CN107330870A - A kind of thick fog minimizing technology accurately estimated based on scene light radiation - Google Patents

Abstract

The invention discloses a method for removing dense fog based on accurate estimation of scene light radiation. The method mainly includes: performing maximum value filtering on three channels of a fog image to obtain an initial value (S1) of scene light radiation; Each channel of the sky image is jointly filtered with the initial scene light radiation of the corresponding channel to obtain an accurate estimate of the scene light radiation (S2); each channel of the foggy sky image is divided by the corresponding scene light radiation component to obtain the scene light radiation Foggy image (S3) affected by radiation attenuation; project the image that eliminates the effect of scene light radiation attenuation to a spherical coordinate system, and cluster the pixels according to the angle, and then use the haze-line method to calculate the transmittance of each pixel ( S4); use the obtained transmittance and fog imaging model to obtain the image after defogging (S5). The brightness of the dense fog image processed by the method of the invention is suitable for observation by human eyes, and the details are clear.

Description

A Fog Removal Method Based on Accurate Estimation of Scene Light Radiation

technical field

The invention relates to an image enhancement method, in particular to a dense fog removal method based on accurate estimation of scene light radiation, and belongs to the technical field of digital image processing.

Background technique

The existence of fog in the image makes the visibility of the image greatly reduced, and people cannot obtain accurate information from it, and then make wrong judgments about the surrounding environment, which may even lead to disasters. In particular, under the weather conditions of dense fog, the visibility is extremely reduced and a large amount of image information is missing, which directly affects the effectiveness of the security monitoring system. Due to the image quality requirements of outdoor monitoring, in-depth research on fog image defogging technology has become an urgent problem to be solved for image clarity.

Image defogging algorithms are mainly divided into two categories: one is based on non-physical models, and the other is based on physical models. The difference between these two methods lies in whether to use the fog imaging model.

The non-physical model-based defogging method does not start with the physical causes of image degradation, but enhances the contrast of the image and corrects the color of the image to improve the quality of the image according to the visual experience. The typical fog image enhancement methods include histogram equalization algorithm, wavelet method and curvelet transform, automatic color equalization algorithm. This type of algorithm cannot achieve defog in the true sense, but can only improve the visual effect of the image to a certain extent, and it is easy to cause incomplete defog and color distortion.

The dehazing method based on the physical model essentially starts from the classic atmospheric scattering model, and obtains the scene reflectance or fog-free image by solving the relevant parameters in the model. At present, the fog image restoration methods based on physical models mainly include methods based on polarization characteristics, methods based on partial differential equations, methods based on depth information, and methods based on prior knowledge or assumptions. These methods consider that the light radiation of the scene is sufficient, and only consider the degradation caused by the scattering of suspended particles on the ground surface, which can better deal with haze images.

Contents of the invention

However, when using the haze processing method to process the dense fog image, the image will appear overall dark or color cast, and the details will be lost. This is because under dense fog conditions, aerosol particles gather on the surface. When the optical thickness of the fog gradually increases, the transmittance of visible light gradually decreases, so that the light radiation energy reaching the ground gradually decreases. In addition, under dense fog conditions, the weather is more complicated, which may be accompanied by thickening of the atmosphere. In the low-altitude area close to the surface, the size of aerosol suspended particles is larger than the wavelength, and their attenuation coefficients for different wavelengths of light are the same. However, as the height increases, the gravity of the earth decreases, and the diameter of the suspended particles gradually decreases. In the atmosphere, there are many suspended particles whose diameter is smaller than the wavelength of visible light, and these particles will undergo Rayleigh scattering. Rayleigh scattering causes light with shorter wavelengths to be scattered and dissipated during propagation, and only light with longer wavelengths passes through the atmosphere to reach the ground. Therefore, the light radiation on the ground surface may appear colorful color rendering under dense fog conditions, and the fog is no longer pure white.

In order to remove the dense fog more effectively, the researchers propose to adjust the image brightness first, and then use the prior knowledge of the dark channel to eliminate the fog. This method can improve the brightness of the processed image, but due to the inaccurate estimation of scene light radiation by the method used, the details of the image will be unclear. In addition, since the image after adjusting the brightness may be eliminated some shadow areas, the image no longer fully conforms to the principle of the dark channel prior. If the prior knowledge of the dark channel color is still used to estimate the parameters related to the scattered light, the image layering will be poor, and the noise will be easily highlighted.

In this context, it is particularly important to study a dense fog image enhancement method that can not only maintain the brightness and details of the enhanced image, but also effectively remove the image fog.

The technical problem to be solved by the present invention is to provide a method for removing dense fog based on accurate estimation of scene light radiation, so as to realize removing dense fog in a single image.

