A cutterhead image detection and correction method under uneven illumination in a shield machine

Abstract

The present invention relates to a method for detecting and correcting cutterhead images under uneven illumination within a shield machine. The method comprises: first, inputting an unevenly illuminated image under low illumination, converting the image from RGB to HSV space, and performing continuous convolution operations on the extracted illumination components using a small-scale Gaussian convolution kernel to obtain the illumination component distribution of the scene; second, determining the gamma value of each pixel position in combination with the local contrast and gradient magnitude, dynamically adjusting the gamma value according to the local features of the image, and converting the corrected image from HSV to RGB space; finally, performing a weighted adaptive histogram equalization algorithm on the image, performing guided filtering to enhance the image, and outputting the image. The method has a significant effect on correcting cutterhead images in unevenly illuminated areas within the dim space of a shield machine, improving local contrast, taking into account both the global characteristics and local detail information of the cutterhead image, suppressing the influence of noise, and making the cutterhead image enhancement result clearer and more accurate.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent equipment manufacturing and detection, in particular to computer vision and image processing intelligent system technology, and relates to an image correction algorithm for uneven illumination, and is particularly suitable for a cutterhead image correction method under uneven illumination in a shield machine. Background Art

[0002] With the development of intelligent equipment manufacturing and computer vision technologies, the design, manufacturing, and operation of shield machines have been greatly improved. Intelligence, automation, and digitization have become the main trends in the development of shield machines. Shield machines utilize intelligent design, using sensors, control systems, and human-machine interfaces to achieve intelligent monitoring, control, and management. Computer vision technology is used to accurately identify geological conditions, obstacles, and the working environment through image processing and analysis. As a representative example of intelligent equipment manufacturing and testing, shield machines integrate key technologies from computer vision, safety and environmental testing services, and mechanical equipment, providing strong technical support and assurance for underground engineering construction.

[0003] With the advancement of underground tunnel construction technology, shield machines (TBMs) have been widely used in the construction of urban rail transit, highways, railways, subways, and tunnels in my country. They offer advantages such as high excavation speed, short construction periods, minimal ecological damage, safe excavation, and high excavation efficiency. However, due to the unique working environment, shield machines often lack sufficient natural light, resulting in dim interior spaces and limited image acquisition. Furthermore, the shape and position of equipment such as the cutterhead can cast shadows, further reducing image quality and usability. This results in low image brightness and makes it difficult to accurately locate worn cutterheads. Clear images not only help operators accurately locate the cutterhead and related equipment, ensuring precise operation during cutter changes, but also reduce operational errors and time wasted due to blurry or inaccurate images. Therefore, when capturing images of the cutterhead inside a shield machine, it is crucial to ensure clear and accurate image data.

[0004] Currently, the more commonly used image enhancement methods can be roughly divided into two categories: spatial domain-based methods and frequency domain-based methods. Spatial domain methods primarily process image pixels in the time domain. By mapping pixel points, the altered pixel values replace the original pixels, thereby achieving the goal of image improvement. Common spatial domain enhancement algorithms include histogram grayscale transformation, histogram correction, f-segment method, contrast-limited adaptive method, and local sub-block method. Frequency domain methods, on the other hand, are based on altering the image's variation domain. After selecting parameters, the inverse operation is performed back to the spatial domain to obtain the improved image. Common frequency domain enhancement algorithms include FFT and DCT. In summary, eliminating the effects of uneven illumination on images has become a research hotspot in the current field.

[0005] (1) For example, the document "Adaptive Correction Algorithm for Uneven Illumination Images Based on Two-Dimensional Gamma Function" proposes an adaptive brightness correction method based on a two-dimensional gamma function, which uses the distribution characteristics of the illumination components of the image to adaptively adjust the parameters of the two-dimensional gamma function to achieve the purpose of improving the overall quality of the image with uneven illumination. The algorithm proposed in this invention takes into account the local contrast and gradient size to better retain the detail information in the image, and dynamically adjusts the gamma value according to the characteristics of the local image. Therefore, it is more suitable for coping with illumination changes in different areas, and can more accurately control the degree of enhancement when adjusting the brightness of the image to avoid excessive enhancement or reduction;

[0006] (2) The paper "Adaptive Enhancement Algorithm for Images with Uneven Illumination" combines the advantages of gamma correction and Top-Hat transform, ensuring that texture information remains unchanged while improving image contrast and taking into account both global image characteristics and local detail information. The algorithm proposed in this paper uses weighted adaptive histogram equalization based on gradient magnitude to make the pixel value distribution of the image more uniform while preserving image details, thereby improving image contrast and brightness.

[0007] The present invention takes into account local contrast and gradient size, dynamically adjusts the gamma value according to the characteristics of the local image, and adopts a weighted adaptive histogram equalization method that takes into account the gradient size, taking into account both the global characteristics and local detail information of the image. It has a significant effect on correcting the cutterhead image in areas with uneven lighting inside the shield machine, and the enhancement result is clearer and more accurate. Summary of the Invention

[0008] The purpose of the present invention is to provide a cutterhead image correction method suitable for uneven illumination in a shield machine. Compared with the ordinary two-dimensional gamma transform, the method takes into account the local contrast and gradient size, can better retain the detail information in the image, and can more accurately control the degree of enhancement when adjusting the brightness of the image to avoid excessive enhancement or reduction; weighted adaptive histogram equalization taking into account the gradient size can make the pixel value distribution of the image more uniform while retaining the image details, improve the contrast and brightness of the image, reduce the impact of uneven illumination on the image, make the image more uniform and clear, and make adjustments according to the local features of the image. Therefore, images in different areas can be enhanced in a targeted manner, and the adaptability is stronger.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] S1: Input image and extract scene illumination components;

[0011] S1.1: Convert the original image to the HSV color space and extract the luminance channel as the illumination component of the scene. The luminance value F(x,y) is the result of the combined effect of the illumination component R(x,y) that illuminates the scene and the reflection component S(x,y) from the object surface:

[0012] F(x,y)=R(x,y)×S(x,y) (1)

[0013] Where: R(x,y) is the component of light irradiating the scene, S(x,y) is the component reflected from the object surface;

[0014] S1.2: Use a small-scale Gaussian convolution kernel to perform continuous convolution operations on the extracted illumination components.

