Visual retrieval of digital media image features based on active noise control

Abstract

In order to improve the recall and precision of image retrieval, a visual retrieval method of digital media image features based on active noise is proposed. In this paper, Canny algorithm is used to detect the edge of the image to get the feature information of the edge of the image. The RGB color space model is used to decompose the color information of the image, and the color characteristic information of the image is obtained. Extracting image features to retain the useful information contained in the image as much as possible; In order to facilitate the visual retrieval of image features, reduce the retrieval complexity and further fuse image features, FPCA and ReliefF algorithms are used to reduce the dimensionality of image features, and the active noise control method is used to sharpen the image. After processing the results, a digital media image feature visual retrieval platform is established to realize the visual retrieval of digital media image features. Experimental results show that the proposed method has a high accuracy of over 80% and a high recall rate of 95.2%.

Keywords

Active noise control image features visual retrieval canny algorithm wavelet PCA ReliefF algorithm

1. Introduction

With the rapid development of multimedia technology and network technology, digital media images have gradually become the main carrier of information, and have been widely and deeply applied in all walks of life, and various image databases have gradually formed [1]. Today, most digital media image resources are stored and transmitted in compressed form, and most of the compression coding standards are based on DCT. In order to better manage and use image databases, it is of great significance to study image retrieval techniques [2]. At this stage, the traditional image retrieval technology is carried out in the spatial domain, and the general steps can be divided into decoding compressed data, spatial feature extraction, and re-encoding to restore the compressed format [3]. There are two main problems in the implementation of traditional schemes. First, the codec process requires a lot of calculation; second, the whole process involves compression domain and spatial data, which requires a lot of data and occupies a lot of memory space [4, 5]. These two problems greatly reduce the recall rate and precision rate of image retrieval and fail to meet the requirements of modern people for image retrieval.

In view of the problems existing in the existing image retrieval technology, relevant scholars have improved the image retrieval technology. Wang et al. [6] proposed an image retrieval algorithm based on multi-feature fusion. In this method, a fusion formula FSF based on feature similarity was designed to extract global, main content and style features of clothing based on YOLOv3 model. Through training and metric learning training, CNN’s ability to extract clothing style attributes and features was improved. Then, the three-way CNN features were extracted and the FSF formula was used to calculate the multi-scale CNN fusion features. The Euclidean distance between optimized features is predicted, and semantic drift is suppressed. Finally, the deep semantic features extracted by CNN were supplemented, and traditional features were introduced to constrain texture, color and other features of initial retrieval results. Multi-scale CNN fusion features were combined with traditional features by FSF formula to optimize the ordering of initial retrieval results. The experimental results show that the algorithm can effectively optimize the image sorting results and improve the retrieval accuracy, but there is a problem of low precision in image retrieval. Shi et al. [7] proposed an image retrieval method based on joint weighted aggregation depth convolution feature. In this method, the image is input into the pre-trained convolutional neural network, and the last convolution output is extracted as the depth convolution feature of the image; by calculating the spatial weight matrix, the salient areas of the image are highlighted; according to the principle of maximum channel variance, the corresponding The feature map calculates the spatial weight matrix, and the original depthwise convolution features are weighted and aggregated into column vectors; the channel weight vector is calculated by treating the feature maps of different channels differently, and the final global feature vector is obtained. The experimental results on public data sets show that this method can effectively enhance the expression ability of image features, has obvious advantages in the average accuracy of image retrieval, and can be effectively applied to the related fields of image retrieval. However, this method has the problem of low recall and is easy to miss some images. Cui et al. [8] proposed a hash image retrieval method based on bilinear iterative quantization, which uses compact bilinear projection to map high-dimensional data into two smaller projection matrices; Then the iterative quantization method is used to minimize the quantization error and generate an effective hash code. Experiments on cifar-10 and caltech256 data sets show that this method can reduce the impact of high-dimensional data and can widely serve the hash image retrieval application of long coding bits of high-dimensional data. However, the retrieval results of this method are easy to include different images and the retrieval effect is poor.

