A fire scene low-quality image restoration method for rescue robots

Abstract

The present invention proposes a method for restoring low-quality images of fire scenes for rescue robots. The method of the present invention segments the flame area based on the regional threshold segmentation algorithm to avoid the influence of the flame light source on the global atmospheric light estimation; designs an atmospheric light detection operator to obtain accurate atmospheric light parameters based on superpixel blocks, solving the problem of global atmospheric light estimation distortion; constructs a transmittance estimation optimization module, realizes the refinement of transmittance based on the bilateral weighted guided filtering method, and solves the problems such as halo caused by traditional methods. The method proposed by the present invention helps to improve the clarity of fire scene image restoration and provides a data basis for improving the on-site detection, identification, environmental mapping and path planning of rescue robots.

Description

Technical Field

[0001] The invention belongs to the field of rescue robot image processing, and in particular relates to a fire scene low-quality image restoration method for a rescue robot. Background Art

[0002] In a fire rescue environment, combustion is usually accompanied by uneven flames and smoke. When rescue robots perform tasks in such scenarios, the collected images are usually affected by environmental factors such as light, flames, and fog.

[0003] The existing methods for processing fire scene images have estimation distortion problems due to the uneven distribution of ambient light and uneven fog density. Through the study of the method of sharpening low-quality images in the rescue fire environment, the low-quality degraded images are restored to high-quality images, providing a data basis for the rescue robot's on-site detection, recognition, environmental mapping and path planning. Summary of the invention

[0004] In order to solve the problem in the background technology, the present invention proposes a low-quality image restoration method for a fire scene for a rescue robot.

[0005] The technical solution adopted by the present invention is as follows:

[0006] 1. A fire scene low-quality image restoration system for rescue robots

[0007] It includes a regional atmospheric light estimation module, a transmittance estimation optimization module and an image reconstruction and restoration module, wherein the regional atmospheric light estimation module includes a fire light regional atmospheric light estimation module and a global atmospheric light estimation module;

[0008] The fire area atmospheric light estimation module is used to divide the original image into a fire area image and a non-fire area image;

[0009] The global atmospheric light estimation module is used to perform superpixel segmentation on the non-fire area image and design an atmospheric light detection operator to obtain the estimated value of the global atmospheric light;

[0010] Transmittance estimation optimization module, used to estimate the rough transmittance and calculate the precise transmittance using the bilateral weighted guided filtering method;

[0011] The image reconstruction and restoration module is used to reconstruct and restore a clear image.

[0012] 2. Low-quality image restoration method for fire scenes for rescue robots using the above system

[0013] Step 1: Use the fire area atmospheric light estimation module to segment the original image I into the fire area image I F and non-fire area image INF ;

[0014] Step 2: The global atmospheric light estimation module calculates the global atmospheric light estimation value A0 by constructing an atmospheric light detection operator;

[0015] Step 3: Input the dark channel image and the global atmospheric light estimation value into the transmittance estimation optimization module to estimate the rough transmittance, and use the bilateral weighted filtering transmittance optimization to obtain the accurate transmittance t;

[0016] Step 4: Restore the final clear image through the image reconstruction and restoration module.

[0017] The step 1) is specifically as follows:

[0018] 1.1) Segment the original image I by combining the RGB color space criterion and the HLS color space criterion to obtain the initial fire area image I1;

[0019] 1.2) The initial fire area image I1 is subjected to morphological processing of opening first and then closing, the isolated points in the image are deleted and the holes inside the area image I1 are filled. The morphologically processed image is further subjected to Gaussian filtering to obtain the final fire area image I1. F ; and according to the original image I and the fire area image I F Obtain non-fire area image I NF .

