Image feature extraction algorithm in big data environment

Abstract

The purpose of this article is to explore effective image feature extraction algorithms in the context of big data, and to mine their potential information from complex image data. Based on the BRISK and SIFT algorithms, this paper proposes an image feature extraction and matching algorithm based on BRISK corner points. By combining the SIFT scale space and the BRISK algorithm, a new scale space construction method is proposed. The BRISK algorithm extracts the corner invariant features. Then, by using the improved feature matching method and eliminating the mismatching algorithm, the exact matching of the images is realized. A large number of experimental verifications were performed in the standard test Mikolajczyk image database and aerial image database. The experimental results show that the improved algorithm in this paper is an effective image matching algorithm. The highest accuracy of actual aerial image matching can reach 85.19%, and it can realize the actual aerial image matching that BRISK and SIFT algorithms cannot complete. The improved algorithm in this paper has the advantages of higher matching accuracy and strong robustness.

Keywords

Big data image feature extraction corner point brisk algorithm sift algorithm

1 Introduction

With the advent of the era of big data, image data has exploded [1]. In the face of increasing image data, how to effectively mine the effective information from huge image data, to accurately and quickly find the images that users need has gradually developed into a research hotspot in the field of computer vision. At the same time, image feature extraction is also a difficult point in this field, especially in the matching of drone aerial images and Google earth aerial images. Because there are differences in the same object in aerial images at different resolutions, the matching result is affected. And the image obtained by Google earth is not updated in real-time. Compared with the aerial image obtained by the drone, the ground information of the building in the image has obvious differences. In some scene matching, image content is not only easily affected by scale changes, rotation changes, lighting changes, and noise interference, but also limited by the characteristics of the shooting camera. These objective conditions restrict the performance of matching to a certain extent. Therefore, it is very difficult to achieve accurate image matching under such harsh conditions.

With the continuous development of the Internet and information technology, a large number of digital images continue to appear in people’s lives through various channels every day. Therefore, it is of great theoretical significance and practical value to study the large-scale image classification problem in depth. In the practical application of aerial image matching, the matching of some aerial building scenes is affected by unfavorable factors such as scale, rotation, lighting, noise, etc., which restricts the matching performance to a certain extent. In current image feature extraction and matching algorithms, compared with other texture and color information, corner points provide a two-dimensional comparison relationship more accurately. Corner point features are distributed on the edges of building structures, maintaining the importance of object edge features. It has high positioning information in the image. In the case where the ground information has changed significantly, the correspondence between the corner points can still be found in some areas that have not changed significantly, which can better achieve the matching of such images, and the calculation is simple and the image has better Robustness [2].

Qin’s team researched the latest image feature detection algorithms and focused on scalar invariant feature transformation (SIFT). It is one of the most important local feature detection algorithms and has been widely used in different fields, including object recognition, image matching, Robot mapping. They analyze and model the privacy requirements in outsourced SIFT calculations, and propose a secure scalar invariant feature transform (SecSIFT), a high-performance privacy-protected SIFT feature detection system. Contrary to previous work, the proposed design is not limited by the efficiency of current homomorphic encryption schemes. In their design, they decompose and distribute the calculation process of the original SIFT algorithm to a set of independent, collaborative cloud servers, and simplify the outsourced calculation process as much as possible to avoid the use of computationally expensive homomorphic encryption schemes. The proposed SecSIFT enables the implementation with practical computational and communication complexity. A large number of experimental results show that the performance of SecSIFT in image benchmarking is comparable to that of original SIFT, and at the same time, it can protect privacy in an effective manner [3]. Li’s team proposed a distance-Doppler image rotation parameter extraction method based on the time difference and image edge detection. This method first uses the method of image edge detection to detect the motion trajectory of a moving pixel caused by the rotating part in the time-domain differential distance-Doppler image. A Doppler slow-time image is then generated from the edge pixels on the motion trajectory. Finally, the rotation parameters are extracted from the Doppler slow-time image. The method is simple, fast and practical. Computer simulation and experimental results show that the method is effective in terms of calculation time compared with existing methods [4]. Lei Tian’s team has proposed a multi-feature extraction method for objects in an image to improve image-based tactile perception. Multi-feature extraction includes contour shape extraction, pattern extraction and detail texture extraction. First, they used an inherent decomposition method to decompose the image into a shadow image and a reflectance image. The reflectivity image describes a pattern of colors scattered on the surface that is unaffected by illumination. The shadow image is then used for contour shape and detail texture extraction. Contour shape extraction is based on partial differential equations (PDE) to reconstruct a three-dimensional (3D) surface model in a virtual environment. Detailed texture extraction is also based on the fractional difference method. Finally, the various features extracted above are subjected to tactile rendering by different methods. Experimental results show that the method improves human tactile perception and recognition in a virtual environment [5].