To solve the above problems, a digital media image feature visual retrieval method based on active noise control is proposed to improve recall and precision of image retrieval and improve image retrieval effect. In this paper, Canny algorithm is firstly used to detect image edges to obtain image edge feature information, and then image features are extracted and fused. FPCA and ReliefF algorithms are used to reduce the dimension of image features, and active noise control method is used to sharpen the image. Finally, a digital media image feature visual retrieval platform is established to realize image visual retrieval. It is hoped that this paper can provide a good reference for modern image retrieval research.

2. Digital media image preprocessing

The digital media image feature visual retrieval method proposed in this paper draws lessons from the processing process of biological vision, and divides the image preprocessing into three steps: the first step is to detect the edge of the image; The second step decomposes the color information of the image to provide more features for the similarity comparison of the image; The third step is to measure the similarity of the processed image.

2.1 Image edge detection

Before image retrieval, first detect the edge of the image to obtain the image edge features, which is realized by Canny algorithm [9]. Canny algorithm uses Gaussian noise to describe the image edge, which is represented by a mathematical model of step change. The problem of edge detection is transformed into the problem of finding the maximum value of gradient function. The calculation process is divided into the following four steps:

(1)
Smooth the image with a Gaussian filter, which aims to remove the influence of noise on edge extraction [10]. The expression of the transfer function of the two-dimensional Gaussian filter is $G\left({x,y}\right)$ , then the processed image is $G^{\prime}\left({x,y}\right)$ , where $x$ and $y$ are the coordinates of the image. Setting $s^{2}$ is the standard deviation of the probability function and the only parameter of the Gaussian filter. It is proportional to the size of the neighborhood that the filter acts on. The farther the pixel is from the center of the filter, the smaller the impact on the filter. If the pixel distance from the center exceeds $3s^{2}$ , then the effect of the pixel on the filter is negligible.
(2)
Calculate the gradient value and direction corresponding to each pixel by the finite difference of the first-order partial derivative.
(3)
Non maximum suppression of gradient amplitude. For each critical point, search along the gradient direction and compare with the gradient values of all pixels in the gradient direction. If the gradient value of this pixel is the maximum value in the gradient direction, this pixel is considered as a boundary candidate point, otherwise it is regarded as a non boundary point.
(4)
Eliminate false boundary points and connecting edges with a double-threshold algorithm. Given two critical values $L_{1}$ and $L_{2}$ , and set $L_{1}>L_{2}$ , for each critical candidate point obtained in step Eq. (3), if its gradient value is greater than $L_{1}$ , it is determined as a boundary point, and if it is less than $L_{2}$ , it is determined as a non-critical point. If the gradient value of the boundary point is between $L_{1}$ and $L_{2}$ , according to the principle that the candidate boundary point is connected with the pixel that has been determined as the boundary point, that is, there is at least one boundary point in its eight adjacent areas, the candidate point is determined as the boundary point, otherwise it is a non boundary point.

According to the above steps, the edge detection result of digital media is obtained and the image edge feature information is obtained.
2.2 Decomposition of image color information

In addition to image edge information, image color information is also a relatively important image detail information. At present, most images are stored in digital image format, and the expression of color information in digital images is realized in the form of color space. The color space is divided into different models, such as RGB model, HSV model, HIS model, CMY model, Lab model, etc. [11]. In this paper, the RGB color space model is mainly used to decompose the color information of the image.

The RGB color space model utilizes the configuration of red, green, and blue color channels, which can express both color image information and grayscale image information. When expressing grayscale image information, the values of the red channel, green channel, and blue channel are equal. Therefore, the RGB representation of grayscale images can also be regarded as a special case of the RGB representation of color images.

The pixel color information in a digital media image, under the RGB color space model, is expressed as:

$\displaystyle C_{s}=R\times 256+G\times 256\times B\times 256$ (1)

In the formula, $C_{s}$ represents the color information of any pixel in the digital media image, $R$ represents the color information of the red pixel, $G$ represents the color information of the red pixel, and $B$ represents the color information of the blue pixel.

Accordingly, if the color information of any pixel in the digital media image is known, it can be decomposed into three component information of red, green and blue, as shown in Eq. (2):

$\displaystyle\left\{{\begin{array}[]{l}R=C_{s}-B\times 256-C\times 256\\ G=C_{s}-B\times 256/256\\ B=C_{s}/256\\ \end{array}}\right.$ (2)

In order to visually show the effect of a color image after decomposing RGB color information, an example situation as shown in Fig. 1 is given.

