CN107169487A - The conspicuousness object detection method positioned based on super-pixel segmentation and depth characteristic - Google Patents

Abstract

The invention proposes a salient target detection method based on superpixel segmentation and depth feature location. It solves the problem that the target segmentation effect of the traditional salient target detection method is not ideal. Its realization includes that the present invention utilizes color similarity linear iterative superpixel segmentation to raise the image processing unit from a single pixel point to a collective similar area; fully consider image features such as color features, direction features, and depth features, and combine human eyes to more Caring about the center and ignoring the characteristics of the surrounding background, the feature similarity of the region where the saliency image is located, and the prior knowledge of the uniqueness compared to the global feature, generate a localization saliency map and a depth saliency map of the input image, and fuse and boundary them deal with. The present invention has clearer edges in detection image effects, more complete background elimination, and more complete target form segmentation. It is used in face recognition, vehicle detection, moving target detection and tracking, military missile detection, hospital pathological detection and other fields.

Description

Salient Object Detection Method Based on Superpixel Segmentation and Depth Feature Location

technical field

The invention belongs to the technical field of image detection, and mainly relates to a salient target detection method, in particular to a salient target detection method based on superpixel segmentation and depth feature positioning. It is used in face recognition, vehicle detection, moving target detection and tracking, military missile detection, hospital pathological detection and other fields.

Background technique

As the amount of data continues to grow, the amount of data accumulated per unit of time increases exponentially. The huge amount of data requires better computer technology and algorithm theory to process and refine data information. With the emergence of high-resolution images, it brings great visual enjoyment. People's understanding of complex images has reached a very high level. Traditional image processing separates pixels, or completely analyzes the meaning of information transmitted by images. Faced with a huge amount of data, traditional image processing methods are far from meeting the requirements of high efficiency and real-time. At the same time, only by considering the relevant features of the human eye attention mechanism, such as simple features such as color features and direction features, can no longer meet the desired effect of extracting salient target detection. Or manually process the image to be detected, the work is difficult, stressful, and heavy. How to let the computer simulate the visual mechanism of the human eye and realize the saliency attention mechanism similar to human beings to process image information has become a hot topic that needs to be solved urgently.

Some of the existing salient target detection methods only consider the characteristics of the image itself to find the difference between the target area and the background area of the image, so as to distinguish the target position and the background area. In addition, the Markov chain is used to process the saliency map, and the mutual influence relationship between the central saliency area and the surrounding background area is found. There is also a method of using the convolution of the amplitude spectrum and the filter to realize the redundant information and finally find the salient region. Furthermore, there are various methods focusing on local contrast and global contrast. Although these methods are effective in detecting salient targets, the detection effect is not satisfactory in terms of edge segmentation, background removal, and target shape extraction, and has certain limitations. Moreover, most of the image features are processed in the form of individual pixels, which is far from satisfying the status quo.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art and provide a salient target detection method based on superpixel segmentation and depth feature location with clearer edges, more complete background removal, and more complete target shape segmentation.

The present invention is a salient target detection method based on superpixel segmentation and depth feature location, characterized in that it includes the following steps:

Step 1: The input image is linearly iteratively clustered and segmented. Input the target image to be detected, first divide it into K regions, find the local gradient minimum point in the neighborhood of each region as the center point, and set a label number for the same region; find the five-dimensional Euclidean distance in the neighborhood of the pixel point The smallest center point, and assign the center point label to the pixel to be processed; continuously iterate the process of finding the center point with the smallest distance from the pixel point, stop the iteration when the label value of the pixel point does not change, and complete the superpixel segmentation;

Step 2: Construct difference of Gaussian to generate localization saliency map.

2a: Carry out Gaussian function filtering processing according to the input original image, and generate 8 layers of scale images of the original image;

2b: Combine the constructed 8-layer scale map with the original image to form a nine-layer scale map, extract the red-green color difference map and the blue-yellow color difference map of the nine-layer scale image, a total of 18 color difference maps; Extract the intensity map of the nine-layer scale map, a total of 9 intensity maps; extract the Gabor filter direction map of the nine-layer scale map, a total of 36 direction maps, forming three types of feature maps;

2c: Because the sizes of similar features of the nine-layer scale map are different, the three types of feature maps are first interpolated and then differentially processed;

2d: Different types of feature maps have different metrics for their features, so it is necessary to normalize the different types of features first and then fuse them into a localization saliency map;

Step 3: Generate a deep feature saliency map. First, perform a positioning process on the superpixel-segmented image according to the positioning saliency map in step 2, and then collect the nearest neighbor area information, global area information, and corner background area information for each segmented area and its adjacent areas. Feature information, generating a deep feature saliency map for the detection of salient objects;

Step 4: Combine the localization saliency map and depth feature saliency map finally determined by step 2 and step 3, perform fusion and boundary processing on the localization saliency map and depth feature saliency map, generate the final saliency target map, and complete superpixel segmentation and salient object detection for deep feature localization.