Due to the unfavorable factors such as scale, rotation, light, noise, etc., there is often a loss of edge information in aerial buildings, which results in that only a small number of speckle features can be extracted on the structural edges of the image. In contrast, the corners can maintain the stability of the edge features of the object. However, the current corner algorithm has disadvantages in matching accuracy. To this end, this paper proposes an image feature extraction algorithm based on BRISK corners. By combining the SIFT scale space with the BRISK algorithm, a new scale space construction method is proposed, and then the BRISK algorithm is used to extract the corner invariant feature of the scale. Then, by using the improved feature matching method and eliminating the mismatching algorithm, the exact matching of the images is realized. The experimental results show that the algorithm in this paper is an effective image matching algorithm, which can realize the actual aerial image matching that BRISK and SIFT algorithms cannot complete.

2 Design of image feature extraction algorithm in big data environment

2.1 Image feature extraction

(1) Image feature classification

Image feature extraction is a concept in the field of computer vision and image processing, which mainly refers to the process of obtaining certain visual characteristics in an image through a feature extraction algorithm [6]. There is also a process of feature extraction in the human visual system: when people see different things through the eyes, the human visual system automatically extracts different feature information in the visual field, and the visual system sends these feature information to the brain. The brain understands these characteristics, and thus distinguishes different things and scenes. Similarly, by implementing some feature extraction algorithms on a computer to extract the visual features in an image, tasks such as image recognition and matching can be realized according to different visual features [7, 8]. Because of the complexity of natural conditions, the following problems will be encountered when extracting image features: images belonging to the same category will have large differences due to changes in lighting, deformation, and angle; and images between different categories will appear more A similar situation [9]. Therefore, it is usually hoped that the selected features can meet the condition that the feature vectors of the images belonging to the same category have similar expressions; and the features of the images of different categories are as different as possible, and the feature should also have good external interference robustness [10].

(2) Global feature extraction

According to the different points of interest in image features, image feature extraction can be divided into two categories: global feature extraction and local feature extraction. One of the global features of an image is the image color feature [11, 12]. Generally speaking, color characteristics are statistics on the color information of each pixel in a certain area. One advantage of color features is that they are very robust to changes in image rotation and scale. However, when the database to be processed contains many images and images of different categories have similar color features, only color features are extracted for image retrieval or classification, which can easily cause misdetection. The color histogram is usually used to extract the color information of the image. The color histogram represents the distribution of color information in the image in the form of histogram [13]. Its advantage is that it is very robust to the rotation and translation of the image. The normalization operation can also reduce the impact of scale and illumination changes. Therefore, when it is not necessary to consider the spatial information of the image or the relative position of the objects in the image, a color histogram can be used. But its disadvantages are also obvious.The color histogram cannot describe the color space distribution of the image. It is not suitable for scenes that need to accurately describe the local content of the image, and it is not suitable for image matching applications of image occlusion and aliasing [14]. Another commonly used global feature is the texture feature. Unlike color features, texture features are not based on each pixel, and are usually performed in image blocks. In image classification, this region-based feature has higher recognition ability than other features, because even if there is a deviation in a local area of the image, the matching failure will not occur. Texture features are very robust to image rotation, because they are extracted based on image regions, so image rotation does not extract image texture features [15]. But when the scale of the image changes greatly, the texture features extracted at this time will be very different, that is, it is not very robust to the scale of the image [16].

(3) Local feature extraction

In contrast to global features, local features are feature expressions of local areas of the image, reflecting the local characteristics of the image. Local feature excellent noise immunity, there is a certain robustness to illumination variation, image rotation, scale changes, changes in perspective, etc., suitable for object recognition, image matching, retrieval FIELD [17, 18]. Local feature includes local spatial distribution image, in such methods is the most representative point feature, image feature extraction and is currently the most widely used during the matching process. Because information such as color, texture, and shape is easily affected by changes in lighting, color distribution, and rotational affine, in some practical applications, it is not possible to extract feature points that meet the matching conditions, resulting in loss of image information. However, point-based features can maintain certain invariance under various changing conditions. Feature points extracted in practical applications are more stable, and image matching is completed under conditions that meet the matching needs [19, 20]. Local point features can be robust to most image changes, and local feature descriptions can obtain important and unique image information contained in the detection area [21].

