CN106709501B - Scene matching area selection and reference image optimization method of image matching system - Google Patents

Abstract

The invention relates to a scene matching area selection and reference image optimization method of an image matching system, which is characterized in that blocks are initialized aiming at a remote sensing satellite picture, edge characteristics of each area are extracted according to an edge characteristic extraction algorithm, and block areas with concentrated edge gradient points are screened; calculating the autocorrelation degree of each region by using a repetitive pattern index measurement method, sequencing the regions from small to large according to the inner correlation index, and outputting an optimal candidate region; establishing a spatial feature vector of a reference map of the candidate region by adopting a spatial distribution description method; establishing a scene matchability measurement evaluation index set, analyzing the relevance between the evaluation index set and the matching probability, and outputting a matching evaluation index set; and statistically analyzing the matching correlation surface of the reference image and the satellite image, and optimizing the reference image according to the reference image optimization method. According to the invention, when the matching system works in an unknown environment, the optimized reference image is utilized to realize high matching performance and high reliability, and the image matching system is ensured to successfully complete tasks.

Description

A scene matching area selection and reference map optimization method for an image matching system

technical field

The invention relates to the technical field of image processing system performance evaluation, in particular to a scene matching area selection and reference map optimization method of an image matching system.

Background technique

In the scene matching system, the essence of the matching process is to calculate the similarity between the real-time image and the reference image. Therefore, the quality of the reference image is closely related to the scene matching accuracy. As the pre-planning information, the benchmark map is mainly derived from remote sensing images, and there are many differences in time and imaging system, which seriously affect the quality of the benchmark map. It is of great significance to improve the performance of the scene matching system to measure the evaluation indicators and criteria for high matching, and select the matching area and optimize the reference map according to these indicators.

The specific process of the selection and analysis of the adaptability of the scene area is essentially to determine the region of interest (ROI) that meets the basic indicators of the adaptability of the area. In these areas, a certain feature attribute is different from its adjacent areas, find a feature set that can fully reflect the regional adaptation performance and quantify the feature index, fuse the information of various feature indexes to form a comprehensive feature quantity, and propose various comprehensive features. The evaluation method for the adaptability of the scene area of the feature quantity, and the design principles of the feature index include:

(1) The degree of richness of scene information, the scene matching area must contain enough information to be able to match, and the richer the information, the more conducive to successful matching, such as the method of scene matching area selection based on information entropy proposed by Zhang Xiaochen et al.

(2) For stable features in the scene, due to the poor imaging quality, the features of the ground objects become blurred or even disappear, resulting in the failure of matching. Therefore, the feature index must ensure that the matching area has stable features, and the extraction targets such as Meth R. are stable. to estimate the stability boundary of the target;

(3) Reflect the uniqueness of features in the scene. If there are multiple similar obvious features in the selected matching area, the probability of successful matching will be greatly reduced. The feature index should be able to reflect the uniqueness of features, such as the introduction of Caves R G. Digital map, calculate its high matching area with the imaging image, as the index of feature selection;

(4) To reflect the obvious features in the scene, in order to achieve high matching accuracy, the matching position is obviously different from the non-matching position, the correlation peak is large enough, and the correlation peak is sharp enough. The number of suspected targets in the scale image, weighting the images at all levels, and designing the target salience index;

The suitability evaluation of comprehensive feature quantities is mostly included in the two basic theoretical systems of multi-attribute decision-making and pattern classification:

(1) Adaptability evaluation of scene area based on multi-attribute decision-making theory

The basic idea of this type of evaluation method is to abstract the adaptive evaluation process into a decision-making strategy, take each feature index as the basic attribute of decision-making, and construct a selected function through a specific decision-making model to form a comprehensive feature quantity. For example, Zhao Fengwei et al. The simple weighting of each feature extreme value in the region realizes the decision-making process. Cao Zhiguo et al. used the modified D-S theory to complete the multi-attribute decision-making, obtained the weighting coefficient of each global feature index through the orthogonal experimental design, and obtained the weighting coefficient through feedback correction to realize the decision-making process.