[0020] The step 2) is specifically as follows:

[0021] 2.1) Perform minimum filtering on the R, G, and B channels of the original image I to obtain the dark channel image I of the image d ;

[0022] 2.2) For the original image I, construct the atmospheric light detection operator score:

[0023] score=(1-S)I d

[0024] Where S is the saturation component of the original image I;

[0025] 2.3) For the non-fire area image I NF Perform superpixel segmentation to obtain segmentation map I s , calculate the segmentation map I s (s i ∈I s ) in each superpixel block s i The atmospheric light detection operator score

[0026]

[0027] in, is a superpixel block s i The number of pixels in the , x is the super pixel block s i Pixels in ;

[0028] 2.4) The corresponding The values ​​are sorted in descending order, and the superpixel block with the largest atmospheric light detection operator score is selected, denoted as s max , calculate superpixel block s max The average pixel value of all pixels in the image is used to obtain the estimated value of global atmospheric light A0:

[0029]

[0030] in, is a superpixel block s max The number of pixels in the image, I(x) is the pixel block s max The pixel value of pixel x in .

[0031] The step 3) is specifically as follows:

[0032] 3.1) Calculate the adaptive confidence t through the L channel of the original image I1 * (x):

[0033]

[0034] Where Ω is the minimum filter interval, p is the confidence adjustment parameter, and L(y) is the L channel value of pixel y in region Ω;

[0035] 2) According to the dark channel image I of the original image I1 d The rough transmittance feature map t0 of the image is calculated using the global atmospheric light estimation value A0. The calculation formula is as follows:

[0036]

[0037] 3) Use bilateral weighted guided filtering to optimize t0 and obtain the image transmittance t. The calculation formula is as follows:

[0038] t=a*I g +b

[0039] Among them, I g is the grayscale image of the original image I, a and b are the guided filter coefficients, and the calculation formula is as follows:

[0040]

[0041]

[0042] Among them, ε is the tolerance factor, d is the filter interval, ω(i,j,k,l) ​​is the filter weight coefficient, (k,l) represents the center coordinates of the filter window, (i,j) represents the other coordinates of the window, ω m is the kernel function, m takes the value of {1,2,3,4}, and the calculation formula is as follows:

[0043]

[0044] Among them, σ d is the spatial weight, σ r is the range weight;

[0045] Among them, I m (i,j),I m (k,l), m∈{1,2,3,4} represents the pixel value of the corresponding image at point (i,j), (k,l), and the calculation formula is as follows:

[0046]

[0047] I2=t0

[0048] I3=I g

[0049]

[0050] in, represents the matrix Hadamard product.

[0051] In step 4), the image restoration module outputs the final clear image J, which is expressed as:

[0052]

[0053] Beneficial effects of the present invention:

[0054] The low-quality image restoration method for fire scenes for rescue robots proposed in the present invention segments the flame area through a regional threshold segmentation algorithm, avoids the distortion effect of the flame light source on the global atmospheric light estimation, designs an atmospheric light detection operator, obtains accurate atmospheric light parameters based on superpixel blocks, and solves the problem of global atmospheric light estimation distortion; constructs a transmittance estimation optimization module, realizes the refinement of transmittance based on the bilateral weighted guided filtering method, and solves the problems such as halo caused by the general method. The method proposed in the present invention helps to improve the clarity of fire scene image restoration and provides a data basis for improving the on-site detection, identification, environmental mapping and path planning of rescue robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A flow chart for clarifying images of fire scene for rescue;

[0056] Figure 2 is the original image in the embodiment of the present invention;

[0057] Figure 3 It is the fire area segmentation module and the fire area map after binarization processing;

[0058] Figure 4 is a rough transmittance estimation diagram;

[0059] Figure 5 It is the refined transmittance estimation map after bilateral weighted guided filtering. DETAILED DESCRIPTION

[0060] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0061] 1. The present invention includes a fire scene rescue image clarity system, including a regional atmospheric light estimation module, a transmittance estimation optimization module and an image reconstruction and restoration module. The regional atmospheric light estimation module includes a fire light regional atmospheric light estimation module and a global atmospheric light estimation module.