Therefore, this local feature points are studied. Local feature extraction occurs in two steps: detecting expression of a feature point and descriptors. A first step of extracting local feature is a local feature points in the image is detected, local feature points in common with a corner and two spots. It refers to a corner of the object image corner or intersection between the lines. Generally it refers to spots surrounding region has a significant difference in color and gradation, such as a bird or a cloud in the sky [22]. By definition, the image spot is an area, it is stronger than the anti-noise performance spots corner point, high stability. The methods of speckle detection mainly include the method using the Gaussian Laplacian operator detection (LOG) and the method using the pixel Hessian matrix and its determinant value (DOH). Among them, the improved LOG operator DOG (Difference of Gaussians) is used in SIFT.) detecting a local feature point operator, the DOH SURF method is used. The second step of local feature extraction is the expression of local features. After extracting a series of local feature points from the original image, the properties of these feature points need to be described mathematically for subsequent feature point matching [23].

2.2 BRISK algorithm principle

Due to the impact of unfavorable factors, scale, rotation, light, noise, image (such as aerial building) shot there is often a loss of edge information, resulting in the structure of the edge of the image that can only extract a small amount of spots feature. In contrast, it is possible to hold the object corner edge feature stability [24].

The corner point is an important local feature of the image that is invariant to rotation. It does not change with the change of lighting. Without loss of image information, extracting corner features can reduce the computational complexity of the data and shorten the processing time [25]. There are two different definitions of a corner point: a corner point is the intersection of two edges; a corner point is a point in the neighborhood with two main directions and different edge directions. A large number of corner detection algorithms have been proposed, the most representative of which are the Harris algorithm and the FAST algorithm. The BRISK algorithm uses the AGAST algorithm, which is an optimization of the FAST algorithm, and its detection principle is roughly the same as FAST. It only transforms the tri-tree into a binary tree in the FAST algorithm, and can dynamically and efficiently allocate a decision tree based on the currently processed image information, which makes the performance and repeatability of the AGAST algorithm better than FAST algorithm. Among the many feature extraction and matching algorithms, although the SIFT algorithm is the most widely used and has certain robustness to many changes such as scale, illumination, rotation, affine, and perspective, the algorithm has a limited number of matching points, complicated calculations, and Disadvantages such as poor matching accuracy [26]. BRISK algorithm is a new image feature extraction and matching algorithm with scale and rotation invariance. Compared with the SIFT algorithm, the BRISK algorithm has a faster operation speed and a smaller memory footprint. However, the comprehensive performance of the BRISK algorithm in actual aerial image application is not as good as the SIFT algorithm. Based on the analysis of the BRISK algorithm and SIFT algorithm, this paper proposes an image feature extraction and matching algorithm based on BRISK corners. The BRISK algorithm includes three stages of feature detection, feature description, and matching. Scale-space feature point detection, feature descriptor construction and feature matching [27].

To have scale invariance, the BRISK algorithm builds an image scale-space pyramid. In the BRISK feature detection frame structure, the scale-space pyramid is composed of 4 inner layers c_i and 4 middle layers d_i (i = 0, 1, 2, 3). Each inner layer image is obtained by 0.5 times the downsampling of the previous inner layer image, where the original image corresponds to the c0 layer. Each intermediate layer di is located between two adjacent inner layers c_i and c_i+1. The first intermediate layer d₀ is obtained by 1.5 times downsampling the original image c₀, and the remaining intermediate layers are obtained by 0.5 times the downsampling of the previous intermediate layer.

Assuming δ represents the scale, then: $δ (c_{i}) = 2^{i}$ (1) $δ (d_{i}) = 1.5 \times 2^{i}$ (2)

After the image scale space is constructed, the AGAST algorithm is used for feature point detection. Since the method is optimized for FAST algorithm detection principle and FAST algorithm is substantially the same, therefore, BRISK the feature detection implemented process:

On each inner layer c_i and each intermediate layer d_i, use the AGAST algorithm with the same threshold value to detect feature points and identify potential areas of feature points.

For these feature points in the potential area, non-maximum suppression is performed on the scale space, that is, the score value of the feature point to be detected in the same layer must be greater than the eight points adjacent to it in the same layer. The score value. And the scores of the other AGAST points in the corresponding areas of the upper and lower layers must also be smaller than this point, and then the stable feature points are selected. The area used is a 3 * 3-pixel area centered on the corresponding feature point.

The maximum value of each detected feature point is fitted using a two-dimensional quadratic function of least squares to correct the sub-pixels. And one-dimensional parabola fitting along the scale coordinates is used for continuous scale correction.

Re-interpolate the image coordinates to obtain the corrected feature point coordinates and scale information.

In the construction of feature descriptors, the BRISK algorithm describes each feature point by using a simple comparison of pixel gray values to obtain a cascaded binary bit string. The neighborhood sampling mode is adopted in the BRISK algorithm, that is, using the feature points as the center, a plurality of discrete Bresenham concentric circles with different radii are constructed, and then N sampling points with the same pitch are obtained on each concentric circle.