Figure 1.

Decomposition effect of digital media image color information.

From the results in Fig. 1, it can be seen that after RGB decomposition, the original digital media color image becomes three component images, which can provide more features for similarity comparison. However, turning one image into three images increases the time for feature extraction and similarity comparison.

2.3 Image similarity measure

Based on the decomposition results of the color information of digital media images, the image similarity measure is carried out according to the wavelet basis, and the overall similarity measure is composed of two similarity sub-measures [12]. The first similarity sub-measure is used to compare the approximate wavelet-based feature $\mu_{1}$ of the image to be retrieved and the database image, and the second similarity sub-metric is used to compare the approximate wavelet-based feature $\mu_{2}$ of the image to be retrieved and the database image, respectively as Eqs (3) and (4) shows:

$\displaystyle S\left(\mu_{1}\right)=\textit{sim}\left[{\sum\limits_{i=1}^{N}{% \mu_{1}\left(i\right)-\sum\limits_{j=1}^{N}{\mu_{1}\left(j\right)}}}\right]$ (3)

In the formula, $\mu_{1}\left(i\right)$ represents the $i$ -th approximate wavelet-based feature $\mu_{1}$ extracted from the database image; $\mu_{1}\left(j\right)$ represents the $j$ -th approximate wavelet-based feature $\mu_{1}$ extracted from the image to be retrieved; $S\left({\mu_{1}}\right)$ represents the similarity sub-measure related to the feature $\mu_{1}$ .

$\displaystyle S\left({\mu_{2}}\right)=\textit{sim}\sqrt{\left({\sum\limits_{i=% 1}^{I}{\mu_{2}\left(i\right)-\sum\limits_{j=1}^{I}{\mu_{2}\left(j\right)}}}% \right)^{2}}$ (4)

In the formula, $\mu_{2}\left(i\right)$ represents the $i$ -th approximate wavelet-based feature $\mu_{2}$ extracted from the database image; $\mu_{2}\left(j\right)$ represents the $j$ -th approximate wavelet-based feature $\mu_{2}$ extracted from the image to be retrieved; $S\left({\mu_{2}}\right)$ represents the similarity sub-measure related to feature $\mu_{2}$ .

The two sub-measures are fused together to form an overall similarity measure, and the result is shown in Eq. (5):

$\displaystyle S\left({\mu_{1},\mu_{2}}\right)=\frac{E_{1}S\left({\mu_{1}}% \right)+E_{2}S\left({\mu_{2}}\right)}{R+G+B}$ (5)

In the formula, $E_{1}$ and $E_{2}$ respectively represent the contribution of the two similarity sub-measures $S\left({\mu_{1}}\right)$ and $S\left({\mu_{2}}\right)$ to the overall similarity measure. Here, $E_{1}=E_{2}=0.5$ is set. As the images to be retrieved and the database images are compared one by one, the database image with the greatest similarity is ranked first in the retrieval result, and so on.

3. Visual retrieval of digital media image features

Through the preprocessing steps in the second section, the image edge features and color information are obtained, and the image similarity measurement results are obtained. Based on the digital media image preprocessing results, the digital media image features are further visually retrieved.

3.1 Image feature extraction based on wavelet PCA

Image feature extraction is the basis and core of image retrieval technology. Broadly speaking, image features include text-based features (such as keywords, annotations, etc.) and visual features (such as color, texture, shape, object surface, etc.). Text-based image feature extraction has been deeply studied in the fields of database systems and information retrieval. On this basis, this paper uses the wavelet PCA method to extract image features [13]. The wavelet decomposition is to act on each pixel. Therefore, in the spatial domain, wavelet-based dimensionality reduction cannot distinguish different classes between adjacent pixels, while PCA can provide more local spatial information of different classes between adjacent pixels. Therefore, a new feature extraction algorithm combining wavelet decomposition and PCA dimensionality reduction is proposed, which is helpful for better image retrieval.

Assuming that the dimension of the original space $D$ of the digital media image is $m\times n_{1}\times n_{2}$ , in order to perform PCA transformation, it is necessary to convert $D$ into an observation sample set, $D=\left\{{d_{1},d_{2},\ldots,d_{N}}\right\}$ , where $N=n_{1}\times n_{2}$ , represents the total number of pixels.