Compared with prior art, the beneficial effect of the present invention is:

1. Most of the existing salient target detection algorithms process image features in units of individual pixels, and the edge separation effect between the detected target area and the complex background is not ideal. The present invention uses five-dimensional Euclidean The linear iteration of the distance and color similarity performs a superpixel segmentation preprocessing on the input image, which solves the problem of unsatisfactory target edge segmentation in the traditional salient target detection method, and provides a more intelligent, efficient, and robust Strong salient object detection method.

2. The method of the present invention fully considers image features such as color features, direction features, and depth features. At the same time, it fully considers that it is more concerned about the center and ignores the surrounding background, and the similarity of features in the area where the target is located. Experimental knowledge; and then realize the detection of salient objects, making the computer more logical and artificial.

3. The method of the present invention draws from the detection results that the detection target is not limited to specific features, conditions such as the environment, by shooting images to be detected in multiple scenes such as office scenes, campus areas, parks, etc., the method of the present invention can detect significant Objects can be detected, and the detection effect is more in line with the salient effect of the human eye. The background removal is more complete, and the position and shape of the target extraction are more complete.

Description of drawings

Fig. 1 is the flowchart of the inventive method;

Fig. 2 is the effect figure after being segmented by superpixel in the method of the present invention, and wherein Fig. 2 (a) is the segmentation effect figure of office wall corner, and Fig. 2 (b) is the segmentation effect figure in the library scene;

Fig. 3 is for selecting ten pictures, the detection effect display and effect comparison diagram of the present invention and other methods in recent years, wherein Fig. 3 (a) is the original image selected, Fig. 3 (b) is the detection effect diagram of the present invention, Figure 3(c) is the rendering of the GS method, Figure 3(d) is the rendering of the GBMR method, Figure 3(e) is the rendering of the RARE method, Figure 3(f) is the rendering of the HSD method, and Figure 3(g) is Effect diagram of STD method, Figure 3(h) is a manual marking diagram;

Fig. 4 is for the selected 500 pictures, the curve results of the precision and recall rate of the present invention and other methods in recent years.

detailed description

The present invention is described in detail below in conjunction with accompanying drawing

Example 1

Some of the existing salient target detection methods only consider the characteristics of the image itself to find the difference between the target area and the background area of the image, so as to distinguish the target position and the background area. In addition, the Markov chain is used to process the saliency map, and the mutual influence relationship between the central saliency area and the surrounding background area is found. There is also a method of using the convolution of the magnitude spectrum and the filter to realize the redundant information and finally find the salient target area. Although these methods are effective in detecting salient targets, the detection effect is not satisfactory in terms of edge segmentation, background removal, and target shape extraction, and has certain limitations.

Aiming at these defects of the prior art, after discussion and innovation, the present invention proposes a salient target detection method based on superpixel segmentation and depth feature location, see Figure 1, including the following steps:

Step (1) Carry out linear iterative clustering segmentation of the input image: input the target image to be detected, that is, the original image, first divide it into K regions, find the local gradient minimum point in the neighborhood of each region as the center point, and Set a tag number for the same area. Find the center point with the smallest five-dimensional Euclidean distance in the neighborhood of the pixel point, and assign the center point label to the pixel to be processed; continuously iterate the process of finding the center point with the smallest distance from the pixel point, and assign a label to the pixel point to ensure that the pixel point The superpixel segmentation is completed until the label numbers of the pixels do not change. In this example, the 5*5 neighborhood is used to find the neighborhood of each area, and the 2S*2S neighborhood is used to find the distance pixel neighborhood.

Step (2) Use the difference of Gaussian method to generate a localization saliency map:

(2a) Perform Gaussian function filtering processing according to the input original image to generate 8 layer scale images of the original image.

(2b) Combine the scale maps of these 8 layers with the original image to form a nine-layer scale map, and extract the red-green color difference map and the blue-yellow color difference map of the nine-layer scale image. A total of 18 images of the color difference map; extract the intensity map of the nine-layer scale map, a total of 9 images of the intensity map of the nine-layer scale map; extract the Gabor filter four direction maps of the nine-layer scale map, these four directions are 0 °, 45°, 90°, 135°, the nine-layer scale map has four kinds of orientation maps, a total of 36 maps, forming three types of feature maps: color difference map, intensity map and direction map.

(2c) Since the obtained nine-layer scale maps have different sizes among similar features, the three types of feature maps need to be interpolated first, and then differentially processed.

(2d) Different types of feature maps have different feature metrics, so a single amplitude cannot reflect the importance of saliency. Therefore, it is necessary to normalize different types of features and then fuse them into a localized saliency map. .