(2) Evaluation of scene area adaptability based on pattern classification theory

The basic idea of this type of evaluation method is to use each feature index value as perceptual information and design a classifier according to pre-selected classification criteria, thereby transforming the problem of matching probability estimation into the problem of classifying pixels or regions, as proposed by Li Jun et al. The Mumford-Shah model, through the evolution of the level set curve, obtains the optimal division of the two sets of matching stable local regions and unstable local regions.

Therefore, the current benchmark map selection is mainly focused on the analysis of image grayscale information or the evaluation of feature quality, and is not related to the matching performance to optimize the benchmark map in advance, especially in the image matching system, most of the work of the image matching system In the infrared imaging system, the preparation of the prior reference map of the image matching system should be calibrated with the homologous test image, that is, the infrared image sequence. However, in the actual unknown environment, it is difficult to obtain homologous infrared image sequences, while heterologous images of visible light are very easy to obtain, such as satellite images and aerial images. The analysis error brought by this imaging system is less concerned. , all of which lead to a decrease in the matching probability. This requires the establishment of a reference map selection and evaluation method for heterologous images, which can adapt to the conditions of heterologous image preparation, and at the same time has the function of optimizing the reference map in advance, so that the reference map preparation is close to the high-quality target. Methods for preparing and optimizing benchmark maps have not been published.

SUMMARY OF THE INVENTION

In view of the deficiencies of the prior art, the present invention provides a method that utilizes the edge feature similarity of heterologous images, uses remote sensing satellite images, establishes a set of internal correlation and external correlation evaluation indicators, supervises the preparation process of the reference map, and completes the selection of regions and the optimization of the reference map. Methods.

The technical scheme that the present invention adopts for realizing the above-mentioned purpose is:

A scene matching area selection and reference map optimization method for an image matching system, comprising the following steps:

Step 1: Initialize the segmentation for remote sensing satellite images, extract the edge features of each area according to the edge feature extraction algorithm, and screen the segmented areas in the edge gradient point set;

Step 2: utilize the repeated pattern index measurement method to calculate the autocorrelation degree of each region, and according to the internal correlation index, the regions are sorted from small to large, and the optimal candidate region is output;

Step 3: Use the spatial distribution description method to establish the spatial feature vector of the candidate area reference map;

Step 4: establishing a set of evaluation indicators for the scene matchability measurement, analyzing the correlation between the evaluation indicator set and the matching probability, and outputting a matching evaluation indicator set;

Step 5: Statistically analyze the matching correlation surface between the benchmark map and the satellite image, and optimize the benchmark map according to the benchmark map optimization method.

The edge feature extraction algorithm is a Canny operator edge feature extraction algorithm.

The said screening of the block area in the edge gradient point set includes the following process:

Step 1: For each block sub-region, use the Canny operator to extract edge features;

Step 2: Calculate the gradient change histogram distribution of each sub-region;

Step 3: Compare the gradient point set distribution of each sub-region, and select one or more regions in the point set set as candidate regions according to the gradient change from small to large.

The method for measuring repetition pattern indicators includes the following steps:

Step 1: For each block area, select a sub-image i, and the calculation formula of the repetition pattern index is:

where s is the number of tiles involved in the matching calculation when sub-picture i is matched point by point on the candidate region; Pi is the number of tiles whose edge point correlation coefficient between s tiles and sub-picture i is greater than the threshold during the matching process;

Step 2: The formula for calculating the repetition pattern of the reference graph is:

Among them, cf is the measure of the repetition pattern of the entire block; n is the number of selected sub-blocks.