Calculate the covariance matrix of the original space $D$ as:

$\displaystyle P=\sum\limits_{i=1}^{N}{\frac{r_{i}+u_{i}}{C_{f}}}$ (6)

In the formula, $C_{f}$ represents the mean vector of all observed sample sets in the original space; $r_{i}$ represents the covariance estimation; $u_{i}$ represents the correlation coefficient matrix.

The eigenvalue $p_{t}$ and eigenvector $p_{u}$ of covariance matrix $P$ are obtained by solving the characteristic formula:

$\displaystyle\eta\left({p_{t}-p_{u}}\right)=0$ (7)

In the formula, $\eta$ represents the identity matrix.

Arrange the eigenvalues in descending order, that is, $\eta_{1}>\eta_{2}>\ldots>\eta_{m}$ , and its corresponding eigenvector is $p_{u1},p_{u2},\ldots,p_{um}$ , and the PCA transformation matrix can be obtained:

$\displaystyle D=\left\{{P_{u1},P_{u2},\ldots,P_{uM}}\right\}$ (8)

The final PCA transform can be expressed as:

$\displaystyle S=\frac{D^{T}Z}{\partial_{i}}$ (9)

In the formula, $D^{T}$ represents the original spatial data of the transformed digital media image; $Z$ represents the residual; $\partial_{i}$ represents the transform domain.

After PCA transformation, the first few principal components contain most of the information, but they are a linear combination of individual information, while the latter principal components contain a small amount of information, and the last few are basically noise. After the traditional wavelet decomposition, the high frequency part has both noise and detailed information in the image. If all high frequency components are discarded, a lot of detail information will be lost. The method proposed in this paper is to perform PCA dimensionality reduction on the premise of fully retaining the image details. Its advantage is that it can retain the useful information contained in the image to the greatest extent. The specific implementation steps are shown in Fig. 2.

Figure 2.

Dimensionality reduction of digital media image PCA.

3.2 Digital media image feature fusion

According to the digital media image feature extraction results obtained in Section 3.1, in order to facilitate visual retrieval of image features and reduce retrieval complexity, image features are further fused. Assuming that $g$ color features, $k$ texture features, and $h$ shape features are extracted from the original image, these features are mapped into feature vectors and can be expressed by Eq. (10):

$\displaystyle\left\{{\begin{array}[]{l}P_{U_{1}},P_{U_{2}},\ldots,P_{U_{g}}=Q_% {U}\left({u_{1},u_{2},\ldots,u_{g}}\right)\\ P_{X_{1}},P_{X_{2}},\ldots,P_{X_{k}}=Q_{X}\left({x_{1},x_{2},\ldots,x_{k}}% \right)\\ P_{J_{1}},P_{J_{2}},\ldots,P_{J_{h}}=Q_{J}\left({j_{1},j_{2},\ldots,j_{h}}% \right)\\ \end{array}}\right.$ (10)

In the formula, $P_{U_{g}}$ represents the $g$ -th color feature of the original image; $P_{X_{k}}$ represents the $k$ -th texture feature of the original image; $P_{J_{h}}$ represents the $h$ -th shape feature of the original image; $x_{k}$ represents the feature vector mapped from feature $P_{X_{k}}$ ; $u_{g}$ represents the feature vector mapped from feature $P_{U_{g}}$ ; $j_{h}$ represents the feature vector mapped from feature $P_{J_{h}}$ .

The features mapped by Eq. (10) are fused, and the result is shown in Eq. (3.2):

$\displaystyle\left\{{\begin{array}[]{l}Q_{U}\left({u_{1},u_{2},\ldots,u_{g}}% \right)\\ Q_{X}\left({x_{1},x_{2},\ldots,x_{k}}\right)\\ Q_{J}\left({j_{1},j_{2},\ldots,j_{h}}\right)\\ \end{array}}\right.$ (11) $\displaystyle\to Q\left({q_{1},q_{2},\ldots,q_{g},q_{g+1},\ldots,q_{g+k},q_{g+% k+1},\ldots,q_{g+k+h}}\right)$

In the formula, $Q\left(\right)$ represents the feature vector after fusion; $q_{1}$ represents the form mapped by the feature vector after fusion, and so on.

