CN109543568A - A kind of vehicle-logo location method - Google Patents

Abstract

The present application belongs to the technical field of image processing, and in particular relates to a vehicle logo positioning method. Most of the existing vehicle logo localization methods use traditional methods, and the localization rate and robustness are relatively poor. The present application provides a method for locating a vehicle logo. The method includes the following steps: Step 1): acquiring an image of the front face of the vehicle; Step 2): extracting primary features of the image; Step 3): using a multi-scale spectral residual method to calculate and fuse to obtain In the final saliency map, the saliency map is divided into multiple sub-salient regions to form multiple focus points; Step 4): Calculate the complexity of each region; Step 5): Use the improved ant colony algorithm to optimize the focus transfer path; Step 6 ): Determine the vehicle logo area. The vehicle logo localization algorithm of the present application has better localization rate and robustness.

Description

A method of car logo positioning

technical field

The present application belongs to the technical field of image processing, and in particular relates to a method for locating a vehicle logo.

Background technique

Car logo refers to the logos of various car brands, which are mainly used for vehicle identification. These logos often become the representatives of automobile enterprises. The logo of a car includes the car's trademark or factory logo, product label, engine model and factory serial number, vehicle model and factory serial number, and vehicle identification code. According to my country's national regulations, these signs must be checked during new car registration and annual inspection. The car should have its own logo to indicate the car manufacturer, model, engine power, load-bearing quality, engine and vehicle serial number, etc. Their role is to facilitate the identification of the vehicle's "identity" by sellers, users, maintenance personnel, and traffic management departments.

The car logo is an important attribute of a vehicle. Different from the license plate, the car logo is difficult to replace and alter. It has a wide range of applications in the fields of robbery and robbery, automatic recording of illegal vehicles, unmanned management of parking lots, and automatic toll collection at bridges and intersections. However, most of the existing vehicle logo localization methods use traditional methods, and the localization rate and robustness are relatively poor.

SUMMARY OF THE INVENTION

1. Technical problems to be solved

Based on the car logo is an important attribute of the vehicle, different from the license plate, the car logo is difficult to replace and alter. However, most of the existing vehicle logo positioning methods use traditional methods, and the positioning rate and robustness are relatively poor. The present application provides a vehicle logo positioning method.

2. Technical solutions

In order to achieve the above purpose, the present application provides a method for locating a vehicle logo, the method comprising the following steps:

Step 1): Obtain the front face image of the vehicle;

Step 2): extract the primary features of the image;

Step 3): adopt the multi-scale spectral residual method to calculate and fuse to obtain the final saliency map, and divide the saliency map into multiple sub-salient regions to form multiple focus points;

Step 4): Calculate the complexity of each area;

Step 5): using the improved ant colony algorithm to optimize the focus transfer path;

Step 6): Determine the vehicle logo area.

Optionally, the primary feature in step 2) includes one or more of color, brightness or direction.

Optionally, the color and brightness feature extraction is to extract a saliency map about the color feature and the brightness feature by subjecting the original image to the calculation of a nonlinear anisotropic diffusion equation.

Optionally, the color and brightness feature extraction method is as follows:

The Gaussian filter function G(x, y, σ) is improved to a regularized nonlinear anisotropic diffusion equation, that is, the P-M equation:

The scale σ is 1/2 and 1/4 of the original image, respectively; the original input image I(x, y) is sub-sampling and low-pass filtering step by step under different scale filters, and the gradient modulus value of the image is used at the same time. The filtering is combined with the edge detection of the image, and the diffusion coefficient is changed according to the information of the image And the diffusion coefficient reaches a minimum value at the edge of the image, according to the gradient of each iteration of the image Then use the regularized PM equation to obtain the image after nonlinear diffusion filtering; the solution I _σ (x, y, t) calculated in the above formula is the image after regularization and filtering.

Optionally, the directional feature extraction is to extract orientation-sensitive features through a Gabor filter, and convert the original image into an image of the orientation feature;

The function of the Gabor filter is expressed as:

where x' and y' are respectively:

x′=xcosθ+ysinθ

y′=-xsinθ+ycosθ

Among them, the parameter θ is the orientation of the Gabor filter, f ₀ is the center frequency, σ _x and σ _y are the Gaussian function variances in the x′ and y′ directions of the spatial domain, respectively; four different orientations (θ∈[0°, 45 °, 90°, 135°]) filter the original input image I(x, y) to form 4 orientation feature maps {R _k (x, y), k=1, 2, 3, , 4}, and then normalize It is normalized into an orientation feature saliency map; its output on (x ₀ , y ₀ ) is expressed as:

R(xy)=l(x, y)*h(xx ₀ y-)y ₀ )

In the formula, x ₀ , y ₀ are equivalent to the position of the center of the receptive field, and * represents the convolution operation.