In order to achieve the above object, the present invention provides a method for removing dense fog based on accurate estimation of scene light radiation, comprising the following steps:

(1) Decompose the RGB three color channels of the foggy image I to obtain the three color channel component maps I ^r , I ^g , I ^b , and use the maximum value filter to respectively analyze the three color channel component maps I ^r , I ^g , I ^b perform calculations to obtain the rough estimation map of scene light radiation M ^r , M ^g , M ^b ;

(2) Use the joint edge-preserving filter to calculate the three groups of images I ^r and M ^r , I ^g and M ^g , I ^b and M ^b respectively, and obtain the accurate estimation map of the scene light radiation L ^r , L ^g , L ^b ;

(3) Divide the three color channel component maps I ^r , I ^g , I ^b by the corresponding scene light radiation accurate estimation map L ^r , L ^g , L ^b , and calculate the results of the three color channels Synthesize to obtain a foggy image J that eliminates the influence of scene light radiation attenuation;

(4) Project each pixel value in the foggy image J that eliminates the effect of scene light radiation attenuation on the spherical coordinate system to obtain (r(x,y),θ(x,y),φ(x,y)) Coordinates, where (x, y) represents the coordinates of the pixel point, r(x, y) represents the distance from the point to the origin, that is ||J(x, y)-1||, θ(x, y) and φ(x,y) represent the corresponding latitude and longitude respectively, and then cluster the pixels according to the latitude and longitude to obtain n classified fog lines (P ₁ ,P ₂ ,…,P _n ), and then use the haze line method to obtain each The transmittance of a pixel;

(5) Using the transmittance obtained in step (4), calculate the reflectance of the scene, and then obtain the image after dehazing.

A method for removing dense fog based on accurate estimation of scene light radiation as described above is characterized in that in the step (1), the radius of the maximum value filter is Radius, and the formula used is as follows:

Radius＝|(max(height, width)/100|+1 (1)

Where height and width represent the length and width of the image.

A method for removing dense fog based on accurate estimation of scene light radiation as described above, characterized in that in the step (2), the formula used for joint filtering is as follows:

Where (x, y), (i, j) is the coordinate information, Ω is the block centered on (x, y), and M ^c is the rough estimation map of the scene light radiation obtained after the maximum value filtering of each channel .

η ^c (i,j) is a step function:

is an exponential function:

Among them, I ^c (x, y) represents the c channel of the fog image, and σ represents the variance of pixel values.

A method for removing dense fog based on accurate estimation of scene light radiation as described above is characterized in that in step (3), in order to avoid the situation that the obtained ambient light is equal to 0, when dividing by the ambient light, it is necessary to divide Add a very small constant ε=0.01 to the ambient light components obtained in step (2).

The method for removing dense fog based on accurate estimation of scene light radiation as described above is characterized in that in the step (4), the k-mean algorithm is used to cluster the pixels projected to the spherical coordinate system.

The method for removing dense fog based on accurate estimation of scene light radiation as described above is characterized in that the number of cluster centers used in the clustering method is set within a range of 200-500.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

FIG. 1 is a flowchart of a method for removing dense fog based on accurate estimation of scene light radiation according to the present invention.

FIG. 2( a ) to FIG. 2( c ) are diagrams showing the 3-D effect of the joint filtering method of the present invention.

Fig. 3 (a) to Fig. 3 (f) are according to the defogging method of the present invention and the experimental result comparison of existing typical defogging method on the test image; Wherein, Fig. 3 (a) represents foggy day image, Fig. 3 ( b) shows the result of haze removal by the adaptive histogram equalization method with limited contrast, Fig. 3(c) shows the result of Tarel method after haze removal, Fig. 3(d) shows the result of combined dark channel and guided filter method after haze removal, Fig. 3(e) shows the result after combined dark primary color and post-processing method, and Fig. 3(f) shows the image after the method of the present invention.

detailed description

The dense fog removal method based on the accurate estimation of scene light radiation provided by the present invention first estimates scene light radiation through an edge-preserving filter to eliminate the influence of scene light radiation attenuation. Then, the reflectance of the scene is calculated by using the properties of the pixel spherical coordinate system, and the image after dehazing is obtained. This is described in detail below.

In the present invention, the dense fog removal method based on accurate estimation of scene light radiation includes the following steps, as shown in Figure 1:

Step 1. Decompose the r, g, and b three color channels of the foggy image I to obtain three color channel component maps I ^r , I ^g , I ^b , and use the maximum value filter to respectively analyze the three color channel component maps I ^r , I ^g , I ^b are calculated to obtain a rough estimation map of the scene light radiation M ^r , M ^g , M ^b ;

The estimate of scene light radiation that maintains naturalness needs to meet two conditions: (a) light radiation should be smooth in most areas but should retain alternating edges of light and dark; (b) scene light radiation should not be smaller than reflected light, so as to ensure dehazing The resulting image retains as much scene detail as possible.