Step (3) Generate the depth feature saliency map of the input image: first perform a positioning process on the superpixel-segmented map according to the localization saliency map in step 2, and then fully consider that the center position is much more likely to be a salient target than the surrounding positions of the image And the concentration of salient objects, that is, the salient objects must be concentrated in a certain area, and it is impossible to scatter in all or most areas of the image; so for each area and its adjacent areas that have been segmented in step 1, the collection The nearest neighbor area information, the global area information, and the corner background area information are three types of feature information to generate a deep feature saliency map for the detection of salient objects.

In step (4), the localization saliency map and depth feature saliency map finally determined through steps 2 and 3, in order to make the object segmentation more regular and make the boundary between the salient target and the neglected background clearer, The location saliency map and the depth feature saliency map are fused and boundary processed to generate the final saliency object map, and the salient object detection of superpixel segmentation and depth feature location is completed.

The method of the present invention fully considers image features such as color features, direction features, and depth features, and at the same time fully considers prior knowledge such as caring more about the center and ignoring the surrounding background, feature similarity of the target area, and uniqueness compared to global features. ; and then realize the detection of salient objects, making the computer more logical and artificial.

Example 2

The salient target detection method based on superpixel segmentation and depth feature location is the same as in embodiment 1, and the superpixel segmentation of the target image to be detected in step 1 of the present invention includes the following steps:

1.1 First assume that the target image, that is, the original image has a total of N pixels, and the total area to be segmented is K. Obviously, each divided area has a total of N/K pixels, and the distance between different areas is about It may happen that the set center point just appears on the edge. In order to avoid this situation, find the position with the minimum local gradient around the set center, and move the center position to the minimum local gradient. And set a label number in the same area as a mark.

1.2 Calculate the Euclidean distance value of the five-dimensional feature vector from each pixel point to the determined center point of the surrounding neighborhood, and then assign the center point label number with the smallest value to the currently processed pixel point. Calculate the Euclidean distance of the five-dimensional feature vector C _i =[l _i ,a _i , _bi , _xi ,y _i ] ^T as shown in the following three formulas, in the five-dimensional feature vector, l _i ,a _i ,b _i respectively represent the brightness of the color in the CIELAB space, the position between red and green, and the three color component information values of the position between yellow and blue, x _i and y _i represent the target image where the pixel is located Coordinate position information value.

In the above formula, d _lab represents the Euclidean distance between pixel point k and center point i in CIELAB color space; d _xy represents the Euclidean distance between pixel point k and center point i in space coordinate position; D _i is the evaluation Whether the pixel point k and the center point i belong to a label or not, the larger its value is, the closer the similarity between the two is, the label will be the same; m is a fixed parameter used to balance the relationship between variables; s is the distance between different regions of about

The above is an iterative cycle for setting the label number to which the pixel belongs.

1.3 Continue to iterate according to the steps in 1.2, and further optimize the accuracy of the label number of the pixel point until the label number of each pixel point in the entire image no longer changes. Usually, about 10 times Iteration can achieve results.

1.4 After the iterative process, some problems may occur. For example, a small area is divided into a super pixel area, and only one pixel is isolated to form a super pixel area. In order to eliminate the occurrence of this situation, the small-sized independent area or isolated single pixel point is assigned to the nearest label number to complete the superpixel segmentation of the target image.

The invention adopts color similarity to perform superpixel segmentation, pre-processes the input image, solves the problem of unsatisfactory target segmentation effect of the traditional salient target detection method, and provides a more intelligent, efficient, robust and more Strong salient object detection method.

Example 3

The salient target detection method based on superpixel segmentation and depth feature location is the same as that of Embodiment 1-2. In step 3 of the present invention, the three types of feature information, namely, nearest neighbor region information, global region information and corner background region information, are collected to fully consider Pay more attention to the center and ignore the surrounding background, the feature similarity of the target area, and the prior knowledge compared to the uniqueness of the global feature, including the following steps:

3.1 Considering that the center is more likely to be salience than the surrounding background, saliency targets must be concentrated in a certain area. For each segmented area, the information in the nearest adjacent area is collected, that is, the nearest neighbor area information.

3.2 Consider the degree of influence of the processing area on the entire image, by removing the information contained in the current area, that is, each segmented area, and other areas, that is, the global area information.

3.3 Each segmented area represents the information of the four corner areas of the background feature, that is, the information of the corner background area.

Acquisition is completed through the feature information provided by these three parts.

Example 4

The salient target detection method based on superpixel segmentation and depth feature location is the same as that in Embodiment 1-3, and the generation of depth feature saliency map described in step 3 specifically includes:

For one of the regions R that has been segmented, quantify the saliency of its depth features as:

Among them, s(R) represents the significance of the region R; Π is the multiplication of multiple factors; s(R,ψ ^C ) represents the information of the nearest neighbor region; s(R,ψ ^G ) represents the global region information; s(R ,ψ ^B ) represents the corner background area information.