Optionally, the method for calculating and merging the multi-scale spectral residual method to obtain the final saliency map includes:

The calculation process of saliency under multi-scale is as follows:

in, The input image is I(x), and the Fourier transform of it, the amplitude spectrum A(f) and the phase spectrum P(f) are obtained; h is a convolution kernel of a 3*3 mean filter, and R(f) is the frequency domain. Residual spectrum; S(x) is the significant area map;

The saliency map of the same feature at different scales is supplemented and adjusted to the same scale as the original image, and then weighted and fused to obtain the final feature saliency map; since the saliency map is a grayscale image, the scale is For each feature saliency map of , the remaining pixels are supplemented with a gray value of 0; for the color and brightness feature saliency map of each scale, the calculation formula is:

The different feature saliency maps, that is, the color, brightness and direction feature saliency maps, are normalized and merged to obtain the final global saliency map A:

S ₃ is the saliency map of the direction feature obtained by Gabor filter fusion, the weight γ ₁ +γ ₂ +γ ₃ =1, and γ ₁ =γ ₂ =γ ₃ =1/3.

Optionally, the complexity of each region is calculated by comprehensively analyzing the three complexity indicators of quality symmetry, image composition complexity and shape complexity, and performing linear weighted fusion to obtain the significant sub-regions A ₁ and A respectively. ₂ ... A _j ... _An of the area complexity C ₁ , C ₂ ... C _j ... C _n ;

The area complexity algebra is expressed as:

C _j =λQ _dj +κC′ _dj +μE _j

Among them, λ, κ, and μ are standardized coefficients.

Optionally, the improved ant colony algorithm is to traverse multiple focal points, and judge the reliability of the vehicle logo of this focal point according to the regional complexity and edge information, and use the regional complexity to drive the ants to preferentially access the points with higher complexity. , speed up the convergence speed, and introduce the ant error rate at the same time, so that the ants do not go to the target with more pheromone according to a certain probability, so as to avoid the algorithm from falling into the local optimal solution;

There are m ants, select the next target according to the probability function with the target distance as the heuristic factor and the number of pheromone on the path as the variable, establish the ant tabu _k table, and add the focus visited by the ants to the taboo list; set the first An ant is a mistake ant and only goes to the target with low pheromone, so the probability of the kth ant transferring from target i to target j is as follows:

where C _j is the regional complexity of the next target j, ω is the importance coefficient of the regional complexity, allowed _k ∈({1, 2,...n}-tabu _k ) is the target that ants are allowed to choose, η _ij =1/d _0j is the heuristic function of the path (i, j), d _0j is the length of the path from the origin to the target j, τ _ij is the pheromone concentration on the path; α represents the importance coefficient of the path, and β represents the relative heuristic factor. Equilibrium coefficient, ρ is the pheromone persistence factor, that is, the existence intensity of pheromone;

At the initial moment, the number of pheromone on each path is the same, that is, τ _ij (0)=C is a constant, and each time the ants complete a search, the pheromone content on each path is updated. The pheromone increment on the target path with large area complexity is given additional enhancement, and the importance factor is introduced That is, when the complexity of the next transfer target is greater than the previous target, the pheromone increment is larger, otherwise, the pheromone increment is suppressed to a certain extent, and the ant-week model is used to update the global information. The pheromone increment and pheromone update method are as follows: :

The kth ant passes through (ij) in this cycle

Among them, Δτ _ij represents the pheromone increment on the path ij in this cycle; L _k represents the path length of the kth ant to travel around a week, and Q is the information intensity, which means that the ants release on the path they traveled during a round trip total amount of pheromones.

Optionally, the optimized focus transfer path is to number each attention focus from 1 to n respectively; to initialize parameters, m ants are randomly placed on n vertices, and τ _ij (t) is initialized as Δτ _ij (0 )=C, initialize the pheromone increment Δτ _ij (t), N _c ←0 (N _c is the number of iterations), set the upper limit of the number of iterations; set the tabu _k of the ant k, and place the initial starting point of each ant in the current In the taboo list; calculate the probability of each ant moving to the next target j and move the ants, add this target point j to the taboo list; calculate the path length L _k of each ant (k=1, 2, 3...m) and the pheromone increment on the path (i, j), modify the pheromone intensity on the path (i, j) according to the pheromone update equation; for each path (i, j), set Δτ _ij =0, nc←nc +1; if N _c reaches the predetermined number of iterations or no better solution appears, exit the loop, otherwise set the tabu _k of the ant k, and put the initial starting point of each ant in the current tabu table; output the current best solution, that is, the optimal path U.

