CN104715476B - A kind of well-marked target detection method based on histogram power function fitting - Google Patents

Abstract

A kind of well-marked target detection method based on histogram power function fitting, including the classification of histogram power function fitting, super-pixel, marking area problem and well-marked target detect four steps.Beneficial effect of the present invention：Using FT notable figures, figure manifold ranking method, SLIC methods super-pixel classify, histogram power function fitting ask gray threshold image multiple target and scene complexity in the case of, detection efficiency is high, performance is good, high precision, solve a great problem of well-marked target detection field, and method provided by the present invention is performed, and speed is fast, algorithm complex is low, while can guarantee that accuracy of detection higher.

Description

A Salient Object Detection Method Based on Histogram Power Function Fitting

technical field

The invention relates to the field of image salient target detection, in particular to a salient target detection method based on histogram power function fitting.

Background technique

As we all know, the rapid development of computer performance and functions provides reliable and feasible conditions for machine intelligence. With the deepening of disciplines such as machine learning and pattern recognition, people increasingly hope that computers can complete tasks more autonomously and intelligently. To achieve this goal, the computer needs to be able to understand the surrounding environment. The main way for humans to perceive external information is through vision, so the key to a computer's understanding of the surrounding environment is the ability to process visual perception.

The salient target is the target that people pay the most attention to in the image, and generally contains more interesting and useful information. Therefore, salient object detection is widely used in object recognition, image segmentation, image retrieval and other fields. Commonly used salient object detection technologies mainly include salient region detection technologies based on local contrast, such as: based on local contrast and fuzzy growth technology, multi-scale center-surrounding histogram and color space distribution comparison technology; and salient region detection based on global contrast Technologies, such as: Achanta proposes a salient region detection method based on global contrast (Frequency-tuned salient region detection, referred to as FT method) from the perspective of the frequency domain. This method will pass through each pixel in the Gaussian low-pass filter image The Euclidean distance between the value and the average pixel value of the entire image is taken as the saliency value of the point. But it will fail in the following two cases:

(1) The color of the salient area accounts for most of the image. After calculation by this method, the background will have a higher salient value;

(2) The background contains a small amount of prominent color, so the saliency value of this part of the color in the background will be very high.

Related literature: ACHANTA R, HEMAMI S, ESTRADA F, et al. Frequency-tunedsalient region detection[C] // IEEE Conference on Computer Vision and Pattern Recognition, 2009: 1597–1604.

In addition, although the performance of many salient object detection models in single salient object and simple background scenes is close to the standard of the test set, they cannot achieve better performance in multi-object and complex backgrounds, especially in the background of target fusion. which performed.

Graph Based Manifold Ranking (Graph Based Manifold Ranking) is a clustering method that has appeared recently. It obtains Laplacian regularization or non-regularization matrix by calculating the adjacency matrix and degree matrix of the graph. Different variants can be applied in different environments. . Chuan Yang et al. applied graph manifold sorting to salient target detection, segmented the image by SLIC, and used the segmented superpixels as graph nodes, and used the image edge nodes as the seeds of correlation query to detect the background, and then obtained the contrast Significant area. The detection effect of this method is better in the case of single target and simple background, but the detection effect is not ideal when the salient target is located at the edge of the image, multi-target scene, complex background or foreground and background blending.

Related literature: Chuan Yang, Lihe Zhang, Huchuan Lu, Minghsuan Yang, Saliency Detection via Graph-Based Manifold Ranking, CVPR2013, P3166-3173.

Contents of the invention

The technical problem to be solved by the present invention is to provide a salient target detection method based on histogram power function fitting, by calculating the histogram data of the saliency map obtained by the FT algorithm to find a gray threshold, which can extract the salient target area Superpixels, and these superpixels are used as query seeds for graph manifold sorting, and then all superpixels that may have salient pixels are extracted by adaptive binarization method, as a supplement to query seeds, to achieve a salient graph that is close to the standard of the test set , so as to realize the detection of salient targets that are difficult to detect, such as those located at the edge of the image, multi-target scenes, complex backgrounds, or fusion of foreground and background.