It should be pointed out that the dimensions of different eigenvectors are different, and the value ranges are also quite different. When they are merged into one eigenvector, they need to be processed according to certain rules. For example, the normalization process can greatly weaken the influence of the difference in the value range on the importance of different feature vectors. In addition, according to the actual situation of the image information, it is also possible to determine which feature vector accounts for the proportion of the fused vector.

3.3 Dimensionality reduction of digital media image features

Since the dimension of digital media images will increase accordingly after feature fusion, the eigenvalues and eigenvectors are usually solved accurately by the matrix method to achieve feature dimension reduction, but the biggest drawback of this method is that when the dimension scale increases, the amount of calculation is huge, which obviously reduces the efficiency of the algorithm. The basic idea of image feature dimensionality reduction method based on FPCA and ReliefF algorithm is: for a given image feature data, the feature is projected by using the linear transformation principle of FPCA algorithm, and the high-dimensional data is mapped to a low-dimensional subspace. This newly formed low Viterbi feature subset can effectively represent the original data information and eliminate the redundant information in the high-dimensional data, it can effectively reduce the dimension of image features.

FPCA algorithm is a multivariate statistical method that can transform multiple features into several comprehensive features with less information loss. The transformed comprehensive features (principal components) are linear combinations of uncorrelated original features, so as to reduce the dimension of digital media image features.

Assuming that a vector $E$ needs to be compressed from $w$ dimension to $w^{\prime}$ dimension, $w^{\prime}\leqslant w$ , a relatively simple implementation method is to remove the $w-w^{\prime}$ dimension data in the vector $E$ , then the resulting mean square error is the sum of the variances of all components removed. It is required that an invertible linear transformation $F$ must be obtained so that the truncation of $F_{e}$ has the optimal mean square error. If the data to be reduced in dimension is a matrix $E\in F^{n\times m}$ , the process can be expressed as:

$\displaystyle E\left\{{F(j_{n})\leqslant R(j_{n})\left|{F(j_{n})=R(j_{0}),R(j_% {1}),\ldots,R(j_{n})}\right.}\right\}$ (12)

In the formula, $R$ represents the removed part. Express Eq. (12) in matrix form as:

$\displaystyle E=\frac{F^{T}+R}{R^{2}}$ (13)

Then the pivot variable $F$ can be expressed as:

$\displaystyle F=\textit{EZ}$ (14)

It can be proved that when the variance of $F$ is the largest, $z_{i}$ is exactly the eigenvector corresponding to the $i$ -th eigenroot arranged from large to small among all the eigenvalues of the $E$ covariance matrix. For $f_{i}\left({i=1,2,\ldots,n}\right)$ , it is a linear function of $E$ , which can ensure that the change of $E$ can be fully reflected with the minimum correlation between them, so as to achieve the purpose of removing redundant information.

The implementation steps of the digital media image feature dimension reduction algorithm based on FPCA and ReliefF algorithm proposed in this paper are as follows:

(1)

Input the original data set $E\in F^{n\times m}$ , initialize the number of iterations, and set the dimension $w^{\prime}$ to be reduced to;

(2)

Under the condition of ensuring the minimum correlation between the principal components, use the FPCA algorithm to reduce the dimension of $E$ , and obtain a new low-dimensional subspace data set $E^{\prime}\in F^{n\times z}$ ;

(3)

Use the ReliefF algorithm to calculate the classification weight $\lambda_{i}\left({i=1,2,\ldots,n}\right)$ of each feature in the dataset;

(4)

Arrange $\lambda_{i}$ in descending order, remove the features corresponding to $\lambda_{i}\left({\min}\right)$ from the dataset, and obtain a new low-dimensional subspace dataset $E^{\prime\prime}\in F^{n\times q}$ ;

(5)

Judging the dimension of the subspace data, if $q>z$ , go to 3); otherwise, the algorithm stops.