Optionally, the reliability of the vehicle logo area determined as a sub-area is defined as:

The constraint condition is the focus transfer path U optimized by the ant colony, and the reliability of the sub-regions is determined one by one according to the focus transfer path U. When CRE _U = 3 in a certain region, it is determined that this region is the vehicle logo region, and the determination is terminated; Among them, P _Ci is the regional complexity, P _Ei is the boundary ratio, and unicity _i is the uniformity of the vehicle logo in the front face image of the vehicle.

3. Beneficial effects

Compared with the prior art, the beneficial effects of a vehicle logo positioning method provided by the present application are:

The vehicle logo positioning method provided by this application combined with the improved ant colony algorithm visual attention mechanism, uses the center-of-mass symmetry characteristics of the car logo, and proposes a new index of "mass symmetry", and uses the mass symmetry, image composition complexity and shape complexity. Three indicators measure the regional complexity, and an improved ant colony algorithm is proposed. The regional complexity and ant error rate are introduced into the ant colony algorithm to speed up the algorithm convergence speed and alleviate the influence of easily falling into local optimum; The positioning algorithm, the vehicle logo positioning algorithm proposed in this application, the vehicle logo positioning rate can reach up to 98.43%; it is better than the traditional symmetry detection vehicle logo positioning, and can also meet the real-time requirements; the stability and robustness of the algorithm are better than the other two. an algorithm. The vehicle logo localization algorithm of the present application has better localization rate and robustness.

Description of drawings

FIG. 1 is a schematic flowchart of a method for locating a vehicle logo according to the present application.

Detailed ways

Hereinafter, specific embodiments of the present application will be described in detail with reference to the accompanying drawings, from which those skilled in the art can clearly understand the present application and be able to implement the present application. Without departing from the principles of the present application, the features of the various embodiments may be combined to obtain new embodiments, or instead of certain features of certain embodiments, to obtain other preferred embodiments.

The saliency-based Gaussian pyramid model proposed by itti-koch, which uses the calculation method of the central peripheral difference based on non-uniform sampling to obtain the saliency measurement, but the calculation method only considers the local characteristics of the salient area, and does not consider the whole image Global information for the image.

The target-based computational model using multiscale analysis and grouping proposed by Luo Siwei et al. uses differential operators to extract edges, and contour grouping derived from Gestalt perceptual organization organizes edges into perceptual targets, but the model is computationally more efficient but grouping The main clue is occlusion, which limits the scope of application of this model.

The spectral residual significant region detection method proposed by H Xiaodi based on frequency domain signal analysis adopts the information theory information structure method to filter redundant information from the frequency domain to obtain useful information, but the calculation saliency effect is obvious and the real-time performance is strong, but this method The saliency is only calculated at a single scale, and other local features of the image such as color features cannot be well described.

Radharkrishna et al. proposed that the saliency value is directly defined by the color difference between the pixel and the average color of the whole image, which has the advantages of good real-time performance, simplicity and speed, but the calculation effect is not ideal for images with complex backgrounds. In terms of attention focus shift, the currently used hierarchical search mechanism from coarse scale to fine scale is not efficient and has poor real-time performance.

In machine vision, saliency is a pattern of image partitioning, and a saliency map is an image that shows the uniqueness of each pixel. The goal of a saliency map is to simplify or change the representation of a general image into a style that is easier to analyze. For example, a pixel with a higher gray level in a color map will be displayed more prominently in the saliency map. From the point of view of visual stimuli, if certain features are particularly able to capture attention, such properties are called saliency in psychology.

Ant Colony Algorithm (AG) is a simulation optimization algorithm that simulates the foraging behavior of ants. It was first proposed by Italian scholar Dorigo M et al. in 1991, and was first used to solve TSP (Traveling Salesman Problem).

The basic principle of ant colony algorithm:

1. Ants release pheromones on the path.

2. When encountering an intersection that has not yet been passed, choose a road at random. At the same time, the pheromone related to the path length is released.

3. Pheromone concentration is inversely proportional to path length. When later ants encounter the intersection again, they choose the path with higher pheromone concentration.