Usually when seeing an image like a bunch of flowers on the prairie, people will instantly focus on the flowers and ignore the surrounding background. The probability of appearing in the whole image is relatively low for the target flower with high significance; the high probability causes low attention because of its generality, while the low probability causes low attention because of its uniqueness. High degree of attention; this coincides with Shannon's information theory, low probability means high information content, and high probability means that the amount of information it brings is low. Thus define s(R,ψ ^m ) as follows:

s(R,ψ ^m )=-log(p(R|ψ ^m ))

In the above formula, s(R,ψ ^m ) represents the nearest neighbor area information extracted under the depth feature, the global area information, and the corner background area information; p is a probability value.

For the nearest neighbor area information, the global area information, and the corner background area information, the three types of area information are respectively simplified by their area average values as follows:

In the above formula, d represents the average depth of the currently processed region block R; is the depth average value of ψ ^m mentioned above, where d _i ^m represents the depth average value of the i-th block area of the m-th type area information feature, n _m has three cases, and n _C represents the nearest neighbor area information The total number, n _G represents the total number of global area information, and n _B represents the total number of corner background area information.

Estimated with a Gaussian function Implementation of:

In the above formula, Represents the impact factor of the depth difference of different regional blocks, d, D ^m , The meanings of n and _m have been explained above.

The depth feature saliency model of the present invention fully considers that the human eye is more concerned about the center and ignores the characteristics of the surrounding background, and uses the feature similarity of the area where the target is located. Compared with the uniqueness of the global feature, it takes into account the depth differences of different area blocks Prior knowledge such as the impact factor of the present invention makes the edge of the target detection effect clearer, the background removal is more complete, and the target form segmentation is more complete. It makes the computer more logical and artificial.

Provide comparatively detailed example below, the present invention is further described in detail:

Example 5

The salient target detection method based on superpixel segmentation and depth feature location is the same as in Embodiment 1-4, and its core steps include the following steps:

Step (1) Carry out linear iterative clustering segmentation of the input image: first divide the input image into K regions, find the local gradient minimum point in the neighborhood of each region as the center point, and set a label number for the same region , and different regions are set to different tag numbers, which can also be called tag values in this field. For each pixel point, find the center point with the smallest five-dimensional Euclidean distance from the pixel point neighborhood, and assign the center point label to the pixel point to be processed. Set the current pixel point as the label of the center point with the smallest distance, see Figure 1, iteratively find the center point with the smallest distance from the pixel point, and assign a label to the pixel point. Taking a K area as a unit, by comparing the distance between the pixel point and the center point, set the label of the pixel point, complete the setting of the pixel point label in the K area through iterative optimization, traverse the entire image, and perform judgment operations during the iterative optimization process , specifically to determine whether the label value of the pixel has changed. If the label value has changed relative to the previous iteration process, repeat the iterative operation. Otherwise, if the label value does not change compared to the previous iteration process, the too small size Independent areas or isolated single pixel points are assigned to the nearest label number. In this example, when the number of iterations is 10, the label value will no longer change. The superpixel area formed by the isolated point is removed to complete the superpixel segmentation. After the segmentation is completed, a controllable number of regions is generated, not necessarily K regions. In this example, the neighborhood of each area to be searched is 3*3, and the neighborhood of distance pixels to be searched is 2S*2S.

Step (2) Use the difference of Gaussian method to generate a localization saliency map of the input image:

(2a) Perform Gaussian filter processing on the input image to generate a scale map of 1/2 of the original image, a scale map of 1/4 of the original image, and a scale map of 1/256 of the original image, a total of 8 levels of scale maps.

(2b) Combine the constructed 8-layer scale map with the original map, that is, add the 8-layer scale map to the original map to form a nine-layer scale map. A total of 18 color difference maps are extracted from the red-green color difference map RG and the blue-yellow color difference map BY of the nine-layer scale map. A total of 9 intensity maps are extracted from the intensity map I of the nine-layer scale map. Extract the Gabor filter direction map O of the nine-layer scale map, and extract a total of 36 directions in four directions of 0°, 45°, 90°, and 135° from the nine-layer scale map.

The above process is to extract feature maps from three aspects of the nine-layer scale map, including color difference map, intensity map and direction map.

(2c) Since the sizes of similar features of the obtained three-dimensional feature maps are different, it is necessary to interpolate the similar features of the three types of feature maps first, and then perform differential processing, see Figure 1.

(2d) Different types of feature maps have different measurement standards for their features, so a single amplitude cannot reflect the importance of saliency. First, different types of features are normalized and then fused into a localized saliency map. Get the localization saliency map of the input image.