The described establishment of the spatial feature vector of the candidate region reference map by using the spatial distribution description method includes the following processes:

Step 1: Extract features from the benchmark map, express them as directional features and scale features, and establish a feature sensitivity analysis test;

Step 2: According to the minimum distinguishable interval of the directional feature, the reference image of the reference image area is divided into intervals;

Step 3: Count the number of scale points in the interval, and establish a multi-dimensional feature vector in the form of {X ₁ , X ₂ , X ₃ , X ₄ , X ₅ , X ₆ }.

The benchmark graph optimization method is:

Step 1: Use the Harsdoff distance matching method to complete the matching process, and calculate the matching metrics;

Step 2: After judging the pros and cons of the matching metrics, calculate the main influence interval of the main peak and the secondary peak;

Step 3: Correct the number of edges in the interval, optimize the benchmark graph, and recalculate the matching metrics until the matching threshold requirements are met.

The present invention has the following beneficial effects and advantages:

1. The method of the present invention adopts the image feature analysis method, and for the image edge feature information, uses the self-similar index to search and evaluate the most matchable candidate region, and uses the reference map spatial distribution description to establish a feature vector, establish an evaluation index, and measure the matching system. Match the relevant surface, modify and optimize the eigenvector;

2. The method of the present invention is particularly suitable for the preparation of the reference map of the heterologous matching system. Through the analysis and evaluation of the internal correlation and the external correlation, the matchability is improved, and a supervision index system is established to measure the quality of the reference map in real time, and guide and evaluate the quality of the reference map in real time. Optimize the datum map to realize the preparation of high-quality datum map.

Description of drawings

Fig. 1 is the overall flow chart of the present invention;

Fig. 2 is a sub-region histogram comparison diagram of the present invention;

Figure 3 is the spatial distribution analysis diagram of the benchmark map, (a) is the direction feature sensitivity analysis diagram, (b) is the scale feature sensitivity analysis diagram;

FIG. 4 is a matching index curve diagram after the optimization of the method of the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

As shown in Figure 1, the scene matching area selection and reference map optimization method of the image matching system of the present invention includes the following steps:

(1) The edge features of each region are extracted in blocks, and the block regions with stable edge gradient points are screened;

(2) The degree of autocorrelation of each region is measured by the repetition pattern index, and the regions are sorted according to the internal correlation index, and the optimal candidate region is output;

(3) Using the spatial distribution description method to establish the spatial feature vector of the candidate region reference map;

(4) Establish a set of evaluation indicators for scene matchability measurement, and analyze the correlation between the indicator set and the matching probability;

(5) Statistically analyze the matching correlation surface between the benchmark map and the satellite image, calculate the matchability metrics, correct the feature vector, guide and optimize the benchmark map, and end the process of benchmark map preparation and optimization.

The process of extracting the edge features of each region and screening the block regions with stable edge gradient points:

According to the satellite image of the target area, it is divided into multiple sub-areas, as shown in Figure 2, the edge detection of the image is to use the discretized gradient approximation function to find the grayscale transition of the image grayscale matrix according to the gradient vector of the two-dimensional grayscale matrix. Change the position, and then connect the points at these positions in the image to form the so-called image edge. The selection strategy here is to select the sub-regions where the gradient edges are concentrated and stable. This selection method is also to ensure that the heterologous infrared system imaging can also have stable edge information and ensure the similarity of heterologous images.

The gradient is divided into 4 levels, and the sample statistics of the 4 levels are compared, as shown in Table 1. Select areas with obvious concentration, and the candidate areas selected for this target area are area 5 and area 7.

Table 1 Histogram statistics of gradient features of sub-regions

sub area code Gradient Level 1 Gradient Level 2 Gradient Level 3 Gradient Level 4 1 401695 210796 120432 77159 2 456190 199114 109368 69573 3 611629 146493 68119 41373 4 505308 172192 93492 61521 5 710924 146338 41584 22355 6 517751 171202 83510 51910 7 782992 91640 30488 18005 8 583435 151043 68775 43351 9 419121 187095 111749 74036

The process of measuring the autocorrelation degree of each area by using the repeating pattern index, sorting the areas according to the internal correlation index, and outputting the optimal candidate area;

Step 1: For each block area, select a sub-image i, and the calculation formula of its repetition pattern is:

In the formula, S is the number of tiles participating in the matching calculation when the sub-image i is matched point by point on the candidate region;

Pi is the number of blocks whose edge point correlation coefficients between s blocks and sub-image i are greater than the threshold during the matching process.