4. The pheromone concentration on the optimal path is increasing.

5. Finally, the ant colony finds the optimal foraging path.

Ant-Cycle Model (Ant-Cycle Model)

Ant-Quantity Model (Ant-Quantity Model)

Ant-Density Model (Ant-Density Model)

the difference:

1. The ant week model uses global information, that is, the ants update the pheromone on all paths after completing a cycle;

2. The ant quantity and ant density models use local information, that is, the ants update the pheromone on the path after completing one step.

The so-called traversal (traversal) refers to following a certain search route, one and only one visit to each node in the tree in turn. The operations performed by the access node depend on the specific application problem. Traversal is one of the most important operations on a binary tree, and it is the basis for other operations on a binary tree. Of course, the concept of traversal is also suitable for multi-element collections, such as arrays.

Gabor is a linear filter for edge extraction. Its frequency and direction expression are similar to those of the human visual system. It can provide good direction selection and scale selection characteristics, and is insensitive to illumination changes, so it is very suitable for texture analysis.

Referring to FIG. 1, the present application provides a method for locating a vehicle logo, and the method includes the following steps:

Step 1): Obtain the front face image of the vehicle;

Step 2): extract the primary features of the image;

Step 3): adopt the multi-scale spectral residual method to calculate and fuse to obtain the final saliency map, and divide the saliency map into multiple sub-salient regions to form multiple focus points;

Step 4): Calculate the complexity of each area;

Step 5): using the improved ant colony algorithm to optimize the focus transfer path;

Step 6): Determine the vehicle logo area.

Optionally, the primary feature in step 2) includes one or more of color, brightness or direction.

Optionally, the color and brightness feature extraction is to extract a saliency map about the color feature and the brightness feature by subjecting the original image to the calculation of a nonlinear anisotropic diffusion equation.

Optionally, the color and brightness feature extraction method is as follows:

The Gaussian filter function G(x, y, σ) is improved to a regularized nonlinear anisotropic diffusion equation, that is, the P-M equation:

The scale σ is 1/2 and 1/4 of the original image, respectively; the original input image I(x, y) is sub-sampling and low-pass filtering step by step under different scale filters, and the gradient modulus value of the image is used at the same time. The filtering is combined with the edge detection of the image, and the diffusion coefficient is changed according to the information of the image And the diffusion coefficient reaches a minimum value at the edge of the image, according to the gradient of each iteration of the image Then use the regularized PM equation to obtain the image after nonlinear diffusion filtering; the solution I _σ (x, y, t) calculated in the above formula is the image after regularization and filtering. After improved Gaussian filtering, six saliency maps in total are obtained under three scales σ (original image, 1/2 of the original image, and 1/4 of the original image).

Optionally, the directional feature extraction is to extract orientation-sensitive features through a Gabor filter, and convert the original image into an image of the orientation feature;

The function of the Gabor filter is expressed as:

where x' and y' are respectively:

x′=x cosa+y sinθ

y′=-x sinθ+y cosθ

Among them, the parameter θ is the orientation of the Gabor filter, f ₀ is the center frequency, σ _x and σ _y are the Gaussian function variances in the x′ and y′ directions of the spatial domain, respectively; four different orientations (θ∈[0°, 45 °, 90°, 135°]) filter the original input image I(x, y) to form 4 orientation feature maps {R _k (x, y), k=1, 2, 3, , 4}, and then normalize It is normalized into an orientation feature saliency map; its output on (x ₀ , y ₀ ) is expressed as:

R(x, y)=I(x, y)*h(xx ₀ , yy ₀ )

In the formula, x ₀ , y ₀ are equivalent to the position of the center of the receptive field, and * represents the convolution operation.

Optionally, the method for calculating and merging the multi-scale spectral residual method to obtain the final saliency map includes:

The calculation process of saliency under multi-scale is as follows:

in, The input image is I(x), and the Fourier transform of it, the amplitude spectrum A(f) and the phase spectrum P(f) are obtained; h is a convolution kernel of a 3*3 mean filter, and R(f) is the frequency domain. Residual spectrum; S(x) is the saliency area map; using the above algorithm, a total of 7 saliency maps of color, brightness, and direction features at each scale can be obtained.

The saliency map of the same feature at different scales is supplemented and adjusted to the same scale as the original image, and then weighted and fused to obtain the final feature saliency map; since the saliency map is a grayscale image, the scale is For each feature saliency map of , the remaining pixels are supplemented with a gray value of 0; for the color and brightness feature saliency map of each scale, the calculation formula is:

The different feature saliency maps, that is, the color, brightness and direction feature saliency maps, are normalized and merged to obtain the final global saliency map A:

S ₃ is the saliency map of the direction feature obtained by Gabor filter fusion, the weight γ ₁ +γ ₂ +γ ₃ =1, and γ ₁ =γ ₂ =γ ₃ =1/3.