The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a method for detecting salient objects based on histogram power function fitting, characterized in that: the method for detecting objects includes the following steps:

Step 1: Histogram power function fitting: use the FT algorithm to generate the FT saliency map of the original image and calculate the gray histogram data of the saliency map, and use the least square method to fit the power function curve equation according to the gray histogram data to obtain Get the grayscale threshold x ₀ for superpixel classification in the FT saliency map;

Step 2: Superpixel classification: The original image is divided into n superpixels by the SLIC algorithm, and the superpixels are divided into significant superpixels and background superpixels according to the gray threshold obtained in step 1;

Step 3: Salient region localization: find superpixels with salient pixels;

Step 4: Salient object detection: Calculate the superpixel correlation matrix and the correlation ranking value of each superpixel by graph manifold sorting method, and obtain each superpixel by normalizing the correlation ranking value of each superpixel The saliency of each superpixel is assigned to all the pixels it contains to generate the final saliency map.

In step one of the present invention, the method of fitting the power function curve equation with the least squares method to ask the gray threshold value is: according to the gray histogram data of the FT saliency map, use the least squares method to fit the power function curve equation; the obtained power Derivation of the function curve equation, the point with a derivative of -1 (x ₀ , y ₀ ) is used as the inflection point between the background gray level and the salient gray level; x ₀ is used as the gray level threshold for separating the background and the salient target in the FT saliency map;

The method for dividing superpixels into significant superpixels and background superpixels in step 2 of the present invention is as follows:

1. Calculate the average gray level mean_gray(i) of all pixels belonging to the same superpixel i in the FT saliency map;

2. Generate an indicator vector Y1=[y1,y2,...yn]T for all superpixels by number, and classify all superpixels whose average grayscale mean_gray(i) is greater than the grayscale threshold as significant superpixels, and their value yi (i=1,2,…,n) is set to 1, otherwise it is classified as a background superpixel, and its value yi (i=1,2,…,n) is set to 0;

The method for locating the salient area in step 3 of the present invention is:

1. Binarize the FT saliency map: use the adaptive binarization method, set the gray level of pixels whose gray level is higher than 2 times the average gray level of the FT saliency map to 255, and set the gray level of pixels lower than this gray level degree is set to 0;

2. Generate an indicator vector Y ₂ =[y ₁ ,y ₂ ,…y _n ] ^T for all superpixels by number, and count the superpixel number of the pixel whose gray level is 255 in the FT significant binary image, and put the significant The indicator value yi of the superpixel i of the pixel is set to 1, and the rest are set to 0.

The calculation method of the superpixel correlation matrix in the step 4 of the present invention is: the superpixels after the original image are divided into a graph G=(V, E), where V represents all superpixel sets of the graph G, and E represents all nodes The fully connected edge set of the graph manifold is used to calculate the superpixel correlation matrix C=(D-αW) ^-1 , where D is the degree matrix of graph G, W is the adjacency matrix of superpixels, and α is the correlation coefficient .

The calculation method of the correlation ranking value of each superpixel in step 4 of the present invention is:

1. Generate an indicator vector Y=[y ₁ ,y ₂ ,…y _n ] ^T for all superpixels by number, let Y= Y ₁ .| Y ₂ , that is, Y takes Y ₁ and Y ₂ bitwise OR operation value;

2. Obtain the correlation ranking value of each superpixel according to the formula f*=(D-αW) ^-1 Y.

In the second step of the present invention, the number n of superpixels in which the original image is divided by the SLIC algorithm is 180-230.

The beneficial effects of the present invention are: (1) use the FT saliency map to obtain the salient superpixels by using the histogram data power function fitting method to calculate the target and background segmentation thresholds, and use the FT saliency binary map to locate all areas where the salient objects may exist, Improve the accuracy of salient target detection; (2) Use the graph manifold sorting method for salient target detection, which can quickly realize salient target detection and binarization of salient areas in simple scenes with single targets, simple backgrounds, multi-targets, and complex scenes Graph segmentation further ensures the accuracy of target detection and makes up for the possible failure of FT algorithm; (3) FT saliency graph combined with graph manifold sorting detection is faster than most salient target detection methods and has lower algorithm complexity. At the same time, high detection accuracy can be guaranteed; (4) The superpixel classification involved in the present invention adopts the pixel clustering technology with better cutting-edge performance—SLIC method, and the clustered superpixels can effectively preserve the significant target edges, ensuring The finally generated saliency map can show the outline of the target more clearly; (5) the saliency region location involved in the present invention adopts the method of combining the binary image obtained by the FT saliency map with the adaptive binarization method and the superpixel information, which can maximize the Maximally ensure that the area where the salient pixels are located is located, effectively improving the detection accuracy; (6) The salient target detection involved in the present invention uses a graph manifold sorting method similar to the spectral clustering method, and the non-regular laplacian matrix can Effectively start from the query seed to search for the nodes related to itself in the graph nodes, and sort them according to the size of the correlation, so that the generation speed of the salient graph is fast and the precision is high; (7) the histogram power function involved in the present invention is simulated The combined gray threshold method is an error theory based on the least square method, which belongs to the category of numerical analysis, not a simple digital image processing technology. (8) The present invention has high detection efficiency, good performance, and high precision in the case of multi-target images and complex scenes, and solves a major problem in the field of salient target detection.