Because this neighborhood sampling mode has the effect of image gray aliasing during sampling, the BRISK algorithm first performs Gaussian smoothing on the image. And the Gaussian function standard deviation σi is proportional to the distance between the dots on the concentric circles. Suppose: $C_{N}^{2} = N \times (N - 1) / 2$ (3)

An arbitrary pair of sampling points (pi, pj) is selected from the sampling points, and the smoothed gray values are I (pi, σi) and I (pj, σj), respectively. The local gradient between the two points is: $g (Pi, Pj}) = (Pj - Pi) \cdot \frac{I (Pj, σ j) - I (Pi, σ i)}{∥ Pj - Pit ∥^{2}}$ (4)

Suppose that the set of all sample point pairs is denoted as A, then $A = {(Pi, Pj) \in R^{2} \times R^{2} | i < N \land j < i \land i, j \in N}$ (5)

Then the set S consisting of short-distance sampling point pairs and the set L consisting of long-distance sampling point pairs are: $S = {(Pi, Pj) \in A | ∥ Pj - Pi ∥ < δ max} \subseteq A$ (6) $L = {(Pi, Pj) \in A | ∥ Pj - Pi ∥ > δ min} \subseteq A$ (7)

Among them, the distance threshold is usually set to δ max = 9.75δ, δ min = 13.67δ and δ are the scales of the feature points.

Because the pair of long-distance sampling points contains more feature point angle information and the local gradients cancel each other out, the feature mode direction of the feature points can be calculated in the set L as: $g = (\begin{matrix} gx \\ gy \end{matrix}) = \frac{1}{L} \cdot \sum (Pi, Pj) \in Lg (Pi, Pj)$ (8)

Then rotate the sampling mode around the feature point $α = arctan 2 (gy, gx)$ (9)

Furthermore, the feature descriptor has rotation invariance.

Finally, in the rotated short-distance sampling point set S, the grayscale values of pixels are compared for all feature points $(P_{i}^{α}, P_{j}^{α})$ , and finally a 512-bit binary string descriptor is formed. $b = {\begin{matrix} 1, I (P_{j}^{α}, σ j) I (P_{i}^{α}, σ i) \\ 0, \end{matrix}$ (10)

Since the BRISK algorithm uses the idea of the BRIEF algorithm to construct the feature descriptors, the Hamming distance is used to calculate the degree of matching between the feature descriptors, that is, the corresponding feature descriptors are XORed bit by bit, and then the results are counted The smaller the number of 1, the better the degree of matching. And if the value is less than the set threshold, the match is considered successful.

2.3 Improved BRISK corner image feature extraction algorithm

Based on the research of the BRISK algorithm and SIFT algorithm, this paper proposes improvement strategies, which are embodied in several aspects.

(1) Construction of a new-scale spatial pyramid

The construction of the image pyramid is a key step to achieve the invariance of the image scale. To extract feature points that do not change with scale from multiple scales, this paper proposes a new method of constructing scale space by using the SIFT and BRISK algorithms to build a pyramid. Therefore, the stability of feature point detection of the BRISK algorithm on multiple scales is improved. The specific implementation process of building a new-scale spatial pyramid is:

1) Assuming that the size of the input image is M × N, the SIFT method is used to establish a scale space, and a Gaussian pyramid image with 4 groups and 5 layers in each group is constructed. The different layers in the same group are formed by convolving the image with a Gaussian kernel (σ, kσ, k²σ, k³σ, k⁴σ) of different scales. There is a constant scale factor k = 2^1/-S between the Gaussian kernel function scales of adjacent layers. The first layer images in each group are obtained by 0.5 times downsampling from the last second layer images in the previous group.

2) Take each layer in each group in the Gaussian pyramid as the initial image for constructing the BRISK scale space, construct a subgroup scale-space containing 2 inner layers and 2 intermediate layers, and use the layer image as each subgroup The first inner image.

3) In each subgroup, each intermediate layer is distributed between two adjacent inner layers. Each inner layer image is obtained by 0.5 times the downsampling of the previous inner layer image. The first-intermediate layer is obtained by 1.5 times the downsampling of the first inner layer image, and the other intermediate layers are obtained by 0.5 times the downsampling of the previous intermediate layer.

4) Repeat steps 2and 3 until the images in all groups of the SIFT Gaussian scale-space pyramid are traversed, and finally a new scale-space pyramid is generated.

(2) Scale-invariant feature point detection

In SIFT algorithm, the feature point detection is performed on the image Gaussian difference pyramid, and the calculation is complicated. The BRISK algorithm uses the AGAST algorithm for feature point detection, and the calculation is simple. Therefore, you can use the BRISK algorithm to detect feature points to perform feature point detection on the newly constructed image pyramid.The specific implementation process is:

Use the AGAST algorithm with a template of 5–8 to perform feature point detection on the first inner layer image in each subgroup. That is, the gray value of at least 5 consecutive pixels in the 8 pixels on the circular structure is larger or smaller than the gray value of the center pixel.