Optionally, the complexity of each region is calculated by comprehensively analyzing the three complexity indicators of quality symmetry, image composition complexity and shape complexity, and performing linear weighted fusion to obtain the significant sub-regions A ₁ and A respectively. ₂ ... A _j ... _An of the area complexity C ₁ , C ₂ ... C _j ... C _n ;

The area complexity algebra is expressed as:

C _j =λQ _dj +κC′ _dj +μE _j

Among them, λ, κ, and μ are standardized coefficients.

According to the multi-scale global saliency map calculation method, the saliency map A (grayscale map) of the front face image of the vehicle is obtained, and after segmentation, n sub-saliency regions A ₁ , A ₂ . . . A _j . . . A _n are obtained. In the subsequent focus transfer, the sub-salient regions need to be identified to discriminate the vehicle logo region. It can be intuitively obtained from human vision that the distinction between the complexity of the vehicle logo area and the background is relatively large, so the computational area complexity is proposed as one of the indicators to measure the vehicle logo area.

The present application proposes to use three indicators of quality symmetry, image composition complexity, and shape complexity to measure the regional complexity.

(1) Mass symmetry. Since the center of mass of the car logo is on the central axis of the car logo area, the quality difference between the left and right sides of the central axis of the significant sub-region is calculated, that is, the difference in the number of pixels whose gray value is (0, T ₁ ), and the threshold is set to 0<T ₁ ≤255, Th _D ≥0, define the quality symmetry degree

in is the sum of the masses on the left side of the central axis, is the sum of the masses on the right side of the central axis.

(2) Image composition complexity. According to the information entropy theory proposed by Zhang Xuewen, the composition complexity can be defined by the richness of the internal state of the generalized set. The specific method is: if there are k different flag values x ₁ , x ₂ ,... _xi ,…x _k , and the corresponding number of individuals are n ₁ , 2 _n ,…n _i ,…n _k , then the complexity of the generalized set can be obtained Therefore, in the salient sub-region, the generalized probability of the occurrence of each gray value pixel can define the image composition complexity, that is, The unit is bit. where p _i is P(x _i )=n _i /N, which represents the probability of the occurrence of a pixel with a grayscale value of i. The larger C is, the more complex the image composition is.

(3) Shape complexity. The shape of the car logo is special and has certain regularity. The shape complexity includes not only the external contour information, but also the internal topology of the car logo. The internal topological structure of the car logo is complex and the boundary line is long, so the edge ratio is used to describe the shape complexity. Boundary ratio n _edge is the length of the border, that is, the number of border pixels, and N is the total number of pixels in the sub-region. The larger the _Ej , the larger the shape complexity.

This application obtains a global saliency map A through multi-scale saliency calculation, and divides it into n salient sub-regions, that is, there are n attention points. By traversing the n focal points and judging the reliability of the vehicle logo of this focal point according to the regional complexity and edge information, it can actually be transformed into a path planning problem. According to the specific characteristics of the vehicle front face picture, this application introduces an ant colony algorithm guided by regional complexity and ant error rate to improve its convergence speed and avoid falling into a local optimal solution.

The purpose of this application is to quickly and accurately discriminate the vehicle logo area, and the termination condition of the discrimination is that a certain target satisfies the area complexity and edge information. According to the conventional ant colony algorithm, the vehicle logo area may be visited later in the optimal path, thus prolonging the convergence time of the entire algorithm. This application makes use of the high complexity of the vehicle logo area, and uses the area complexity to drive ants to preferentially access the more complex points to speed up the convergence speed. At the same time, the ant error rate is introduced, so that the ants do not go to the target with more pheromone according to a certain probability. , to avoid the algorithm falling into a local optimal solution. The focus transfer mechanism of this application follows the following three principles: 1) the principle of no return, that is, the focus is not repeatedly accessed; 2) the principle of complexity priority, which drives the ants to visit the focus with greater complexity first; 3) the principle of adjacent second priority, in the On the basis of prioritizing complexity, the planned path should better follow the law of human eye focus transfer as much as possible.