3.4 Digital media image sharpening based on active noise control

First, assume that the source image has $\varphi$ gray levels. $f\left({x,y}\right)$ is the gray value of the input pixel, which together with the gray value of the adjacent pixels constitutes a 3 $\times$ 3 window template, then the output $g\left({x,y}\right)$ of the pixel gray value is defined as follows:

$\displaystyle g\left({x,y}\right)=\frac{\sum\limits_{i=1}^{N}{f\left({x,y}% \right)\times\theta_{i}}}{\left|{f\left({x,y}\right)}\right|}$ (15)

The form of $g\left({x,y}\right)$ is basically the same as that of $f\left({x,y}\right)$ , the only difference is that $g\left({x,y}\right)$ is a nonlinear function. The graph of the function $g\left({x,y}\right)$ is composed of two partial circles passing through the origin of the coordinates, and it is centrosymmetric. The horizontal axis $o_{i}$ is the difference between the center pixel $O\left({x,y}\right)$ and the adjacent pixel $O^{\prime}\left({x,y}\right)$ , that is, $o_{i}=O\left({x,y}\right)-O^{\prime}\left({x,y}\right)$ . The vertical axis $o_{j}$ is the output of the pixel difference through the nonlinear function, namely $O\left[{f\left({x,y}\right),g\left({x,y}\right)}\right]$ . A schematic diagram of the nonlinear function is shown in Fig. 3.

Figure 3.

Schematic diagram of nonlinear function.

In Fig. 3, $A$ represents the angle between the tangent line passing through the center $O$ and the horizontal axis, and $B$ and $-B$ represent the intersection points of the partial circle and the horizontal axis, respectively. When $0\leqslant Y\leqslant B$ , the intersection points of the partial circle and the horizontal axis are $\left({0,0}\right)$ and $\left({B,0}\right)$ respectively, and the radius is $r$ , then the formula of the circle is:

$\displaystyle H\left(Y\right)=\left({\varphi-\vartheta}\right)\left[{\frac{B}{% 2}\tan\left({90^{\circ}-A}\right)\times r}\right]$ (16)

In the same way, when $-B\leqslant Y<0$ , the intersection points of part of the circle and the horizontal axis are $\left({0,0}\right)$ and $\left({-B,0}\right)$ respectively, and the radius is $r^{\prime}$ , then the formula of the circle is:

$\displaystyle H\left(Y\right)=\left({\varphi-\vartheta}\right)\left[{\frac{% \sqrt{\frac{-B}{2}\tan\left({90^{\circ}-A}\right)\times r^{\prime}}}{B^{2}}}\right]$ (17)

It can be seen from Eqs (16) and (17) that the parameter $\varphi$ is mainly used to control the sharpening intensity. As the parameter $\varphi$ increases, the output $g\left({x,y}\right)$ obtained from the input $f\left({x,y}\right)$ through the nonlinear function is larger, and the sharpening enhancement effect is greater. obvious. The parameter $\vartheta$ mainly plays the role of controlling noise, that is, the difference between the adjacent pixel $O^{\prime}\left({x,y}\right)$ and the central pixel $O\left({x,y}\right)$ greater than $\vartheta$ is regarded as noise to be eliminated, and the difference less than $\vartheta$ is regarded as the edge detail for sharpening and enhancement. Specifically, when $0\leqslant f\leqslant B/2$ , the closer $f$ is to $B/2$ , the greater the contribution to the overall output $g\left({x,y}\right)$ ; when $B/2<g\leqslant B$ , the closer $f$ is to the noise limit, the smaller the contribution to the overall output $g\left({x,y}\right)$ ; when $f>B$ , $f$ exceeds the noise limit, which is considered to be noise to be eliminated and has no effect on the overall output $g\left({x,y}\right)$ .

To sum up, the nonlinear function can effectively implement active noise control and reduce the sensitivity to image noise [14]. By controlling the size of the parameters $\varphi$ and $\vartheta$ , the image noise limit can be effectively controlled, the image sharpening and enhancement strength can be improved, and a better image sharpening and enhancement effect can be obtained.

3.5 Visual retrieval of digital media image features

According to the above image processing results, a visual retrieval platform for digital media image features is established [15]. The platform provides multiple retrieval portals, including tree list retrieval portals, image interface retrieval portals, and keyword retrieval portals. The image interface retrieval is realized by an Applet on the client side. The user can change the image center node through the mouse to visually browse the characteristics of digital media images. The visual image interface actually shows the class represented by the current node to the user. When the user selects and clicks a specific class in the tree list, the image display area will redraw the entire graph with the class name as the center.