Use the AGAST algorithm with the same threshold value of 9–16 on each of the remaining inner and middle layers to detect feature points. That is, at least 9 consecutive pixel gray values of the 16 pixels on the circular structure are larger or smaller than the gray value of the center pixel.

Non-maximum suppression detection is performed on the region where the feature points are located on the scale space. This step can be realized by using the idea of BRISK feature point detection.

Perform sub-pixel correction on the feature points with the maximum value detected in the previous step. On the same scale, the characteristic response values of one inner layer corresponding to each extreme value and two adjacent middle layers are respectively fitted with a two-dimensional quadratic function of the least squares, thereby obtaining three maximum.

The continuous extremes of the corrected extreme points are corrected by fitting a 1D parabola along the scale coordinates to obtain the extreme response estimates and the optimal scale estimates.

Re-interpolate the image coordinates to obtain the corrected feature point coordinates and scale information, and finally achieve accurate positioning of the feature points.

(3) Scale-invariant feature description and matching

After the feature point detection is completed, a feature description is required for each feature point. Among the existing local feature descriptors, the SIFT feature descriptor performs better. Therefore, the SIFT algorithm is also used to construct 128-dimensional feature vector descriptors. Using the scale and position information of each feature point, a suitable local neighborhood is selected to generate the SIFT feature vector. Then establish the Euclidean distance relationship between the feature points, use the two-way matching method combining the nearest-neighbor-neighbor distance ratio method and the vector space cosine similarity method to obtain candidate matching feature points, and finally use the improved RANSAC algorithm to eliminate errors matching points, thereby improving the matching accuracy.

3 Simulation analysis of image feature extraction algorithm in the big data environment

3.1 Data source

To verify the performance of the improved algorithm in this paper, a large number of experimental verifications were performed in the standard test Mikolajczyk image database and aerial image database under different image conditions. Mikolajczyk image database is a set of 8 high-resolution image sequence sets taken under different camera parameter settings. Each group contains 6 images that change gradually, ranging from simple to complex changes in scale scaling, rotation, viewing angle, blur, lighting and compression loss, etc., which can be used to verify the performance of the algorithm. Here, only the “Boat”, “Wall”, “Bikes” and “Leuven” image libraries in the Mikolajczyk image database are selected for comparative analysis. It is a comprehensive evaluation in terms of scale and rotation changes, perspective changes, Gaussian blur changes, and lighting changes. The first image in each group is used as the reference image, and the fourth image is used as the target image. To test the performance of the improved algorithm of this paper in practical applications, in the testing of aerial images, the images obtained by Google Earth and drone aerial photography were used, and a large number of verification experiments were performed under different conditions.

3.2 Image preprocessing

In the matching of Google earth image and Google Earth image, the Google earth image library generated by this article is used. Through a large number of simulation experiments such as scale, rotation, brightness, and noise immunity in this image library, three groups of images were selected for comparative analysis. In the first group of image matching, the original target image was enlarged by 4 times as the target image, and its image size was 643 * 656 pixels. Its image size is 1024 * 665 pixels. In the second set of image matching, the original target image was rotated by 45 degrees, while increasing the brightness and contrast by 30, and the image after adding 5% Gaussian noise was used as the target image, and its image size was 189 * 176 pixels. In the third group of image matching, the original target image was subjected to three-dimensional spatial rotation. The brightness and contrast were increased by 50 at the same time, and the image after adding 15% Gaussian noise was used as the target image. The three sets of image reference images are all original reference images. In the matching application of the actual aerial image and the Google earth image, the image taken by the drone is mainly used as the target image, and the reference image is the image captured by Google earth in the corresponding area. It is also divided into three groups for comparison.

3.3 Performance analysis in Mikolajczyk database

(1) Experimental results of the improved BRISK algorithm in the Mikolajczyk database

The matching results obtained by using the improved BRISK algorithm in the “Boat”, “Wall”, “Bikes”, and “Leuven” image databases in the Mikolajczyk image database are shown in Table 1 and Fig. 1.

Table 1
Experimental results of the improved BRISK algorithm in the Mikolajczyk database

Image library Total match points Matchpoints correctly Correct rate Running time (s)

Boat 125 100 80% 1.68

Wall 102 100 98.04% 1.89

Bikes 121 113 93.39% 2.4

Leuven 110 105 95.45% 2.35

Image library	Total match points	Matchpoints correctly	Correct rate	Running time (s)
Boat	125	100	80%	1.68
Wall	102	100	98.04%	1.89
Bikes	121	113	93.39%	2.4
Leuven	110	105	95.45%	2.35

Fig. 1

Performance of our algorithm in Mikolajczyk database.