Step (3) extract the depth feature saliency map D of the input image: first perform a positioning process on the superpixel-segmented map according to the location saliency map in step 2, and then fully consider that the center position is much more likely to be a saliency target than the surrounding image Concentration of location and salient objects, that is, salient objects must be concentrated in a certain area, and cannot be scattered in all or most areas of the image. For each area segmented in step 1 and its adjacent areas, three types of feature information, namely nearest neighbor area information, global area information, and corner background area information, are collected to generate a depth feature saliency map of the input image for the detection of salient objects. .

Step (4) In order to make the object segmentation more regular and make the boundary between the salient target and the neglected background clearer, the positioning saliency map obtained in step 2 and the depth feature saliency map finally obtained in step 3 are used for The location saliency map and the depth feature saliency map are fused and boundary processed to generate the final saliency object map, and the salient object detection of superpixel segmentation and depth feature location is completed.

Through the target detection of the present invention, the target detection effect of clearer edges, more complete background elimination and more complete target shape segmentation can be obtained.

Below in conjunction with accompanying drawing and simulation data technical effect of the present invention is described in detail:

Example 6

The salient target detection method based on superpixel segmentation and depth feature location is the same as that in Embodiments 1-5. In this example, the effect display and result analysis of the superpixel segmentation part of the method of the present invention are performed through simulation.

The simulation conditions are: PC, AMD A8-7650K Radeon R7, 10Compute Cores 4C+6G, 3.3GHz, 4GB memory; MATLAB R2015b.

The simulation content is to use the present invention to perform superpixel segmentation on office corners and library scenes.

Fig. 2 is an effect diagram after superpixel segmentation in the method of the present invention, wherein Fig. 2(a) is a segmentation effect diagram of an office corner, and Fig. 2(b) is a segmentation effect diagram in a library scene.

In Fig. 2, the original image without grid is the original image selected by the present invention, and the grid is the effect diagram after superpixel segmentation by the method of the present invention.

The composition of the image often appears independently of each pixel, but the detected target cannot be a single pixel but occupies a certain area, has multiple pixels, and has certain commonality between pixels rather than one independent. Therefore, in view of these characteristics, the present invention divides image superpixels into same regions with certain commonality and different regions with differences. A large number of single independent pixels are replaced by super-pixel blocks to reduce the complexity of computer processing of images.

See Figure 2(a) which is a scene diagram of a potted plant in the corner of an office. Except for a potted plant in the corner, the rest of the figure is a simple background. It can be seen from the detection effect of the present invention that for other places except potted plants, due to their single characteristics, the present invention first performs superpixel segmentation on the input, and the segmentation results are very regular in both size and shape. When the computer processes this image, because the same kind of segmentation has already been carried out, there is no need to perform detailed processing one by one, thereby reducing the complexity of processing. For the location of the potted plant, because its features are changeable, its green leaves and white pots can also be finely divided according to similar and different features by the method of the present invention. Then, the image is processed in units of the same type and different types of regions, which improves the computer's processing speed of the image. Figure 2(b) is a scene diagram of a library. Eight exhibits are placed in a single background of the library, and one of the exhibits is placed in the middle of the scene. Although the scene in (b) is much more complicated than that in (a), the There is still a single wall area. According to the effect display, it can be seen that for a wall with a single feature, the segmentation is very regular in size and shape. For places where exhibits are placed, the method of the present invention can also clearly segment similar features and different features. In any image to be detected, there will be to some extent the same type of area similar to a single wall. The invention takes the region as the unit when processing by the computer, which effectively reduces the complexity of processing the image by the computer.

Compared with the traditional superpixel segmentation method, the clustering segmentation method realized by the linear iteration of the present invention has regular segmentation in both size and shape, and the segmentation processing effect of boundaries between different areas is clearer.

Example 7

The salient target detection method based on superpixel segmentation and depth feature location is the same as in Embodiments 1-5. In this example, target detection and result analysis are carried out through simulation.

The simulation conditions are the same as those in Embodiment 6.

Simulation content: for the selected ten images, through the present invention and the global contrast significant detection GCS (that is, RARE), geodesic distance significant detection GS, graph theory-based significance detection GBMR, hierarchical significance detection HSD, statistical texture significance The five methods of detecting STD are compared in the same image. The selected images include indoor and outdoor scene images, office scenes, campus areas, parks, etc.

Referring to Fig. 3, Fig. 3 (a) is the selected original image, Fig. 3 (b) is the detection effect diagram of the present invention, Fig. 3 (c) is the GS method effect diagram, Fig. 3 (d) is the GBMR method effect diagram, Fig. 3(e) is the rendering of the RARE method, Figure 3(f) is the rendering of the HSD method, Figure 3(g) is the rendering of the STD method, and Figure 3(h) is the manual labeling graph.