The graphic display area provides the visual navigation and retrieval function of the ontology [16]. Figure 4 is the architecture diagram of the visual retrieval platform for digital media image features.

Figure 4.

Architecture diagram of visual retrieval platform for digital media image features.

As can be seen from Fig. 4, the bottom of the user interface in the visual retrieval platform for digital media image features is the retrieval result output area. The user selects “detailed information” in the shortcut menu, and both keyword retrieval and SPARQL retrieval will be performed in this area. Output retrieval results, thereby realizing visual retrieval of digital media image features.

4. Experimental verification

In order to verify the effectiveness of the visual retrieval method of digital media image features based on active noise control proposed in this paper, experiments are carried out.

4.1 Experimental environment and parameter settings

This test is carried out in the CloudSim cloud computing environment, the network bandwidth is 80 MI.S-1, the memory is 8 GB, and the operating system is Windows 8. The corel standard library is used as the experimental object to carry out experimental research. The Corel standard library is a standard open source library widely recognized in the field of image retrieval. There are 100 small categories and 10,000 images in the image library. Figure 5 shows some of the experimental images selected from the database.

Figure 5.

Experimental sample image.

According to the above experimental conditions, the visual retrieval results of digital media image features by the method in this paper, the image retrieval algorithm based on multi-feature fusion and the image retrieval method based on joint weighted aggregation depth convolution features are compared. In the experiment, the precision rate, recall rate and retrieval effect are used as comparison indicators to compare the retrieval effect of different methods, and draw the experimental conclusion.

4.2 Experimental results

4.2.1 Precision comparison

Three methods are used for image feature retrieval, and the precision is the experimental index to compare the retrieval effects of the three methods. The results are shown in Fig. 6.

It can be seen from the comparison results in Fig. 6 that the precision of the three methods shows an increasing trend with the increase of the number of iterations. Although the precision of this method is lower than that of the image retrieval method based on the joint weighted aggregation depth convolution feature when the number of iterations is less than 3, the highest precision of this method is more than 80%, the highest precision of image retrieval algorithm based on multi feature fusion and image retrieval method based on joint weighted aggregation depth convolution feature are 56% and 45% respectively, which is significantly lower than that of this method. It shows that in the visual retrieval of data media images, this method can obtain more accurate retrieval results and provide users with more accurate retrieval results.

4.2.2 Recall rates comparison

Three methods are used for image feature retrieval. Taking recall as the experimental index, the retrieval effects of the three methods are compared. The results are shown in Table 1.

Table 1
Recall comparison results

Iterations/ time	The method of this paper	Multi-feature fusion	Joint weighted aggregation depth convolutional features
1	94.3	89.7	78.4
2	95.0	87.6	80.1
3	93.8	86.3	81.9
4	92.1	84.4	83.5
5	94.6	83.0	84.2
6	95.2	86.9	78.8
7	90.6	82.5	90.3

Figure 6.

Precision comparison results.

Figure 7.

Comparison of retrieval effects.

According to the data in Table 1, when using the method in this paper to retrieve data media image features, the highest recall rate is 95.2%; when using the multi-feature fusion-based image precision retrieval algorithm to retrieve data media image features, the highest recall rate is 89.7%; when the image retrieval method based on joint weighted aggregation depth convolution features is used to retrieve data media image features, the highest recall rate is 90.3%. From the above data comparison results, it can be seen that the recall rate of the method in this paper is higher, indicating that it can retrieve more image features and obtain more comprehensive retrieval results.

4.2.3 Image feature retrieval effect

In order to further verify the performance of the method in this paper, three methods are used to retrieve the sample images shown in Fig. 5, and the results are shown in Fig. 7.

In the sample images shown in Fig. 5, the top images are all buildings, the middle images are all plants, and the bottom images are all animals. According to the retrieval results obtained in Fig. 7, the order of the images retrieved by the method in this paper is consistent with that in Fig. 5, indicating that the retrieval results are more accurate. The multi feature fusion method is mixed with some different types of images in image retrieval, and the retrieval results are biased. The retrieval results of the joint weighted aggregation depth convolution feature method are similar to those of the multi feature fusion method, and there is also the problem of confusion between different types of images. According to the above experimental results, this method has a very prominent retrieval effect.