Description of drawings

Fig. 1 is the general flowchart of the present invention;

Fig. 2 is a power function fitting of the present invention to ask the grayscale threshold flow chart;

Fig. 3 is a flow chart of superpixel classification in the present invention;

Fig. 4 is a flow chart of prominent area positioning in the present invention;

Fig. 5 is a flow chart of salient target detection based on graph manifold sorting in the present invention;

Fig. 6 is a flow chart of binarization of saliency map in the present invention;

Fig. 7 is an example diagram of the basic flow of the present invention.

detailed description

A salient object detection method based on histogram power function fitting described in the present invention comprises steps such as histogram power function fitting, superpixel classification, salient region positioning, and salient object detection.

In order to illustrate the specific implementation of the salient target detection method based on the histogram power function fitting of the present invention, it is described as follows in conjunction with the embodiments and accompanying drawings:

Fig. 1 is the general flowchart of the present invention, and it realizes a complete process of the salient target detection method based on histogram power function fitting through the following steps, comprising:

Step 1: Use the FT algorithm to generate the FT saliency map of the original image, and the FT saliency map data FTimname according to the formula Calculate the histogram data hist(i) after normalization, where i=0,1,2,...,255, hist(i) indicates the number of pixels whose grayscale is i;

Step 2: In order to eliminate the adverse effects of low grayscale and high frequency on the calculation of the inflection point, before fitting, all the pixel number data hist(i) of the FT saliency map histogram are divided by 100 (0≤i≤255), Then use the least squares method to fit the power function curve y=Ax ^B with the data set {(i,hist(i))|0≤i≤255} to obtain the coefficients A and B;

Step 3: Deriving the power function y=Ax ^B to get dy/dx=A*Bx ^B-1 , let dy/dx=-1 solve to get x ₀ , and use x ₀ as the gray threshold for superpixel classification;

Step 4: Use the SLIC algorithm to divide the original image into n superpixels (the value of n is 180-230) superpixels, and obtain the attribution information matrix superpixels of the original image pixel belonging to the superpixel, superpixels(i,j) means that the coordinates are (i , j) the superpixel number to which the pixel belongs;

Step 5: Define a one-dimensional _vector ^Y ₁ = [y ₁ , y ₂ , . The average mean_gray, and compared with x ₀ , the superpixel i whose average value is greater than x ₀ is regarded as a significant superpixel, the value of y _i is set to 1, otherwise it is set to 0;

Step 6: Binarize the FT saliency map, set the gray level of pixels whose gray level is greater than 2 times the average gray level of the FT saliency map to 255, and set the rest to 0. Define a one-dimensional _vector ^Y ₂ = [y ₁ , y ₂ , . The value y _i is set to 1, and the others are set to 0;

Step 7: Treat superpixels as a node composition graph G=(V, E), where V represents the node set of graph G, that is, all superpixels, and E represents the fully connected edge set of all nodes. First seek the adjacency matrix W of the superpixel, each element w of W, ci and cj represent the color mean value of the superpixel i and j in the CIELAB color space, σ is a weight coefficient, which can be a constant between 0 and 1, the present invention Take 0.1, according to the formula Calculate the adjacency weight of each node to obtain the adjacency matrix of graph G, according to the formula Obtain the degree matrix D = _diag {d ₁₁ , . ) ^-1 to obtain the correlation matrix of graph G;

Step 8: Combine the salient indicator vector and the positioning indicator vector calculated in step 5 and step 6 by bitwise OR operation Y=Y ₁ .|Y ₂ , and obtain the indicator vector Y as a query, according to the formula Find the saliency of each superpixel, according to the formula Normalize the saliency, assign the normalized saliency to all pixels contained in the superpixel, and generate the final saliency map;

Step Nine: Follow the formula The salient map is binarized to obtain a binarized segmentation map of the salient region.