[0062] 2. If Figure 1 As shown, the input original image I estimates the atmospheric light through the regional atmospheric light estimation module, designs the regional threshold segmentation algorithm to segment the flame area, further segments the image superpixel and designs the atmospheric light detection operator to obtain the global atmospheric light; constructs the transmittance estimation optimization module to estimate the rough transmittance, and further calculates the refined transmittance based on the bilateral weighted guided filtering method; and inputs the image restoration module to restore the high-quality image.

[0063] 3. Fire area atmospheric light estimation module Figure 2 The original image I shown is segmented into the fire area image I F and non-fire area image I NF .

[0064] Specifically include:

[0065] Step 1: Combine the original image I with the RGB color space criterion and the HLS color space criterion to segment the fire area and obtain the regional image I1.

[0066] The RGB color space criterion is:

[0067] R>R T

[0068] R≥G≥B

[0069] The HLS color space criterion is:

[0070]

[0071] L min ≤L≤Lmax

[0072] R, G, and B are the red, green, and blue components of the image RGB color space, respectively. T is the red threshold, S and L are the saturation and brightness components of the image HLS color space, respectively. min is the minimum brightness threshold, L max is the maximum brightness threshold.

[0073] Step 2: Perform morphological processing of opening first and then closing on the regional image I1, delete the isolated points in the image and fill the holes inside the target area, and further perform Gaussian filtering on the result of image morphological processing to obtain Figure 3 The fire area image I shown F , non-fire area image I NF .

[0074] 4. The global atmospheric light estimation module calculates the global atmospheric light value A0 by constructing an atmospheric light detection operator.

[0075] Specifically include:

[0076] Step 1: Perform minimum filtering on the R, G, and B channels of the input image I to obtain the dark channel image I d ;

[0077] Step 2: Input image I and design the atmospheric light detection operator score:

[0078] score=(1-S)I d

[0079] Where S is the saturation component of image I.

[0080] Step 3: Non-fire area image I NF Superpixel segmentation, get segmentation map I s , calculate each s i ∈I s Atmospheric light detection operator score for superpixel patches

[0081]

[0082] in, is a superpixel block s i The number of pixels.

[0083] Step 4: The values ​​are sorted in descending order, and the superpixel block s with the largest score value is selected. max The estimated value A0 of the maximum atmospheric light value is calculated by taking the average value of all pixels in the superpixel.

[0084]

[0085] in, is a superpixel block s max The number of pixels.

[0086] 5. Transmittance estimation and optimization module, which inputs the dark channel map and the global atmospheric light estimation value to estimate the rough transmittance, and optimizes the precise transmittance t by bilateral weighted guided filtering transmittance.

[0087] Specifically include:

[0088] Step 1: Calculate the adaptive confidence t through the L channel of the image * (x):

[0089]

[0090] Among them, Ω is the minimum filter interval, and p is the confidence adjustment parameter:

[0091] Step 2: If Figure 4 As shown, the rough transmittance feature map t0 of the image is calculated, and the calculation formula is as follows:

[0092]

[0093] Step 3: If Figure 5 As shown, bilateral weighted guided filtering is used to optimize t0 and estimate the accurate transmittance t of the image. The calculation formula is as follows:

[0094] t=a*I g +b

[0095] a, b are the bilateral weighted guidance coefficients, and the calculation formula is as follows:

[0096]

[0097]

[0098] Among them, ε is the tolerance factor, d is the filter interval, ω(i,j,k,l) ​​is the filter weight coefficient, (k,l) represents the window center coordinates, (i,j) represents the other coordinates of the window, ω m is the kernel function, m takes the value of {1,2,3,4}, and the calculation formula is as follows:

[0099]

[0100] Among them, σ d Spatial weight, σ r is the value range weight.

[0101] Among them, I m (i,j),Im (k,l) represents the pixel value of the corresponding image point, and the calculation formula is as follows:

[0102]

[0103] I2=t0

[0104] I3=Ig

[0105]

[0106] represents the matrix Hadamard product.