5. Conclusion

In order to improve the recall rate and precision rate of image retrieval and improve the effect of image feature retrieval, a visual retrieval method of digital media image features based on active noise control is proposed. The edge feature information of the image is obtained by edge detection, the color feature information of the image is obtained by using the RGB color space model, the image features are fused, the fusion result is obtained, and the dimensionality reduction process is performed on the fusion result. The image is sharpened by the active noise control method. According to the image processing results, a visual retrieval platform of digital media image features is established to realize the visual retrieval of digital media image features. Experimental results show that the proposed method has the highest accuracy of more than 80%, and the highest recall rate is 95.2%. The precision and recall rate of the proposed method are higher than those of the traditional method, and it has better retrieval effect. Although this paper has achieved some research results, its image retrieval accuracy is still low. This is because when extracting image features, this paper only considers its color and other features, without fully analyzing its content. Further research will be carried out in the future.

References

Nawaz

Mehmood

Bilal

Munshi

Saba

. Single and multiple regions duplication detections in digital images with applications in image forensic. J Intell Fuzzy Syst. 2021; 40(6): 10351-10371.

Al-Ameen

. Contrast enhancement of digital images using an improved type-II fuzzy set-based algorithm. Traitement du Signal. 2021; 38(1): 39-50.

Kishore

Rao

. A multi-class SVM based content based image retrieval system using hybrid optimization techniques. Traitement du Signal. 2020; 37(2): 217-226.

Dubey

Roy

Chakraborty

Mukherjee

Chaudhuri

. Local bit-plane decoded convolutional neural network features for biomedical image retrieval. Neural Comput Appl. 2020; 32(12): 1-13.

Andres

Eduardo

Carlos

Fidel

. Image retrieval system based on a binary auto-encoder and a convolutional neural network. IEEE Latin Am Trans. 2020; 18(11): 1925-1932.

Wang

Zhao

Qian

. Accurate retrieval of multi-scale clothing images based on multi-feature fusion. Chin J Comput. 2020; 43(4): 740-754.

Shi

Wang

. Joint weighting aggregation of deep convolutional features for image retrieval. J Xi’an Jiaotong Univ. 2019; 53(2): 128-135.

Cui

Shao

. Hashing image retrieval based on bilinear iterative quantization. Appl Res Comput. 2020; 37(8): 2284-2287.

Wang

Zhang

. Edge detection of agricultural products based on morphologically improved canny algorithm. Math Problems Eng. 2021; (3): 1-10.

10.

Ajith

Latha

. Neutrosophic data formation using Gaussian filter based Costas coding for wireless communication systems. Cognit Syst Res. 2019; 57(10): 115-125.

11.

Nowak

Koscielniak

. What color is your method? Adaptation of the RGB additive color model to analytical method evaluation. Anal Chem. 2019; 91(16): 10343-10352.

12.

Elakkiya

Vanitha

. Interactive real time fuzzy class level gesture similarity measure based sign language recognition using artificial neural networks. J Intell Fuzzy Syst. 2019; 37(5): 6855-6864.

13.

Rezamand

Kordestani

Carriveau

Ting

Saif

. A new hybrid fault detection method for wind turbine blades using recursive PCA and wavelet-based PDF. IEEE Sens J. 2020; 20(4): 2023-2033.

14.

Giouvanakis

Sevastiadis

Papanikolaou

. Measurement of compound sound sources with adaptive spatial radiation for low-frequency active noise control applications. Arch Acoust. 2021; 46(2): 205-212.

15.

Pinilla-Buitrago

Carrasco-Ochoa

Martínez-Trinidad

. Bag of k-nearest visual words for hieroglyph retrieval. J Intell Fuzzy Syst. 2019; 36(5): 4981-4990.

16.

Zhang

Yang

Shao

Yue

. Color image retrieval simulation via generalized local binary pattern. Comput Simul. 2020; 37(6): 457-461.

Visual retrieval of digital media image features based on active noise control

Abstract

Keywords

1. Introduction

2. Digital media image preprocessing

2.1 Image edge detection

3.1 Image feature extraction based on wavelet PCA

4.1 Experimental environment and parameter settings

4.2.1 Precision comparison

4.2.2 Recall rates comparison

Table 1 Recall comparison results

5. Conclusion

References

Table 1
Recall comparison results