Image feature extraction algorithm based on visual information

3.1.1 Visual information processing process

Both biological vision and machine vision face the same processing target, namely, visual information. Its application field is shown in Figure 1. Machine vision is a branch of artificial intelligence that is developing rapidly. In short, machine vision is to use machines instead of human eyes to measure and judge. There are many descriptions of visual information, including brightness, color, and shape [9,10]. However, brightness is essential information. Human eyes can only recognize in the visible area. At the same time, the difference in light intensity also greatly affects the information obtained by the visual system. With the basic information of brightness, the color, shape, and motion state of objects are also essential. For example, it is difficult to find a frog still in the grass. However, if it keeps beating, even if it is wearing protective colors, people can easily find it. For a man walking on a snowy mountain in a white ski suit, his whereabouts are not easy to be found. However, if he changed into a red ski suit, it might immediately attract people’s attention. The research shows that the object appears obvious through the stimulation of various information. Generally, moving objects are more likely to cause the response of the visual system. Color stimulation is a subjective feeling. Existing research shows that cone cells can be divided into three categories in vision. Each type has a strong response to visible light of a specific spectrum, and these three kinds of light are red, green, and blue. Therefore, the basic principle of three primary colors is not difficult to understand [11].

Figure 1

Application field of visual information.

3.1.2 Image segmentation

To better extract the target features in the image, the image should be segmented first, as shown in Figure 2 [12,13]. For the target in the image, its pixel value is generally determined by its gray distribution function. Generally, the area with gray value less than the maximum gray value of the image is taken as the edge, and the area with gray value greater than the maximum gray value is taken as the target. Generally, the target pixel gray distribution function can be described by two parts. One is the mean function, which can be used to reflect the difference between the pixel value and the target gray value. The second is the standard deviation function, which can be used to reflect the difference between pixels. The pixel value is the value given by the computer when the original image is digitized. It represents the average brightness information of a small square in the original, or the average reflection (transmission) density information of the small square. To get better segmentation results, the median filtering algorithm can be used to denoise the image.

Figure 2

Process flow of image segmentation.

3.1.3 Local feature extraction

Local feature is one of the key factors of image segmentation, which is the description of a certain region in the image [14,15]. For the algorithm proposed in this article, more useful information can be extracted after image segmentation because it can avoid introducing additional calculation in region segmentation. In the region of interest, if the pixels in the neighborhood are Gaussian filtered and then these pixels are normalized, the corresponding normalized window becomes smaller. If the pixels in these neighborhoods have similarities with a certain pixel after Gaussian filtering, it can be used to describe the pixel, so that the features to be extracted after image segmentation can be obtained.

3.2.1 Making full use of computer hardware resources

The feature extraction process of visual information is the analysis and processing of visual information, while the computer is responsible for processing data, storing data, and analyzing data. Therefore, in essence, it is a judgment of the utilization of computer resources [16,17]. Computer is a modern electronic computing machine for high-speed computing. It can perform numerical calculation, logical calculation, and memory storage. It is a modern intelligent electronic device that can automatically process massive data at high speed according to the program. There is a large amount of computation in the image feature extraction algorithm. For example, Canny operator in the edge detection operator needs a lot of data operations. For the image feature extraction process of visual information, it is not necessary to consider the amount of data operation, but only the need to consider the analysis and processing of image features to complete the feature extraction. The process of image feature extraction based on visual information does not need to consume a lot of computing resources and storage resources. Therefore, this method has the advantage of making full use of computer resources. In addition, the method is highly fault tolerant and reliable because the computer is used for calculation. If there is an error in this method, it causes problems such as too much computation for recalculation or replacement of this method, which affects the performance of the algorithm. However, a large number of calculations involved in the process of image feature extraction based on visual cues are all completed by computer, so it is highly fault tolerant and reliable. Even if there are errors in the calculation results of this method due to some reasons, it does not cause the performance of the algorithm to decline or fail to work.

3.2.2 Independent of human visual physiological structure and characteristics

Because human visual physiological structure and features are different from machines, in the image feature extraction algorithm based on visual information, it does not depend on human visual physiological structure and features [18,19]. Human perception of the environment is not only a complex process but also an extremely complex and profound process. First, after receiving the information, human senses analyze and process it to form their own understanding of the world. Then, human beings form their own judgment through comprehensive analysis and comparison. In the image feature extraction algorithm based on visual information, it does not depend on human visual physiological structure and features, which means that this method is a kind of imitation of human visual physiological structure and features. When people do not know or observe the environment (such as in the dark environment), the results of the computer image feature extraction algorithm may be wrong or inaccurate. For example, when performing image segmentation, the result of image segmentation is wrong due to ignorance or errors in the observing environment. Therefore, this article uses a dynamic threshold method to determine the accurate threshold of each pixel in the image. The dynamic threshold is calculated according to the visual perception parameters (such as color, brightness, and contrast) of each pixel before the subjects segment the selected image. In the algorithm, these visual perception parameters of each pixel are dynamically updated, and the final test results are obtained.