Optionally, the improved ant colony algorithm is to traverse multiple focal points, and judge the reliability of the vehicle logo of this focal point according to the regional complexity and edge information, and use the regional complexity to drive the ants to preferentially access the points with higher complexity. , speed up the convergence speed, and introduce the ant error rate at the same time, so that the ants do not go to the target with more pheromone according to a certain probability, so as to avoid the algorithm from falling into the local optimal solution;

There are m ants, and the next target is selected according to the probability function that takes the target distance as the heuristic factor and the number of pheromone on the path as the variable. The ant selection strategy is also related to the complexity of the next target area. The focus visited by the ant is added to the taboo list; the first ant is set as the error ant, and only goes to the target with low pheromone, so the probability of the kth ant transferring from target i to target j is as follows:

where C _j is the regional complexity of the next target j, ω is the importance coefficient of the regional complexity, allowed _k ∈({1, 2,...n}-tabu _k ) is the target that ants are allowed to choose, η _ij =1/d _0j is the heuristic function of the path (i, j), d _0j is the length of the path from the origin to the target j, τ _ij is the pheromone concentration on the path; α represents the importance coefficient of the path, and β represents the relative heuristic factor. Equilibrium coefficient, ρ is the pheromone persistence factor, that is, the existence intensity of pheromone;

At the initial moment, the number of pheromone on each path is the same, that is, τ _ij (0)=C is a constant, and each time the ants complete a search, the pheromone content on each path is updated. The pheromone increment on the target path with large area complexity is given additional enhancement, and the importance factor is introduced That is, when the complexity of the next transfer target is greater than the previous target, the pheromone increment is larger, otherwise, the pheromone increment is suppressed to a certain extent, and the ant-cycle model is used to update the global information. The pheromone increment and The pheromone update method is:

The kth ant passes through (ij) in this cycle

Among them, Δτ _ij represents the pheromone increment on the path ij in this cycle; L _k represents the path length of the kth ant to travel around a week, and Q is the information intensity, which means that the ants release on the path they traveled during a round trip total amount of pheromones.

Optionally, the optimized focus transfer path is to number each attention focus from 1 to n; initialization parameters are to randomly place m ants on n vertices, and τ _ij (t) is initialized to τ _ij (0 )=C, initialize the pheromone increment Δτ _ij (t), N _c ←0 (N _c is the number of iterations), set the upper limit of the number of iterations; set the tabu _k of the ant k, and place the initial starting point of each ant in the current In the taboo list; calculate the probability of each ant moving to the next target j and move the ants, add this target point j to the taboo list; calculate the path length L _k of each ant (k=1, 2, 3...m) and the pheromone increment on the path (i, j), modify the pheromone intensity on the path (i, j) according to the pheromone update equation; for each path (i, j), set Δτ _ij =0, nc←nc +1; if N _c reaches the predetermined number of iterations or no better solution appears, exit the loop, otherwise set the tabu _k of the ant k, and put the initial starting point of each ant in the current tabu table; output the current best solution, that is, the optimal path U.

The complexity of the car logo area is relatively large, but the area with high regional complexity is not necessarily the car logo area, that is, the regional complexity should be regarded as a necessary and insufficient condition for determining the car logo area. On this basis, the previously extracted targets are also considered. The edge information, that is, the area with a larger border ratio is also more likely to be a vehicle logo area. In addition, because the car logo in the front face image of the vehicle is not repetitive, that is, unicity, and the headlights, fog lights, turn signals, etc. all appear in pairs, the only area of complexity information is the car logo area. most likely. Therefore, it is necessary to define the area where the three conditions are satisfied as the vehicle logo area.

Optionally, the reliability of the vehicle logo area determined as a sub-area is defined as:

The constraint condition is the focus transfer path U optimized by the ant colony, and the reliability of the sub-regions is determined one by one according to the focus transfer path U. When CRE _U = 3 in a certain region, it is determined that this region is the vehicle logo region, and the determination is terminated; Among them, P _Ci is the regional complexity, P _Ei is the boundary ratio, and unicity _i is the uniformity of the vehicle logo in the front face image of the vehicle. In this way, through the layer-by-layer judgment of the three indicators, the reliability of the determined vehicle logo area is greatly increased.

(1) Extract the primary visual features of the front face image of the vehicle, and use the multi-scale spectral residual method to calculate and fuse to obtain the final saliency map, which comprehensively considers the local information and global information of the image.

(2) Using the symmetry characteristics of the center of mass of the car logo, a new index of "mass symmetry" is proposed, and three indicators of mass symmetry, image composition complexity and shape complexity are used to measure the regional complexity.

(3) The improved ant colony algorithm is used to optimize the focus transfer path, and the regional complexity drives the ants to preferentially visit the points with higher complexity, and introduces the adjustment factor of the ant error rate, which effectively increases the path search efficiency and avoids falling into the local optimal solution.