According to the experimental results of the algorithm in the Mikolajczyk database, it can be seen that the accuracy of the algorithm in the Boat image library is 80%, and the accuracy in the Wall image database is 98.04%. It is shown that the algorithm in this paper is robust to changes in perspective. (2) Performance comparison of three algorithms in Mikolajczyk database.

To verify the performance of the improved algorithm in this paper, during the experiment, the improved algorithm in this paper was compared with the BRISK algorithm and SIFT algorithm in the matching accuracy rate and running time. The specific situation is shown in Table 2 and Fig. 2.

Table 2

Experimental results of three algorithms in Mikolajczyk database

Algorithm	Boat	Wall	Bikes	Leuven
BRISK	20%	24.51%	31.67%	30.04%
	0.24	0.38	0.78	0.73
SIFT	58.12%	79.68%	82.06%	92.03%
	0.68	0.82	1.21	0.94
Improve algorithm	80%	98.04%	93.39%	95.45%
	1.68	1.89	2.4	2.35

Fig. 2

Performance comparison of three algorithms in Mikolajczyk database.

It can be seen that under scale rotation, viewing angle, Gaussian blur, and illumination changes, the matching accuracy of the improved algorithm in this paper is the highest, and the BRISK algorithm is the lowest. However, the running time of the improved algorithm in this paper is the longest, and the processing time used by the BRISK algorithm is the shortest. It can be seen that the improved algorithm in this paper not only has strong robustness, but also maintains a high accuracy rate.

3.4 Performance analysis of aerial image database

(1) Match between Google earth image and Google earth image

To meet the needs of practical applications, in the course of the experiment, this article is mainly divided into the matching of Google earth image and Google earth image and the matching of actual aerial image and Google image. The performance of the algorithms is compared under different types of image matching. In the matching of Google earth image and Google Earth image, the Google earth image library generated by this article is used. This article mainly takes three groups of images for comparison and analysis. For the three groups of experiments, the matching results obtained by different algorithms are shown in Table 3 and Fig. 3.

Table 3
Performance comparison of different algorithms

Algorithm Image group Correct rate Running time (s)

BRISK 1 0 0.76

2 16.13% 0.28

3 9.46% 0.35

SIFT 1 60.98% 0.94

2 86.49% 0.47

3 69.70% 0.63

Improve algorithm 1 91.11% 2.29

2 97.50% 1.60

3 92.86% 1.69

Algorithm	Image group	Correct rate	Running time (s)
BRISK	1	0	0.76
	2	16.13%	0.28
	3	9.46%	0.35
SIFT	1	60.98%	0.94
	2	86.49%	0.47
	3	69.70%	0.63
Improve algorithm	1	91.11%	2.29
	2	97.50%	1.60
	3	92.86%	1.69

Fig. 3

Comparison of the accuracy of different algorithms in Google earth images.

It can be seen that the BRISK algorithm has the lowest matching accuracy rate among the three sets of experiments, and the improved algorithm in this paper has the highest accuracy rate. In the first group of image matching, the accuracy rate of the improved algorithm in this paper is 91.11% higher than that of the BRISK algorithm and 30.13% higher than that of the SIFT algorithm, which shows that the scale invariance of this improved algorithm is optimal. In the second set of image matching, the correct rate of the improved algorithm in this paper is 81.37% higher than the BRISK algorithm, and 11.01% higher than the SIFI algorithm. In the third group of image matching, the accuracy rate of the improved algorithm in this paper is 83.4% higher than that of the BRISK algorithm and 23.16% higher than that of the SIFI algorithm, indicating that the comprehensive performance of the improved algorithm in this paper is optimal.

The improved algorithm in this paper maintains a matching accuracy rate of more than 91% under the conditions of matching requirements, and has the advantages of high matching accuracy and strong robustness.

(2) Match analysis of actual aerial image and Google earth image

In the matching application of actual aerial image and Google earth image, this article mainly uses the image taken by the drone as the target image, and the reference image is the image captured by Google earth in the corresponding area. To verify the matching effect of the improved algorithm, it was also divided into three groups of experiments for comparison. In three sets of image matching experiments, the matching results obtained by different algorithms are shown in Table 4 and Fig. 4.

Table 4

Experimental results of the three algorithms on actual aerial images and Google earth images

Algorithm	Image group	Correct rate	Running time (s)
BRISK	1	0	1.11
	2	0	1.67
	3	0	0.87
SIFT	1	0	2.5
	2	0	5.33
	3	0	3.13
Improve algorithm	1	85.19%	2.68
	2	55.56%	5.46
	3	65.63%	3.33

Fig. 4

Analysis of matching algorithms between actual aerial images and Google earth images.

