage axis to normalize the flight target into an image with the nose up.

[0035] First, extract the center origin of the aircraft

[0036]

[0037]

[0038] Where f(x′, y′) is the image function. When (x′, y′) belongs to the target area, f(x′, y′) = 1; when (x′, y′) does not belong to the target area, f(x′, y′) = 0. a, b, c, and d are the boundary coordinates of the tracking window.

[0039] After obtaining the dot, draw straight lines around it from 0 to 360 degrees, and find the edge points of the aircraft image on this straight line. The angle α corresponding to the line segment with the longest distance from the edge point to the centroid is the axis direction of the aircraft image, and the position where this edge point is located is the nose position. Similarly, search for the target edge points from the nose through the centroid to the tail to find the tail coordinates. Rotating the axis direction along the angle α can convert the flight target image into a standard image with the nose up.

[0040] The preprocessing of the image is divided into two steps. The first step is to perform geometric standardization processing on the picture. Since the position and size of the target in the entire image are different, it will affect the recognition effect. It is necessary to correct the target image through methods such as calibration and positioning. In the present invention, the area of the target to be recognized is enlarged and the image is calibrated to 64×64 pixels, making the target features more obvious. The second step is to perform gray-scale normalization processing. Since different images are greatly affected by the environment, it is necessary to compensate the images obtained under different lighting conditions and light directions to weaken the image signal changes under different environmental conditions and facilitate the extraction of subsequent feature vectors. The comparison diagrams before and after the geometric reconstruction of the flight attitude are as Figure 1 shown.

[0041] 3. Image feature extraction based on Contourlet transform

[0042] Feature extraction of the target image is a crucial step in the recognition process. By extracting feature quantities, the contour features of the target image are characterized, providing input parameters for the subsequent model. This method uses the Contourlet transform to extract low-frequency features and high-frequency energy features respectively.

[0043] The Contourlet transform is an image multi-scale geometric analysis tool based on non-separable filter banks. Its idea is to approximate the image through a basis structure similar to contour segments. Its algorithm inherits the multi-resolution and time-frequency localization characteristics of wavelets, and at the same time has good directionality and anisotropy, overcoming the limitations of wavelet transform that can only reflect the point singularity of signals, cannot accurately describe the direction of image edges, and cannot achieve sparse representation of images. The multi-scale geometric analysis method was precisely developed to overcome this limitation of wavelets. For a twice continuously differentiable target function f ∈ L 2 (R 2 ), if f m is its reconstruction function and m is its coefficient weight, then when performing the Contourlet transform, the error approximation satisfies

[0044]

[0045] When using the wavelet transform, the error approximation satisfies

[0046]

[0047] It can be seen from this that the Contourlet transform has better approximation energy than the wavelet transform, has a more "sparse" expression for curves and straight lines in the image, and is more conducive to preprocessing the image to extract target feature information.

[0048] The low-frequency feature vector contains most of the energy information and can smoothly approximate the contour information of the image. However, in the actual collected pictures, the target images taken under different lighting conditions contain high-frequency noise, and the low-frequency components of the Contourlet transform are not sensitive to light changes. The high-frequency coefficients can reflect information such as edges and details in all directions of the target, and are features with high recognition efficiency. Figure 2 The 16-direction high-frequency features extracted by performing the Contourlet transform on the helicopter image are the distribution of the decomposed target edge and contour information in all directions and positions.

[0049] 4. Modeling based on the hidden semi-Markov model

[0050] The modeling method based on hidden semi - Markov models (HSMM) is a type of HMM that considers the pose dwell probability distribution to be explicit. It adds a time component to the defined HMM structure, allowing each state to have a variable duration. Therefore, it overcomes the limitations of HMM modeling caused by the assumptions of the Markov chain, and both the pattern classification ability and accuracy are improved.

[0051] Compared with HMM, HSMM adds a probability distribution function of the dwell time in form to represent the dwell time of each pose. It is defined as:

[0052] P j (d)=(a n ) d (1 - a n ) (13)

[0053] where d is the number of time units for which a certain pose persists.

[0054] For the initial parameters of the prediction model HSMM

[0055] λ=(A,B,π,P j (d)) (14)

[0056] where: A refers to the transition probability matrix of the hidden poses in the Markov chain during the HSMM process, denoted as:

[0057] A={a ij} N×n (15)

[0058] where:

[0059] a ij =P(q i+1 =S j |q i =S i ),i≤j,j≤N (16)

[0060] B refers to the probability matrix of all possible observed values that the N poses in the Markov chain during the HSMM process may obtain. Denoted as

[0061] B=(b jk ) N×M (17)

[0062] where:

[0063] b jk =P(o t =V k |q t=S t ), 1 ≤ j ≤ N, 1 ≤ k ≤ M (18)

[0064] Among them, the initial parameter A and the initial parameter π have little influence on the training result of the HSMM-based model, while the value of the initial parameter B has a great influence on the training result of the HSMM-based model according to the weight or different settings.