The vehicle logo positioning method provided by this application combined with the improved ant colony algorithm visual attention mechanism, uses the center-of-mass symmetry characteristics of the car logo, and proposes a new index of "mass symmetry", and uses the mass symmetry, image composition complexity and shape complexity. Three indicators measure the regional complexity, and an improved ant colony algorithm is proposed. The regional complexity and ant error rate are introduced into the ant colony algorithm to speed up the algorithm convergence speed and alleviate the influence of easily falling into local optimum; The positioning algorithm, the vehicle logo positioning algorithm proposed in this application, the vehicle logo positioning rate can reach up to 98.43%; it is better than the traditional symmetry detection vehicle logo positioning, and can also meet the real-time requirements; the stability and robustness of the algorithm are better than the other two. an algorithm. The vehicle logo localization algorithm of the present application has better localization rate and robustness.

Although the present application has been described above with reference to specific embodiments, it will be understood by those skilled in the art that many modifications may be made in configuration and detail disclosed herein within the spirit and scope of the present disclosure. The scope of protection of the present application is to be determined by the appended claims, and the claims are intended to cover all modifications encompassed by the literal meaning or scope of equivalents to the technical features in the claims.

Claims (10)

1. A car logo positioning method is characterized in that: the method comprises the following steps:

step 1): acquiring a front face image of a vehicle;

step 2): extracting primary features of the image;

step 3): calculating and fusing by adopting a multi-scale spectrum residual method to obtain a final saliency map, and dividing the saliency map into a plurality of sub saliency areas to form a plurality of attention focuses;

step 4): calculating the complexity of each region;

step 5): optimizing a focus transfer path by adopting an improved ant colony algorithm;

step 6): and judging the car logo area.

2. The emblem positioning method of claim 1, characterized in that: the primary features in the step 2) comprise one or more of color, brightness or direction.

3. The emblem positioning method of claim 2, characterized in that: the color and brightness feature extraction is to extract a saliency map about color features and brightness features from an original image through calculation of a nonlinear anisotropic diffusion equation.

4. The emblem positioning method of claim 3, characterized in that: the color and brightness feature extraction method comprises the following steps:

the gaussian filter function G (x, y, σ) is modified to a regularized nonlinear anisotropic diffusion equation, P-M equation:

the scale σ is 1/2, 1/4 of the original image, respectively; the method comprises the steps of performing sub-sampling and low-pass filtering on an original input image I (x, y) step by step under different scale filters, simultaneously combining the filtering with the edge detection of the image by using the gradient module value of the image, and changing the diffusion coefficient according to the information of the imageAnd the diffusion coefficient reaches a minimum value at the edge of the image according to the gradient of the image obtained by each iterationThe size of the image is subjected to edge judgment, and then a regularized P-M equation is used for obtaining an image subjected to nonlinear diffusion filtering; solution I calculated in the above equation_σ(x, y, t) is the processRegularizing the filtered image.

5. The emblem positioning method of claim 2, characterized in that: the directional feature extraction is to extract orientation-sensitive features through a Gabor filter and convert an original image into an image of the orientation features;

the function of the Gabor filter is expressed as:

wherein x ', y' are respectively:

x′＝xcosθ+ysinθ

y′＝-xsin9+ycosθ

where the parameter θ is the orientation of the Gabor filter, f₀Is the center frequency, σ_xAnd σ_yGaussian function variances in the spatial domain x 'and y' directions, respectively; using 4 different orientations (theta e [0 deg. ], 45 deg., 90 deg., 135 deg. ]]) Filtering the original input image I (x, y) to form 4 orientation feature maps { R }_k(x, y), k is 1, 2, 3, 4}, and then normalized to an orientation feature saliency map; which is in (x)₀，y₀) The output of (c) is represented as:

R(x，y)＝I(x，y)*h(x-x₀，y-y₀)

in the formula, x₀,y₀The position corresponding to the center of the field represents the convolution operation.

6. The emblem positioning method of claim 1, characterized in that: the method for obtaining the final saliency map through calculation fusion by the multi-scale spectrum residual error method comprises the following steps:

the process of calculating the significance under multiple scales is as follows:

wherein,the input image is I (x), Fourier transform is carried out on the input image to obtain an amplitude spectrum A (f) and a phase spectrum P (f); h is a convolution kernel with 3 × 3 mean filtering, and R (f) is a frequency domain residual spectrum; s (x) is a saliency region map;

performing gray level supplement adjustment on the same characteristic saliency map under different scales to the same scale with the original image size, and performing weighted fusion to obtain final characteristic saliency maps; since the saliency map is a grayscale map, for a scale ofSupplementing the rest pixels of each feature saliency map into gray values of 0; the calculation formula of the color and brightness feature saliency map for each scale is as follows:

different feature saliency maps, namely color, brightness and direction feature saliency maps are normalized and merged to obtain a final global saliency map A:

wherein S₃The weight gamma is a direction feature saliency map obtained by Gabor filtering fusion₁+γ₂+γ₃1, take γ₁＝γ₂＝γ₃＝1/3。