The method of least squares power function fitting described in step 2 is as follows:

Known histogram data points (i, hist(i)), where i=1, 2,..., m distribution is roughly a power function curve. As a power function curve fitting two parameters , for the convenience of solving, the Taking the logarithm on both sides, let ;

Instead of passing through all the data points, the curve makes the sum of squared deviations is the minimum, where the deviation of each set of data from the fitted curve is ;

According to the principle of least squares, A and B should be taken so that There is a minimum value, so A and B should meet the following conditions:

That is, the following normal equations are obtained

Solve this system of equations to get A and B, and substitute That is, the fitted power function curve is obtained.

Claims (4)

1. A salient target detection method based on histogram power function fitting, is characterized in that: described target detection method comprises the following steps: Step 1: Histogram power function fitting: use the FT algorithm to generate the FT saliency map of the original image and calculate the gray histogram data of the saliency map, and use the least square method to fit the power function curve equation according to the gray histogram data to obtain Get the grayscale threshold x ₀ for superpixel classification in the FT saliency map; Among them, the specific method of using the least squares method to fit the power function curve equation to obtain the gray threshold value is: according to the gray histogram data, use the least squares method to fit the power function curve equation, and derive the power function curve equation obtained, and the derivative is The point of -1 (x ₀ , y ₀ ) is used as the inflection point between the background gray level and the salient gray level, and x ₀ is used as the gray level threshold to separate the background and the salient target in the FT saliency map; Step 2: Superpixel classification: The original image is divided into n superpixels using the SLIC algorithm, and the superpixels are divided into salient superpixels and background superpixels according to the gray threshold obtained in step 1. The specific method is as follows: 1. Calculate the average gray level mean_gray(i) of all pixels belonging to the same superpixel i in the FT saliency map; 2. Generate an indicator vector Y ₁ =[y ₁ ,y ₂ ,…y _n ] ^T for all superpixels by number, and classify all superpixels whose average gray level mean_gray(i) is greater than the gray level threshold as significant superpixels , its value y _i (i=1,2,…,n) is set to 1, otherwise it is classified as a background superpixel, and its value y _i (i=1,2,…,n) is set to 0; Step 3: Locate the salient region: find out the superpixels with salient pixels, the specific method is: 1. Binarize the FT saliency map: use the adaptive binarization method to set the gray level of the pixel whose gray level is higher than 2 times the average gray level of the FT saliency map to 255, otherwise set it to 0; 2. Generate an indicator vector Y ₂ =[y ₁ ,y ₂ ,…y _n ] ^T for all superpixels by number, and count the superpixel number of the pixel whose gray level is 255 in the FT significant binary image, and put the significant The indicator value yi of the superpixel i of the pixel is set to 1, otherwise it is set to 0; Step 4: Salient object detection: Calculate the superpixel correlation matrix and the correlation ranking value of each superpixel by graph manifold sorting method, and obtain each superpixel by normalizing the correlation ranking value of each superpixel The saliency of each superpixel is assigned to all the pixels it contains to generate the final saliency map. 2. A kind of salient target detection method based on histogram power function fitting according to claim 1, characterized in that: the calculation method of the superpixel correlation matrix in the step 4 is: the superpixel after the original image is divided Composition graph G=(V,E), where V represents all superpixel sets of graph G, E represents the fully connected edge set of all nodes, and the superpixel correlation matrix C=(D-αW is calculated by graph manifold sorting method ) ^-1 , where D is the degree matrix of graph G, W is the adjacency matrix of superpixels, and α is the correlation coefficient. 3. a kind of salient target detection method based on histogram power function fitting according to claim 1, is characterized in that: the calculation method of the correlation ranking value of each superpixel in the described step 4 is: 1. Generate an indicator vector Y=[y ₁ ,y ₂ ,…y _n ] ^T for all superpixels by number, let Y= Y ₁ .| Y ₂ , that is, Y takes Y ₁ and Y ₂ bitwise OR operation value; 2. Obtain the correlation ranking value of each superpixel according to the formula f*=(D-αW) ^-1 Y. 4. A salient target detection method based on histogram power function fitting according to claim 1, characterized in that: the number n of superpixels in the step 2 where the original image is divided by the SLIC algorithm is 180-230.