Step 2: The formula for calculating the repetition pattern of the reference graph is:

Step 3: Sort from small to large according to the repeating pattern index, and select the area with the smallest repeating index as the final output area of the reference image.

Table 2 Correlation comparison table in sub-regions

sub area code Area 5 Area 7 position 1 0 0.020117191 position 2 0 0 position 3 0 0.017970387 position 4 0.013828491 0 position 5 0.017665928 0.030609096 position 6 0.02223386 0.02458329 position 7 0.005933324 0.030570649 position 8 0.015193883 0.02530859 position 9 0.016019976 0.000337711 integrated repeat pattern 0.010097273 0.016610768

According to the principle of the lowest repetition pattern, the sub-region 5 is determined here as the final reference map preparation region.

The described process of using the spatial distribution description method to establish the spatial feature vector of the candidate region reference map is as follows:

The model is extracted from the benchmark map, expressed as a direction mode and a scale mode. The direction attribute is selected according to the quadrant level, and the scale attribute is selected according to the number of points of the contour length. The sensitivity analysis test of the design mode is shown in Figure 3. , which proves that the orientation and scale properties of the benchmark map are highly sensitive to the matching results.

Through the analysis of the actual scene, the targets of interest are all man-made objects, and the edge angles of the man-made objects in all directions are more than 30°. The edge direction distribution is too large to achieve the distribution of template features. Then there may be redundant information, which complicates the template standardization process. Therefore, the division accuracy of the selected direction attribute is selected as follows: [0,30),[30,60),[60,90),[90,120),[120,150),[150,180), and according to the contour lines of each interval The projection length calculates the scale, and finally a six-dimensional feature vector of the form {X ₁ , X ₂ , X ₃ , X ₄ , X ₅ , X ₆ } can be established. The six-dimensional feature vector of this target is: {45, 349, 52, 56, 4545, 167}.

The described selection of the scene matchability metric evaluation index set is used to verify the correlation analysis process between the index set and the matching probability:

Referring to the matching performance measurement method, three indicators for measuring the matching correlation surface are selected from local indicators and global indicators, as shown in Table 3.

Table 3 Set of significant indicators

serial number index Indicator type 1 peak-to-side lobe ratio local indicator 2 sharpness local indicator 3 global saliency Global metrics

Design the index monotonicity verification experiment, decompose the contour lines of the benchmark map, sort the matching scores of each contour line, and delete the contour lines where the sub-peaks of the matching related faces are located in turn to verify the monotonicity of these indicators. The results are shown in Table 4. .

Table 4 Validation results of indicator monotonicity

serial number peak-to-side lobe ratio sharpness global saliency 1 3.34 0.42 37.8 2 3.34 0.31 35.7 3 3.3 0.41 65.3

4 4.55 0.47 58.5 5 5.64 0.48 72.8 6 6.82 0.47 71.6 7 6.81 0.47 75.6 8 7.58 0.43 82.2 9 7.84 0.4 83.3 10 8.09 0.38 83.5 11 8.23 0.38 85.3 12 8.41 0.4 79.5 13 8.54 0.41 74.5 14 8.59 0.44 70 15 9.67 0.45 69.5 16 9.76 0.46 69 17 9.86 0.5 68 18 10 0.51 67 19 9.98 0.51 65 20 10.12 0.52 66 twenty one 9.98 0.53 66 twenty two 10.12 0.54 66 twenty three 10.42 0.55 65 twenty four 10.65 0.56 62 25 9.87 0.67 59 26 8.96 0.75 0

The results of the monotonicity verification show that the local indicators have obvious monotonicity, while the global indicators are non-monotonic. Finally, the peak sidelobe ratio and the sharpness are selected as the supervision indicators for benchmark graph optimization.