For potted plant pictures in office scenes, the present invention and the GBMR method show advantages, not only can detect the position of the salient target, but also display its basic shape. Although other methods generally show the target shape, but The background removal is not complete, especially for the two methods of HSD and STD, a large part of the background remains. For the basketball detection map in a simple scene, these six types of methods can detect the shape of the target object very well, which is close to the result map of manual marking and meets the result requirements. For a single-target red lantern on the roof, because the roof is also red in this scene, the interference effect of the roof is more obvious. In this scenario, these methods can detect the target object, but it can be seen that the GS method and the STD method The red roof with strong interference has not been completely eliminated. For the red lantern image on the roof with two targets, since the method proposed in this paper involves positioning the saliency map, there are certain limitations for the detection of multiple targets, but it focuses more on the detection of the target on the left, and the best detection of the target in this scene The method is the HSD method, but although the HSD method and other methods can detect the position of the target, they cannot completely remove the roof background as a strong interference, and the remaining background areas are too obvious, which is not strongly disturbed by the present invention. The impact of the roof of the item, the background around the detected single target is relatively clean. If there are multiple similar monuments on the wall, it is obvious that the human eye will only pay attention to the middle part of the scene, which is the most central and occupies the largest area of the scene. For this scene, the HSD method puts the interference on the upper and right sides of the scene. It is obviously unreasonable to detect the small similar area of the square as a salient target. The RARE method and the STD method are also in the same situation, while the results of the GS method and the GBMR method are relatively ideal. For the signboard map in the park scene, the present invention has obvious advantages compared with several other methods, and is closest to the artificially marked map. For these three pictures in the museum, the present invention and the GS method and the HSD method can show its advantages more, and it is ideal for the morphological position of the target to be detected, but the GS method and the HSD method are the same as other methods. The culled background area, which is an unnecessary part of the detection result map. For the detection of the number 9 in a relatively simple scene, except for the unsatisfactory effect of the RARE method, the results of the other five types of algorithms are relatively close to the images displayed by the artificially marked images. Generally speaking, the target detection method proposed by the present invention is better than the other five types of methods in terms of edge definition, background rejection, and target shape segmentation in various scenes.

Example 8

The salient target detection method based on superpixel segmentation and depth feature location is the same as that in Embodiments 1-5. In this example, the performance analysis of the target detection result is performed through simulation.

The simulation conditions are the same as those in Embodiment 6.

Simulation content: for the selected five hundred images, through the present invention and the global contrast saliency detection GCS (that is, RARE), geodesic distance saliency detection GS, graph theory-based saliency detection GBMR, hierarchical saliency detection HSD, statistical texture saliency The effect of the five types of sex detection STD methods is compared on the same image. The selected images include indoor and outdoor scene images, office scenes, campus areas, parks, etc.

Referring to Fig. 4, the detection effect performance analysis of the present invention and five kinds of methods, reflect the performance of algorithm by accuracy (Precision) index and recall rate (Recall) index, define as follows:

TP: the intersection of the obtained saliency map and the target area in the manually calibrated saliency map;

TN: the intersection of the obtained saliency map and the non-target area in the manually calibrated saliency map;

FP: the intersection of non-target regions in the artificially calibrated saliency map of the obtained saliency map;

FN: the intersection of the target area in the artificially calibrated salient map of the obtained saliency map;

It can be seen that:

Calculate the Precision and Recall indexes of the present invention and the five types of methods respectively. It is not difficult to see that among these comparisons of salient object detection methods, the present invention exhibits a better effect, with an AUC (Area Under Curve) value of 0.6888, and the second best is the GBMR method, with an AUC value of 0.6093. As the recall rate increases, the overall precision tends to decrease. And when the recall rate is 0 to 0.8, the Precision indicators of the present invention are obviously better than other methods. When the recall rate is close to 0.8, the Precision index of the present invention drops below 0.6, which fully shows that the present invention can achieve better detection, closer to the artificially marked map, more complete background removal, and more complete target shape segmentation.

In short, the present invention designs and proposes a salient object detection method based on superpixel segmentation and depth feature location. Using the color similarity linear iterative superpixel segmentation method of five-dimensional Euclidean distance, the processing unit of the image is raised from a single pixel point to a collective similar area, so that the detected target and the edge of the complex background can be clearly separated, which solves the traditional problem. The problem of unsatisfactory object edge segmentation effect of salient object detection method. Fully consider image features such as color features, direction features, and depth features, combined with the prior knowledge that the human eye is more concerned about the center and ignores the surrounding background, the feature similarity of the salient image area, and the uniqueness compared to global features. Then, the localization saliency map and the depth feature saliency map of the input image are generated by simulating these features through the algorithm, which are fused and boundary processed to generate the final saliency map with a mechanism similar to the human eye attention. It makes the computer more logical, more artificial, and has a logical understanding ability similar to that of a human being, or more intelligent, efficient, and more robust. The edge of the image detected by the invention is clearer, the background is eliminated more completely, and the target form segmentation is more complete. It is used in face recognition, vehicle detection, moving target detection and tracking, military missile detection, hospital pathological detection and other fields.