In Retinex theory, many center-surround methods obtain scene radiance by low-pass filtering the maximum channel of the image. However, this method is not suitable for the case where the scene light radiation is color cast. Moreover, the maximum channel is only a lower limit of the scene light radiation, and using it as an initial estimate of lighting lacks a physical interpretation.

In the previous literature, based on the assumption that the high-brightness area in the image is a white surface or a high-brightness point of the light source, the researchers proposed the Max-RGB algorithm, which used the maximum value of the three channels as an estimate of the illumination. However, this method is not suitable for uneven lighting conditions. In order to make the estimation method robust, the present invention extends the Max-RGB method to local areas. In other words, the present invention considers that the light reflected by objects with higher reflectance in each local area is closer to the ambient light. Assuming that I is the observed image, the rough estimate of the scene light radiation is:

where Ω represents a local window centered at (x, y), and c represents a color channel.

Step 2. Use the joint edge-preserving filtering method to calculate the three groups of images I ^r and M ^r , I ^g and M ^g , I ^b and M ^b respectively, and obtain the accurate estimation map L ^r , L ^g , L of the scene light radiation ^b ;

The method in step 1 can indeed better approximate the scene light radiation. However, similar to other methods that assume local invariance, the lighting estimated by this method will have block effects at the intersection of light and dark in the scene, so it needs to be further optimized. For this reason, the present invention designs a content-adaptive joint edge-preserving filter to estimate accurate scene light radiation L ^c (x, y):

Represents an exponential function, which is used to control the optical radiation to satisfy the condition (a):

Among them, η ^c (i, j) is a step function, which is used to control the scene light radiation to meet the condition (b), defined as follows:

Where σ is the pixel gray level difference.

Unlike other joint edge-preserving filters, this method does not assign the neighborhood relationship of the guide map to the image to be filtered for smoothing. When each window is filtered, only one pixel in the guide map serves as a guide.

Step 3. Divide the three color channel component maps I ^r , I ^g , and I ^b by the corresponding scene light radiation accurate estimation map L ^r , L ^g , L ^b , and calculate the results of the three color channels Synthesize to obtain a foggy image J that eliminates the influence of scene light radiation attenuation;

Using the accurate estimation of scene light radiation, the effect of eliminating scene light radiation attenuation can be obtained by the following formula:

J ^c (x, y) = I ^c (x, y)/L ^c (x, y) (5)

Step 4. Project each pixel value in the foggy image J that eliminates the influence of scene light radiation attenuation on a spherical coordinate system to obtain (r(x,y), θ(x,y), φ(x,y)) Coordinates, where (x, y) represents the coordinates of the pixel point, r(x, y) represents the distance from the point to the origin, that is ||J(x, y)-1||, θ(x, y) and φ(x,y) represent the corresponding latitude and longitude respectively, and then cluster the pixels according to the latitude and longitude to obtain n classified fog lines (P ₁ ,P ₂ ,…,P _n ), and then use the haze line method to obtain each The transmittance of a pixel;

After eliminating the attenuation effect of the scene light radiation, the acquisition of the scene reflectance is transformed into the problem of obtaining the transmittance from the following formula:

J(x,y)=R(x,y)t(x,y)+1-t(x,y) (6)

After transforming the above formula:

J(x,y)-1=(R(x,y)-1)t(x,y) (7)

Express J(x,y)-1 in the spherical coordinate system as:

J(x,y)-1=[r(x,y),θ(x,y),φ(x,y)] (8)

Among them, r represents the distance from the point to the origin, that is, ||J(x,y)-1||, θ and φ represent the latitude and longitude respectively. For the pixel values of the same color in the foggy image, there are different t, in the spherical coordinate system, these pixels have different radii from the origin, but have the same latitude and longitude θ and φ. Based on this principle, through the clustering method, after setting the number of cluster centers, several lines can be obtained, and the lines contain pixel values of similar colors.

Then, the transmittance of the farthest pixel in each cluster is recorded as 1, and the transmittance of other pixels distributed on each line is calculated according to the ratio of the radius of the projected spherical coordinate to the radius of the projection coordinate of the farthest pixel;

Given a point r(x,y) can be expressed as:

r(x,y)=t(x,y)||J(x,y)-1||,0≤t(x,y)≤1 (9)

the longest radius The coordinates are set to t=1. in, P is a clustering line. Then the transmittance of each point can be obtained:

Step 5. Use the transmittance obtained in step (4) to calculate the reflectance of the scene, and then obtain the image after defogging.

The reflectance of the scene can be calculated as the image after dehazing by the following formula:

FIG. 2( a ) to FIG. 2( c ) are diagrams showing the 3-D effect of the joint filtering method of the present invention.