It can be seen that among the three sets of image matching, the improved algorithm has the highest matching accuracy, and the accuracy of BRISK and SIFT algorithms are both 0. Both BRISK and SIFT algorithms can’t complete this kind of image matching, and improved algorithms can complete this type of image matching. This shows that the improved algorithm in this paper has the highest matching accuracy. But in the running time, it takes a long time to improve the algorithm.

4 Conclusions

In summary, the improved algorithm in this paper is an effective image feature extraction and matching algorithm, which has the advantages of strong robustness and high matching accuracy. Compared with the BRISK and SIFT algorithms, although the running time used in this paper is longer, the improved algorithm in this paper has the highest matching accuracy and can realize some aerial images that the BRISK and SIFT algorithms cannot match.

Based on BRISK and SIFT algorithms, this paper proposes an image feature extraction and matching algorithm based on BRISK corners. By using the SIFT scale space and BRISK algorithm, a new scale space construction method is proposed. The corner features with constant scale are extracted. Using the improved feature matching method and the algorithm of removing mismatched points, the accurate matching of aerial images is finally achieved.

Image feature extraction method has been widely used due to its high extraction efficiency. Because of the inherent characteristics of image feature extraction methods, each method has its advantages and also its insurmountable defects. Although the improved algorithm in this paper has the advantages of high matching accuracy and strong robustness, there are still problems such as long running time. The improvement of the classical method and the proposed new method are the directions for future efforts. When conditions permit, multiple extraction methods apply the same feature extraction simultaneously.

References

, Jianmin

, Hou

and Zhang

, Exploring The Innovation Modes and Evolution of the Cloud-Based Service Using the Activity Theory based on Big Data, Cluster Computing 21(1) (2018), 907–922.

Xiao

, Cheng

and Zhang

, Detecting Travel Modes Using Rule-Based Classification System and Gaussian Process Classifier, IEEE Access 7 (2019), 116741–116752.

Qin

, Yan

, Ren

, et al., SecSIFT: Secure image SIFT feature extraction in cloud computing, ACM Transactions on Multimedia Computing Communications and Applications 12(4) (2016), 65.

, Zhao

, Li

, et al., Rotation Feature Extraction for Moving Targets Based on Temporal Differencing and Image Edge Detection, IEEE Geoscience\s&∖sremote Sensing Letters 13(10) (2017), 1512–1516.

Lei

, Aiguo

, Dapeng

, et al., Haptic Display of Image-Based on Multi-Feature Extraction, International Journal of Pattern Recognition & Artificial Intelligence 30(08) (2016), 1459–1472.

Chen

and Shang

, Image Feature Extraction Using Kernel Winner-take-all Based on Independent Component Analysis Algorithm, Acta Metrologica Sinica 38(5) (2017), 576–579.

Chen

Hsing-Chung

, Kuo

Shyi-Shiun

, Sun

Shen-Ching

, et al., A Distinguishing Arterial Pulse Waves Approach by Using Image Processing and Feature Extraction Technique, Journal of Medical Systems 40(10) (2016), 215.

Singh

P.K.

, Plithogenic set for multi-variable data analysis, International Journal of Neutrosophic Science, 1(2) (2020), 81–89.

, Gu

, Wang

, et al., Feature Extraction of Welding Seam Image Based on Laser Vision, IEEE Sensors Journal 18(11) (2018), 4715–4724.

10.

Sun

, Chang

, Wang

, et al., Image Detection Method for Tank and Armored Targets Based on Hierarchical Multi-scale Convolution Feature Extraction, Acta Armamentarii 38(9) (2017), 1681–1691.

11.

Zhao

Wenzhi

and Du

Shihong

, Spectral–spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach, IEEE Transactions on Geoscience & Remote Sensing 54(8) (2016), 4544–4554.

12.

Zaied

A.N.H.

, Ismail

and El-Sayed

, A Survey on Meta-heuristic Algorithms for Global Optimization Problems, Journal of Intelligent Systems and Internet of Things 1(1) (2020), 40–47.

13.

Ding

Yong

, Deng

Ruizhe

, Xie

Xin

, et al., No-Reference Stereoscopic Image Quality Assessment Using Convolutional Neural Network for Adaptive Feature Extraction, IEEE Access (99) (2018), 1–1.

14.

Rastogi

Nitesh

and Chaudhary

Deepak

, Enhanced Semantic Image Retrieval using Feature Extraction and KNN Techniques, International Journal of Computer Applications 136(13) (2016), 23–28.

15.

Wang

, Zhang

S.-Q.

, Li

J.-Z.