Claims (4)

1. a kind of conspicuousness object detection method positioned based on super-pixel segmentation and depth characteristic, it is characterised in that include Following steps：

Step 1：Input picture carries out the cluster segmentation of linear iteraction to it.Input target image to be detected, is first divided into K Region, finds the partial gradient minimum point of regional neighborhood as central point, and set a tag number to the same area；Seek The central point for looking in range pixel vertex neighborhood five dimension Euclidean distances minimum, and central point label is assigned to pending pixel； Continuous iteration finds the process of the minimum central point of Range Profile vegetarian refreshments, stops changing when the label value of pixel will not change In generation, complete super-pixel segmentation；

Step 2：Build difference of Gaussian generation positioning notable figure.

2a：Gaussian function filtering process is carried out according to the artwork of input, 8 depth scalograms of artwork are generated；

2b：To the scalograms of 8 depths of structure in conjunction with artwork nine layers of scalogram of formation, the red green of nine layers of scalogram picture is extracted Color difference figure and blue yellow color differential chart, totally 18 secondary color difference figure；The intensity map of nine layers of scalogram is extracted, totally 9 is secondary strong Degree figure；The Gabor filtering directional diagrams of nine layers of scalogram are extracted, totally 36 auxiliary direction figure, forms three category feature figures；

2c：Because the size between nine layers of scalogram homogenous characteristics is different, interpolation processing is first passed through to three category feature figures, then carry out Difference processing；

2d：Because the module difference of its feature is, it is necessary to first carry out different types of feature between different types of characteristic pattern Normalization is fused to position notable figure again；

Step 3：Generate depth characteristic notable figure.One first is made to the figure after super-pixel segmentation according to the positioning notable figure of step 2 Localization process, then each region completed for segmentation and its adjacent area gather arest neighbors area information, global area and believed Breath, the category feature information of corner background area information three, generate depth characteristic notable figure, the detection for conspicuousness target；

Step 4：By by step 2 and the final positioning notable figure and depth characteristic notable figure for being able to determine of step 3, positioning is shown Work figure and depth characteristic notable figure are merged and BORDER PROCESSING, generate final conspicuousness target figure, complete super-pixel segmentation Conspicuousness target detection.

2. the conspicuousness object detection method according to claim 1 positioned based on super-pixel segmentation and depth characteristic, its It is characterised by, the super-pixel segmentation of target image to be detected is included having the following steps in step 1：

1.1 first assume that target image one has that pixel is N number of, and it is K to expect the overall area of segmentation, it is clear that each area got Domain has the distance between N/K pixel, and different zones aboutIt is possible that the central point of setting Just appear on edge, in order to avoid the generation of such case, the minimum position of partial gradient is found around the center of setting Put, center is moved on at this partial gradient minimum.And the same area is set a tag number, to mark；

1.2 calculate each pixel to five dimensional feature vector Euclidean distance values of the fixed central point of surrounding neighbors respectively, so Currently processed pixel is assigned the minimum central point tag number of value afterwards,.Calculate five dimensional feature vector C_i=[l_i,a_i,b_i,x_i, y_i]^TEuclidean distance as shown in following three formula, in five dimensional feature vectors, l_i,a_i,b_iFace in CIELAB spaces is represented respectively Position between the brightness of color, red and green, three color component information values of the position between yellow and blueness, x_i,y_iGeneration The co-ordinate position information value of target image to be detected where table pixel.

<mrow> <msub> <mi>d</mi> <mrow> <mi>l</mi> <mi>a</mi> <mi>b</mi> </mrow> </msub> <mo>=</mo> <msqrt> <mrow> <msup> <mrow> <mo>(</mo> <msub> <mi>l</mi> <mi>k</mi> </msub> <mo>-</mo> <msub> <mi>l</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>k</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>b</mi> <mi>k</mi> </msub> <mo>-</mo> <msub> <mi>b</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>

<mrow> <msub> <mi>d</mi> <mrow> <mi>x</mi> <mi>y</mi> </mrow> </msub> <mo>=</mo> <msqrt> <mrow> <msup> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>-</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>k</mi> </msub> <mo>-</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>

<mrow> <msub> <mi>D</mi> <mi>i</mi> </msub> <mo>=</mo> <msub> <mi>d</mi> <mrow> <mi>l</mi> <mi>a</mi> <mi>b</mi> </mrow> </msub> <mo>+</mo> <mfrac> <mi>m</mi> <mi>s</mi> </mfrac> <msub> <mi>d</mi> <mrow> <mi>x</mi> <mi>y</mi> </mrow> </msub> </mrow>

In above formula, d_labWhat is represented is the Euclidean distance of pixel k and central point i in terms of CIELAB color spaces；d_xyRepresent It is the Euclidean distance of pixel k and central point i in terms of spatial coordinate location；D_iFor evaluate pixel k and central point i whether institute Belong to the criterion of a label, its value is bigger, and similarity degree between the two is closer to label is just consistent；M is one solid Parameter is determined, for the relation between balance variable；S is that the distance between different zones are about

Above is setting an iteration cycle of the affiliated tag number of pixel.