Fig. 3 (a) to Fig. 3 (f) are according to the defogging method of the present invention and the experimental result comparison of existing typical defogging method on the test image; Wherein, Fig. 3 (a) represents foggy day image, Fig. 3 ( b) shows the result of haze removal by the adaptive histogram equalization method with limited contrast, Fig. 3(c) shows the result of Tarel method after haze removal, Fig. 3(d) shows the result of combined dark channel and guided filter method after haze removal, Fig. 3(e) shows the result after combined dark primary color and post-processing method, and Fig. 3(f) shows the image after the method of the present invention.

It should be noted that what is disclosed above are only specific implementation examples of the present invention. According to the technical idea provided by the present invention, the changes that those skilled in the art can think of should fall within the protection scope of the present invention.

Claims (6)

1. A method for removing dense fog based on accurate estimation of scene light radiation, comprising the steps of: (1) Decompose the RGB three color channels of the foggy image I to obtain the three color channel component maps I ^r , I ^g , I ^b , and use the maximum value filter to respectively analyze the three color channel component maps I ^r , I ^g , I ^b perform calculations to obtain the rough estimation map of scene light radiation M ^r , M ^g , M ^b ; (2) Use the joint edge-preserving filter to calculate the three groups of images I ^r and M ^r , I ^g and M ^g , I ^b and M ^b respectively, and obtain the accurate estimation map of the scene light radiation L ^r , L ^g , L ^b ; (3) Divide the three color channel component maps I ^r , I ^g , and I ^b by the corresponding scene light radiation accurate estimation map L ^r , L ^g , L ^b , and calculate the three color channel components The results are synthesized to obtain a foggy image J that eliminates the influence of scene light radiation attenuation; (4) Project each pixel value in the foggy image J that eliminates the effect of scene light radiation attenuation on the spherical coordinate system to obtain (r(x,y),θ(x,y),φ(x,y)) Coordinates, where (x, y) represents the coordinates of the pixel point, r(x, y) represents the distance from the point to the origin, that is ||J(x, y)-1||, θ(x, y) and φ(x,y) represent the corresponding latitude and longitude respectively, and then cluster the pixels according to the latitude and longitude to obtain n classified fog lines (P ₁ ,P ₂ ,…,P _n ), and then use the haze line method to obtain each The transmittance of a pixel; (5) Using the transmittance obtained in step (4), calculate the reflectance of the scene, and then obtain the image after dehazing. 2. The dense fog removal method as claimed in claim 1, characterized in that: In described step (1), the radius Radius of maximum value filtering is determined by following formula: Radius＝|(max(height, width)/100|+1 (1) Where height and width represent the length and width of the image. 3. The dense fog removal method as claimed in claim 1, characterized in that: In the step (2), the joint edge-preserving filter is represented by the following formula: Where (x, y), (i, j) is the coordinate information, Ω is the block centered on (x, y), and M ^c is the rough estimation map of the scene light radiation obtained after the maximum value filtering of each channel ; η ^c (i,j) is a step function: <mrow><msup><mi>&amp;eta;</mi><mi>c</mi></msup><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mo>=</mo><mfenced open = "{" close = ""><mtable><mtr><mtd><mn>1</mn></mtd><mtd><mrow><msup><mi>M</mi><mi>c</mi></msup><mrow><mo>(</mo><mi>i</mi><mo>,</mo><mi>j</mi><mo>)</mo></mrow><mo>-</mo><msup><mi>I</mi><mi>c</mi></msup><mrow><mo>(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo>)</mo></mrow><mo>&amp;GreaterEqual;</mo><mn>0</mn></mrow></mtd></mtr><mtr><mtd><mn>0</mn></mtd><mtd><mrow><mi>e</mi><mi>l</mi><mi>s</mi><mi>e</mi></mrow></mtd></mtr></mtable></mfenced><mo>,</mo><mi>c</mi><mo>&amp;Element;</mo><mo>{</mo><mi>r</mi><mo>,</mo><mi>g</mi><mo>,</mo><mi>b</mi><mo>}</mo><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>3</mn><mo>)</mo></mrow></mrow> is an exponential function: Among them, I ^c (x, y) represents the c channel of the fog image, and σ represents the variance of pixel values. 4. The dense fog removal method as claimed in claim 1, characterized in that: In the step (3), in order to avoid the situation that the ambient light of the obtained scene light radiation accurate estimation map L ^r , L ^g , L ^b is equal to 0, when dividing by the ambient light, step (2) A very small constant ε=0.01 is added to the ambient light components of the scene light radiation accurate estimation maps L ^r , L ^g , and L ^b obtained in . 5. The dense fog removal method as claimed in claim 1, characterized in that: In the step (4), the k-mean algorithm is used to cluster the pixels projected to the spherical coordinate system. 6. A method for removing dense fog based on accurate estimation of scene light radiation as claimed in claim 5, characterized in that: The range of the number of cluster centers used by the clustering method is set between 200-500.