, et al., Partial discharge feature extraction and optimization based on gray image decomposition, Electric Machines & Control 22(5) (2018), 25–34.

16.

Feng-Ping

, Zhou

Xian-Wei

, BEMD–SIFT feature extraction algorithm for image processing application, Multimedia Tools & Applications 76(11) (2016), 1–20.

17.

Liu

Bing

, Yu

Xuchu

, Zhang

Pengqiang

, et al., Supervised Deep Feature Extraction for Hyperspectral Image Classification, IEEE Transactions on Geoscience & Remote Sensing 99 (2017), 1–13.

18.

Gadicha

and Gadicha

, Implicit Authentication Approach by Generating Strong Password through Visual Key Cryptography, Journal of Cybersecurity and Information Management 1(1) (2020), 5–16.

19.

Longbin

, Xinman

Zhang

, Xuebin

, et al., Discriminative feature extraction of image sequences for lipreading, Journal of Computational & Theoretical Nanoscience 13(5) (2016), 3393–3399.

20.

Zhang

Lanyong

, Cao

and Du

Yixuan

, Study on Image Feature Point Extraction Algorithm and Stereo Matching Technique, Journal of Computational & Theoretical Nanoscience 13(5) (2016), 2938–2944.

21.

Ren

Yuemei

, Liao

Liang

, Maybank

Stephen John

, et al., Hyperspectral Image Spectral-Spatial Feature Extraction via Tensor Principal Component Analysis, IEEE Geoscience & Remote Sensing Letters 14(9) (2017), 1431–1435.

22.

Zou

Hailin

, Liu

Chanjuan

, Shen

Qian

, et al., An improved feature extraction-based color image matching method using the quaternion matrix, International Journal of Communication Networks & Distributed Systems 16(4) (2016), 388.

23.

Xie

Wen

, Jiao

Licheng

, Zhao

Jin

, PolSAR Image Classification via D-KSVD and NSCT-Domain Features Extraction, IEEE Geoscience & Remote Sensing Letters 13(2) (2016), 227–231.

24.

Gururaj

, Jayadevappa

and Tunga

Satish

, Fundus Image Features Extraction for Exudate Mining in Coordination with Content-Based Image Retrieval: A Study, Journal of the Institution of Engineers 99(3) (2018), 313–321.

25.

Wang

Lei

, Gong

Xiaoqing

, Zhang

Yongqin

, et al., Artistic features extraction from Chinese calligraphy works via regional guided filter with reference image, Multimedia Tools & Applications 77(1) (2017), 1–18.

26.

Xian

Yaru

, Xiao

Jun

and Wang

Ying

, A fast registration algorithm of rock point cloud based on spherical projection and feature extraction, Frontiers of Computer Science 13(1) (2019), 170–182.

27.

, You

, Li

, et al., Deep Hierarchical Feature Extraction Algorithm, Pattern Recognition & Artificial Intelligence 30(2) (2017), 127–136.

Image feature extraction algorithm in big data environment

Abstract

Keywords

1 Introduction

2 Design of image feature extraction algorithm in big data environment

2.1 Image feature extraction

2.2 BRISK algorithm principle

3 Simulation analysis of image feature extraction algorithm in the big data environment

3.1 Data source

3.2 Image preprocessing

3.3 Performance analysis in Mikolajczyk database

Table 1 Experimental results of the improved BRISK algorithm in the Mikolajczyk database Image library Total match points Matchpoints correctly Correct rate Running time (s) Boat 125 100 80% 1.68 Wall 102 100 98.04% 1.89 Bikes 121 113 93.39% 2.4 Leuven 110 105 95.45% 2.35

Table 3 Performance comparison of different algorithms Algorithm Image group Correct rate Running time (s) BRISK 1 0 0.76 2 16.13% 0.28 3 9.46% 0.35 SIFT 1 60.98% 0.94 2 86.49% 0.47 3 69.70% 0.63 Improve algorithm 1 91.11% 2.29 2 97.50% 1.60 3 92.86% 1.69

References

Table 1
Experimental results of the improved BRISK algorithm in the Mikolajczyk database

Image library Total match points Matchpoints correctly Correct rate Running time (s)

Boat 125 100 80% 1.68

Wall 102 100 98.04% 1.89

Bikes 121 113 93.39% 2.4

Leuven 110 105 95.45% 2.35

Table 3
Performance comparison of different algorithms

Algorithm Image group Correct rate Running time (s)

BRISK 1 0 0.76

2 16.13% 0.28

3 9.46% 0.35

SIFT 1 60.98% 0.94

2 86.49% 0.47

3 69.70% 0.63

Improve algorithm 1 91.11% 2.29

2 97.50% 1.60

3 92.86% 1.69