1.3 are constantly iterated operation according to 1.2 steps, and the degree of accuracy to the tag number belonging to pixel makees further excellent Change, untill the affiliated tag number of each pixel of whole sub-picture no longer changes.

1.4 pass through iterative process, and the isolated area or the single pixel point that isolates of crossing small size are distributed to mark nearby Sign, completes the super-pixel segmentation of target image.

3. the conspicuousness object detection method according to claim 1 positioned based on super-pixel segmentation and depth characteristic, its It is characterised by that the category feature information of collection three described in step 3 comprises the following steps：

3.1 are directed to each region after the completion of super-pixel segmentation, gather the information in its closest range areas, i.e. arest neighbors Area information；

3.2 split the information that other regions of each region completed are included, i.e. global area letter by removing current region Breath；

The information of four corner areas of 3.3 each the Regional Representative's background characteristics completed for segmentation, i.e. corner background area Domain information.

4. the conspicuousness object detection method according to claim 1 positioned based on super-pixel segmentation and depth characteristic, its It is characterised by that the class area information of the use three generation depth characteristic notable figure described in step 3 has been specifically included：

The one of region R completed for segmentation, quantifying its depth characteristic significance is：

<mrow> <mi>S</mi> <mrow> <mo>(</mo> <mi>R</mi> <mo>)</mo> </mrow> <mo>=</mo> <munder> <mo>&amp;Pi;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mi>C</mi> <mo>,</mo> <mi>G</mi> <mo>,</mo> <mi>B</mi> </mrow> </munder> <mi>s</mi> <mrow> <mo>(</mo> <mi>R</mi> <mo>,</mo> <msup> <mi>&amp;psi;</mi> <mi>m</mi> </msup> <mo>)</mo> </mrow> </mrow>

Wherein, s (R) represents region R significance；Π is multiple fac-tors；s(R,ψ^C) represent arest neighbors area information；s (R,ψ^G) represent global area information；s(R,ψ^B) represent corner background area information；

Low probability is the high expression of information content, and probability height but illustrates that its information content brought is relatively low；By s (R, ψ^m) definition is such as Under：

s(R,ψ^m)=- log (p (R | ψ^m))

In above formula, s (R, ψ^m) represent the arest neighbors area information extracted under depth characteristic, global area information, corner background area Information；P is a probable value.

To arest neighbors area information, global area information, corner background area believes that three class area informations use its zone leveling respectively Value simplifies above formula：

<mrow> <mi>p</mi> <mrow> <mo>(</mo> <mi>R</mi> <mo>|</mo> <msup> <mi>&amp;psi;</mi> <mi>m</mi> </msup> <mo>)</mo> </mrow> <mo>=</mo> <mover> <mi>p</mi> <mi>&amp;Lambda;</mi> </mover> <mrow> <mo>(</mo> <mi>d</mi> <mo>|</mo> <msup> <mi>D</mi> <mi>m</mi> </msup> <mo>)</mo> </mrow> </mrow>

D represents to be presently in the region unit R of reason depth-averaged value in above formula；It is above institute The ψ referred to^mDepth-averaged value, whereinRepresent the depth-averaged value in i-th piece of region of m class area information features, n_mAltogether There are three kinds of situations, n_CRepresent the sum of arest neighbors area information, n_GRepresent the sum of global area information, n_BRepresent corner background The sum of area information；

Estimated with Gaussian functionRealization：

<mrow> <mover> <mi>p</mi> <mi>&amp;Lambda;</mi> </mover> <mrow> <mo>(</mo> <mi>d</mi> <mo>|</mo> <msup> <mi>D</mi> <mi>m</mi> </msup> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <msub> <mi>n</mi> <mi>m</mi> </msub> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <msub> <mi>n</mi> <mi>m</mi> </msub> </munderover> <msup> <mi>e</mi> <mrow> <mo>-</mo> <mfrac> <mrow> <mo>|</mo> <mo>|</mo> <mi>d</mi> <mo>-</mo> <msubsup> <mi>d</mi> <mi>i</mi> <mi>m</mi> </msubsup> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> </mrow> <mrow> <mn>2</mn> <msup> <mrow> <mo>(</mo> <msubsup> <mi>&amp;sigma;</mi> <mi>d</mi> <mi>m</mi> </msubsup> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </mfrac> </mrow> </msup> </mrow>

In above formula,Represent the factor of influence of the depth difference opposite sex of different zones block.