7. The emblem positioning method of claim 1, characterized in that: the complexity of each region is calculated into three complexity indexes of comprehensive analysis quality symmetry, image composition complexity and shape complexity, and linear weighted fusion is carried out to respectively obtain each significant subregion A₁、A₂...A_j...A_nRegion complexity C of₁、C₂...C_j...C_n；

The region complexity algebraic representation is:

C_j＝λQ_dj+κC′_dj+μE_j

wherein λ, κ, μ are normalization coefficients.

8. The emblem positioning method of claim 1, characterized in that: the improved ant colony algorithm is characterized in that through traversal of a plurality of focuses, the reliability of the car logo of the focus is judged according to the regional complexity and the edge information, the regional complexity is adopted to drive ants to preferentially access points with higher complexity, the convergence speed is accelerated, meanwhile, the ant error rate is introduced, the ants do not walk to targets with more pheromones according to a certain probability, and the algorithm is prevented from falling into a local optimal solution;

m ants are arranged, the next target is selected according to a probability function taking the target distance as a heuristic factor and the quantity of pheromones on the path as a variable, and an ant taboo table tabu is established_kAdding the focus visited by the ants into a taboo list; the first ant is set as a miss ant and only goes to the target with low pheromone, so the probability that the kth ant transfers from the target i to the target j is as follows:

wherein C is_jFor the region complexity of the next target j, ω is the importance coefficient of the region complexity, allowed_k∈({1，2，...n}-tabu_k) Goal to allow ants to choose, η_ij=1/d_0jAs a heuristic function of the path (i, j), d_0jLength of path from origin to target j, τ_ijα represents the importance coefficient of the path, β represents the relative balance coefficient of the heuristic factor, and rho is the pheromone persistence factor, namely the pheromone existence intensity;

at the initial time, the number of information elements on each path is the same, i.e. τ_ij(0) And C is a constant, and the pheromone content on each path is updated by the ants each time the ants complete the search. The pheromone increment on a target path with large complexity in a specified area is additionally enhanced, and an importance factor is introducedWhen the complexity of the next transfer target is larger than that of the previous target, the pheromone increment is larger, otherwise, the pheromone increment is inhibited to a certain extent, the global information is updated by adopting the ant surrounding model, and the pheromone increment and pheromone updating mode is as follows:

the kth ant passes through the cycle (ij)

Where Δ τ is_ijIndicating pheromone increment on the path ij in the current cycle; l is_kThe path length of the kth ant in the circumcircle is shown, and Q is the information intensity and represents the total amount of pheromone released on the path by the ant in the circumcircle.

9. The emblem positioning method of claim 8, characterized in that: the optimized focus transfer path is formed by numbering attention focuses from 1 to n; initializing parameters, randomly placing m ants on n vertexes, tau_ij(t) initialization to τ_ij(0) C, initialization pheromone increment Δ τ_ij(t)，N_c←0(N_cIteration times), setting an upper limit of the iteration times; tabu with ant k_kPlacing the initial starting point of each ant in a current taboo list; calculating the probability of transferring each ant to the next target j, moving the ants, and adding the target point j into a taboo list; calculating the path length L of each ant_k(k 1, 2, 3.. m) and pheromone increments on path (i, j), modifying pheromone intensities on path (i, j) according to a pheromone update equation; for each path (i, j), a [ Delta ] tau is set_ijNo. 0, nc ← nc + 1; if N is present_cWhen the preset iteration times are reached or no more optimal solution appears, the loop is exited, otherwise, the taboo table tabu of the ant k is set_kPlacing the initial starting point of each ant in a current taboo list; and outputting the best solution at present, namely the optimal path U.

10. The emblem positioning method of claim 1, characterized in that: the credibility of the car logo area judged as the sub-area is defined as:

wherein the constraint condition is the focus transfer path U optimized by the ant colony, the credibility of the sub-regions is judged one by one according to the focus transfer path U, and when the CRE of a certain region_UDetermining the area as a car logo area and terminating the judgment; wherein P is_CiFor regional complexity, P_EiAs boundary ratio, unity_iThe vehicle logo in the front face image of the vehicle is unity.