[0065] In the embodiment of the present invention, the feature data attitude classification and the initial value setting are performed through different flight postures and the extracted information, generating training samples of multiple groups of observation sequences. For the k postures of the training samples, l groups of observation sequence vectors are extracted: o k =[o(1), o(2), …, o(l)], k = 1, 2 … k.

[0066] Regarding the observation sequence vector o of each posture k =[o(1), o(2), …, o(l)], k = 1, 2 … k as the training data to train the CHSMM i model, thereby obtaining classifiers for k postures.

[0067] 5. Image flight target recognition

[0068] The feature vector is recognized through the CHSMM model. The test sample is input into k classifiers to obtain k output probability values. By comparison, the one with the largest output probability value is selected as the recognized posture.

[0069] The present invention performs image preprocessing, feature extraction and model recognition on 80 different aircraft type flight images in the library, and randomly selects 3 non-repeating images for each type of fighter plane, helicopter and civil airliner for testing. The recognition rates are shown in the following table. The recognition rate is relatively high, which can illustrate the effectiveness of the present invention.

[0070] Table 1 Recognition results of three aircraft type pictures

[0071] Aircraft model Recognition rate Fighter jet 98.13％ Helicopter 99.93％ Civil airliner 99.60％

[0072] The test results show that the target recognition rate of the present invention is high, with good accuracy, and can meet the accurate recognition requirements of different types of air targets. At the same time, the present invention has good robustness to image brightness and illumination changes and target attitude transformations.

Claims

1. An image flight target recognition method based on grid classification and geometric reconstruction, characterized in that: Step 1: Based on the pre-classification method of grid-shaped classification units, perform a chain search within the target contour area to complete the target pre-classification; The pre-classification method based on grid-shaped classification units includes: Set the search interval according to the flying target to be recognized, and use the discriminant function Statistically calculate the means of the matching grids A, B, and C As shown in Equation (1): Among them, p max is the pixel maximum value, and σ cov is the covariance within the grid; The elements in the target connected region are p i , and the pixel mean value is The pixel value within the grid is Using the f A of matrix A as the discrimination threshold for this target, it is necessary to use the matching grids D, E, and F for joint discrimination, and the strategy is shown in Equation (2): Set the pre-classification strategy as shown in Equation (3). When F = 1, the pre-classification is a fighter plane; when F = 2, the pre-classification is a civil aviation aircraft; when F = 3, the pre-classification is a helicopter: Step 2: Perform geometric reconstruction of the flight attitude and image preprocessing on the flight target images in the library. Step 3: Apply a 2-layer Contourlet transform to the target images in the aircraft type library, extract the low-frequency features expressing the contour and the high-frequency feature vectors expressing the detailed information as the training set. Step 4: Modeling of the recognition model based on the hidden semi-Markov model: Classify the feature data postures and set the initial values according to different flight postures and the extracted information, generate training samples of multiple groups of observation sequences, and extract l groups of observation sequence vectors for k postures of the training samples: o k = [o(1), o(2), …, o(l)], k = 1, 2…k. Take the observation sequence vector o k = [o(1), o(2), …, o(l)], k = 1, 2…k of each posture as the training data to train the CHSMM i model, so as to obtain classifiers for k postures; Step 5: Image flight target recognition: The method of target recognition is to recognize the feature vectors through the CHSMM model. The test samples are input into k classifiers to obtain k output probability values, and the one with the largest output probability value is selected as the final attitude recognition result by comparison.

2. The image flight target recognition method based on grid classification and geometric reconstruction according to claim 1, characterized in that: The said Step 2 includes: Extract the center point of the aircraft, and then extract the central axis of the fuselage. Rotating the axis direction by a certain angle α can reconstruct the flight target image into a standard image with the nose up; secondly, perform geometric normalization processing and gray-scale normalization processing on the image.

3. The image flight target recognition method based on grid classification and geometric reconstruction according to claim 1, characterized in that: Group the feature vectors. The number of groups is determined according to the size of the picture library or the number of aircraft types, and use the high-frequency detailed features to correct the contour information expressed by the low-frequency features.

Images