[0107] 6. The image restoration module outputs the final clear image J, which is expressed as:

[0108]

[0109] Compared with other methods, the advantages of the present invention are high-quality restoration of low-quality images of fire scenes while occupying very little computing resources, and can be effectively applied to image pre-processing work such as fire scene detection and identification by rescue robots.

Claims

1. A fire scene low-quality image restoration method for a rescue robot, characterized in that: Step 1: Use the fire area atmospheric light estimation module to segment the original image I into the fire area image I F and non-fire area image I NF ; Step 2: The global atmospheric light estimation module calculates the global atmospheric light estimation value A0 by constructing an atmospheric light detection operator; Step 3: Input the dark channel image and the global atmospheric light estimation value into the transmittance estimation optimization module to estimate the rough transmittance, and use the bilateral weighted guided filtering transmittance optimization to obtain the precise transmittance t; Step 4: Restore the final clear image through the image reconstruction and restoration module; The step 2) is specifically as follows: 2.1) Perform minimum filtering on the R, G, and B channels of the original image I to obtain the dark channel image I of the image d ; 2.2) For the original image I, construct the atmospheric light detection operator score: score=(1-S)I d Where S is the saturation component of the original image I; 2.3) For the non-fire area image I NF Perform superpixel segmentation to obtain segmentation map I s , calculate the segmentation map I s Each superpixel block s i The atmospheric light detection operator score in, is a superpixel block s i The number of pixels in the , x is the super pixel block s i Pixels in ; 2.4) The corresponding The values ​​are sorted in descending order, and the superpixel block with the largest atmospheric light detection operator score is selected, denoted as s max , calculate superpixel block s max The average pixel value of all pixels in the image is used to obtain the global atmospheric light estimation value A0: in, is a superpixel block s max The number of pixels in the image, I(x) is the pixel block s max The pixel value of the pixel point x in ; The step 3) is specifically as follows: 3.1) Calculate the adaptive confidence t through the L channel of the original image I * (x): Where Ω is the minimum filter interval, p is the confidence adjustment parameter, and L(y) is the L channel value of pixel y in region Ω; 3.2) According to the dark channel image I of the original image I d The rough transmittance feature map t0 of the image is calculated using the global atmospheric light estimation value A0. The calculation formula is as follows: 3.3) Use bilateral weighted guided filtering to optimize t0 and obtain the image transmittance t. The calculation formula is as follows: t=a*I g +b Among them, I g is the grayscale image of the original image I, a and b are the guided filter coefficients, and the calculation formula is as follows: ε is the tolerance factor, d is the filter interval, ω m (i,j,k,l) ​​is the filter weight coefficient, (j,l) represents the center coordinate of the filter window, (i,j) represents the other coordinates of the window, and the filter weight coefficient is calculated by the kernel function ω m It is calculated that the value of m is {1,2,3,4}, and the calculation formula is as follows: σ d is the spatial weight, σ r is the range weight; I m (i,j),I m (k,l), m∈{1,2,3,4} represents the pixel value of the corresponding image at point (i,j), (k,l), and the calculation formula is as follows: I2=t0 I3=I g represents the matrix Hadamard product; In step 4), the image restoration module outputs the final clear image J, which is expressed as:

2. The method for restoring low-quality images of fire scenes for rescue robots according to claim 1, characterized in that: The step 1) is specifically as follows: 1.1) Segment the original image I by combining the RGB color space criterion and the HLS color space criterion to obtain the initial fire area image I1; 1.2) The initial fire area image I1 is subjected to morphological processing of opening first and closing later, the isolated points in the image are deleted and the holes inside the initial fire area image I1 are filled. The image after morphological processing is further subjected to Gaussian filtering to obtain the final fire area image I1. F ; and according to the original image I and the fire area image I F Obtain non-fire area image I NF .

Images