Figure 4 shows the test image matching result. The matching used iteratively based on the peak sidelobe ratio and the sharpness, and corrected the reference map of the corresponding quadrant. Until the matchability metric meets the threshold requirements, the final optimized reference map improves the matching probability, and completes the scene matching area selection and reference map optimization process of the matching system.

Claims (7)

1. a scene matching area selection and a reference map optimization method of an image matching system, it is characterized in that: comprise the following steps: Step 1: Initialize the segmentation for remote sensing satellite images, extract the edge features of each area according to the edge feature extraction algorithm, and screen the segmented areas in the edge gradient point set; Step 2: utilize the repeated pattern index measurement method to calculate the autocorrelation degree of each region, and according to the internal correlation index, the regions are sorted from small to large, and the optimal candidate region is output; Step 3: Use the spatial distribution description method to establish the spatial feature vector of the candidate area reference map; Step 4: establishing a set of evaluation indicators for the scene matchability measurement, analyzing the correlation between the evaluation indicator set and the matching probability, and outputting a matching evaluation indicator set; Step 5: Statistically analyze the matching correlation surface between the benchmark map and the satellite image, and optimize the benchmark map according to the benchmark map optimization method. 2 . The scene matching area selection and reference map optimization method of an image matching system according to claim 1 , wherein the edge feature extraction algorithm is a Canny operator edge feature extraction algorithm. 3 . 3. The scene matching area selection and reference map optimization method of a kind of image matching system according to claim 1, it is characterized in that: described screening edge gradient points concentrated block area comprises following process: Step 1: For each block sub-region, use the Canny operator to extract edge features; Step 2: Calculate the gradient change histogram distribution of each sub-region; Step 3: Compare the gradient point set distribution of each sub-region, and select one or more regions in the point set set as candidate regions according to the gradient change from small to large. 4. The scene matching area selection and reference map optimization method of a kind of image matching system according to claim 1, is characterized in that: described repetition pattern index measurement method comprises the following steps: Step 1: For each block area, select a sub-image i, and the calculation formula of the repetition pattern index is:

where s is the number of tiles involved in the matching calculation when sub-picture i is matched point by point on the candidate region; Pi is the number of tiles whose edge point correlation coefficient between s tiles and sub-picture i is greater than the threshold during the matching process; Step 2: The formula for calculating the repetition pattern of the reference graph is:

Among them, cf is the measure of the repetition pattern of the entire block; n is the number of selected sub-blocks. 5. the scene matching area selection and reference map optimization method of a kind of image matching system according to claim 1, is characterized in that: described adopting the spatial distribution description method to establish the spatial feature vector of the candidate region reference map comprises the following process: Step 1: Extract features from the benchmark map, express them as directional features and scale features, and establish a feature sensitivity analysis test; Step 2: According to the minimum distinguishable interval of the directional feature, the reference image of the reference image area is divided into intervals; Step 3: Count the number of scale points in the interval, and establish a multi-dimensional feature vector. 6. The scene matching area selection and reference map optimization method of an image matching system according to claim 5, wherein the multi-dimensional feature vector is {X ₁ , X ₂ , X ₃ , X ₄ , X ₅ ,X ₆ }, where X _i is the number of points in a single interval. 7. The scene matching area selection and reference map optimization method of a kind of image matching system according to claim 1, it is characterized in that: described reference map optimization method is: Step 1: Use the Harsdoff distance matching method to complete the matching process, and calculate the matching metrics; Step 2: After judging the pros and cons of the matching metrics, calculate the main influence interval of the main peak and the secondary peak; Step 3: Correct the number of edges in the interval, optimize the benchmark graph, and recalculate the matching metrics until the matching threshold requirements are met.

Images