3.3.1 K-SVD dictionary learning

In this algorithm, in addition to using dictionary learning to obtain sparse coding for the color, depth, and normal vectors of the original block, SIFT features are constructed and sparsely coded according to the gray information of the image. Normal vector is a concept of spatial analytic geometry. The vector represented by a line perpendicular to a plane is the normal vector of the plane. The sparse coding is solved by the K-SVD update of the original pixel block. However, in the final sparsity coding, K-nearest neighbor is needed to solve the sparsity problem of SIFT features. The advantage of this method is that it can set larger sparsity to preserve more detailed features. For the pixel block features that directly use the original pixel, the Orthogonal Matching Pursuit method is used to obtain the sparse representation of each pixel and set it to a low sparsity to maintain the most important features of each pixel block. As shown in formula (1), the dictionary F is learned from the data matrix B. A represents a sparse coding matrix. Dictionary learning can be achieved by minimizing the objective function.

In formula (1), the square term of the G norm is ‖ . ‖ G 2 , and the zero norm is ‖ . ‖ 0 . They can be used to obtain the sparsity, that is, the number of nonzero elements. Each column in data matrix B has a pixel block feature or SIFT feature of a different information channel. In dictionary F = [ f 1 , . . . , f o , . . . , f n ] ∈ T q × n , column vector f o represents the oth word of the dictionary. Sparse matrix A = [ a 1 , . . . , a u , . . . , a m ] ∈ T n × m can be calculated from dictionary F and sample B. Among them, a u is the sparse encoding of the corresponding b u . Let C = [ D , V , F , M ] be different features extracted from the image. Among them, D, V , F, and M are the feature matrix, dictionary and sparse coding corresponding to the SIFT feature, color, depth, and object surface normal vector information image block.

3.3.2 Extraction and representation of block features

3.3.3 Extraction and representation of image level features

Finally, SIFT is combined with other colors, depth, and other information to achieve sparse coding of feature expression. The article uses the pyramid pooling method to extract SIFT-based sparse coding, block feature-based sparse coding and image block features. These three image features can be combined to obtain the characteristics of the object.

For sparse coding φs based on the SIFT feature, its corresponding subregion feature solution is as follows:

In formula (5), z is the number of layers, whose value is {0, 1, 2}, and i corresponds to the subregions of different layers. u represents the dimension of sparse coding, and p represents the characteristic number of t subregion in the z-layer. δ represents the sparse encoding μ D corresponding to SIFT, block feature corresponding to block α , and sparse encoding α s. All final image features are expressed as follows:

Among them, dg μ D ( h ) , dg α ( v ) , dg α ( f ) , and dg α ( mz ) are features based on gray scale and SIFT sparse coding, color and image blocks, depth and image blocks, and object surface normal vector and image block, respectively. The same feature representation is used for dg α s , but it is obtained by combining with sparse coding of block features.

4.3.1 X algorithm feature matching findings

In this article, the X algorithm was utilized to test the matching effect of similar scenes on images with changes in illumination (A), plane rotation (B), scale (C), and Gaussian blur (D). The results are shown in Figure 4.

Figure 4

Feature matching rate and matching speed of X algorithm in different scenarios. (a) Feature matching rate and (b) matching speed.

It can be learned from Figure 4(a) that the total number of feature matching, the number of wrong feature matching, and the matching rate of X algorithm were 280 pairs, 10 pairs, and 96.4%, respectively, under the change of illumination. When the plane rotation changed, they were 80 pairs, 7 pairs, and 91.3%, respectively. When the scale changed, they were 369 pairs, 30 pairs, and 91.9%, respectively. When Gaussian blur changed, they were 184 pairs, 6 pairs, and 93.4%, respectively. Figure 4(b) shows that the matching speed of X algorithm in the four scenarios was 0.45, 0.17, 0.58, and 0.24 s. Figure 4 shows that the matching rate of the X algorithm in the four scenarios has been maintained at more than 90%, and the time was also within 1 s. The matching effect is flawless.

4.3.2 Y algorithm feature matching result

The experiment remains the same, and only the algorithm is changed. The results are shown in Figure 5.

Figure 5

Feature matching rate and matching speed of Y algorithm in different scenarios. (a) Feature matching rate and (b) matching speed.

Figure 5(a) shows that the total number of feature matching, the number of wrong feature matching, and the matching rate of Y algorithm were 235 pairs, 41 pairs, and 82.5%, respectively, under the change of illumination. When the plane rotation changed, they were 60 pairs, 13 pairs, and 78.3%, respectively. When the scale changed, they were 246 pairs, 57 pairs, and 76.8%, respectively. When Gaussian blur changed, they were 149 pairs, 33 pairs, and 81.1%, respectively. Figure 5(b) shows that the matching speed of Y algorithm in the four scenarios was 1.56, 1.0, 1.97, and 1.81 s. Figure 5 shows that the matching rate of Y algorithm in the four scenarios has been maintained at about 80%. The time was also within 1–2 s, and the matching effect was far less than that of X algorithm.

To sum up, the X algorithm has fast extraction speed and strong stability against rotation changes and image noise. In the pursuit of matching accuracy, matching speed, and real-time, the algorithm has a strong application value.