[0015] With the help of small-scale Gaussian convolution kernel, the distribution of illumination components of the scene is obtained;

[0016] S2: Calculate local contrast and gradient information for the illumination component image after continuous convolution;

[0017] S2.1: Calculate the horizontal and vertical gradient responses by convolving the Sobel operator with the image, combine the horizontal and vertical gradient responses, and calculate the gradient magnitude at each pixel position:

[0018]

[0019] Where: G(x,y) is the gradient size of each pixel position, G x (x,y) and G y (x, y) are the Sobel operator element values in the horizontal and vertical directions respectively;

[0020] S2.2: Illumination component image I after continuous convolution of original illumination component image I c Apply a Gaussian blur to smooth the image and reduce the effects of noise, expressed as a Gaussian blur at position (x,y):

[0021]

[0022] Where: I b (x,y) is the pixel value at position (x,y) after Gaussian blurring, k is the radius of the Gaussian kernel, σ is the standard deviation of the Gaussian kernel, and I c (x+i,y+j) represents the illumination component image I after continuous convolution c The pixel value at position (x+i,y+j), represents the weight of the Gaussian kernel;

[0023] S2.3: For each pixel position (x, y), calculate the standard deviation of the pixel values in the surrounding n×n square area R as an estimate of the local contrast. The local contrast C(x, y) is calculated as follows:

[0024]

[0025] Where: C(x,y) is the local contrast, μ is the average value of the pixel value in the R area, I b (x+i,y+j) is the pixel value at position (x+i,y+j) in the image;

[0026] S2.4: Determine the gamma value for each pixel position based on the local contrast and gradient magnitude. Dynamically adjust the gamma value based on the local features of the image. The formula is as follows:

[0027]

[0028] r=α·C(x,y)+β·G(x,y) (6)

[0029] Where: I (x,y) is the input image, I′ (x,y) is the rectified image, r is the exponent value used for brightness enhancement, C(x,y) is the local contrast, G(x,y) is the gradient size at each pixel position, α and β are adjustment coefficients used to balance the effects of local contrast and gradient size;

[0030] S2.5: After the brightness is corrected, convert it back to RGB space;

[0031] S3: Perform weighted adaptive histogram equalization on the image;

[0032] S3.1: Gradient information calculation weight w(x,y):

[0033] w(x,y)=e c·G(x,y) (7)

[0034] Where: G(x,y) is the gradient size at each pixel position, c is a constant used to control the degree of influence of the gradient size on the weight;

[0035] S3.2: Segment the image. Divide the original image into local regions. The image is divided into continuous, non-overlapping m×n sub-blocks. The values of m and n can be 4, 6, 8, etc. Each sub-block contains N pixels. The larger the sub-block, the more obvious the enhancement effect, but the more image detail information is lost.

[0036] S3.3: Perform local histogram equalization. For each pixel value i in the local area, calculate the number of pixels with pixel value i in the area to form the grayscale histogram H(i) of the local area.

[0037] S3.4: Calculate its cumulative distribution function CDF(i);

[0038] S3.5: For each pixel (x, y) in the local area, map its original grayscale value I(x, y) to the equalized grayscale value I e (x,y), the grayscale value formula after equalization is as follows:

[0039]

[0040] Where: CDF(I(x,y)) represents the value of the cumulative distribution function CDF of the original gray value I(x,y) in the local area, N is the total number of pixels in the local area, L is the gray level of the pixel, is the normalization factor, which is used to ensure the correct grayscale value range after equalization;

[0041] S3.6: For each pixel position (x, y), the grayscale value I after local histogram equalization is e (x,y) is weighted with the corresponding weight w(x,y);

[0042] I f (x,y)=w(x,y)×I e (x,y) (9)

[0043] Where: I e (x,y) is the grayscale value after local histogram equalization, and w(x,y) is the corresponding weight;

[0044] S3.7: Perform guided filtering on the image to obtain an enhanced image;

[0045] S3.8: Output the enhanced image.

[0046] The present invention has the following beneficial effects:

[0047] (1) The present invention discloses a cutterhead image correction method suitable for use in shield machines under uneven illumination. Compared with conventional two-dimensional gamma transforms, this method takes into account local contrast and gradient size to better preserve detailed information in the image. Furthermore, when adjusting the brightness of the image, it can more accurately control the degree of enhancement to avoid over-enhancement or reduction. Information entropy is used to measure the amount of information contained in an image. The greater the information entropy, the more information the image carries. The entropy of this algorithm is significantly improved compared to other algorithms, and is 9.44% higher than that of the MSR algorithm. The images processed by this algorithm also contain more information, and more details and features can be extracted. The average gradient value of this algorithm is also much higher than that of other algorithms, especially the highest GT algorithm. In comparison, the algorithm of the present invention is improved by 45.08%, which can significantly improve the clarity and edge information of the image.

[0048] (2) Weighted adaptive histogram equalization that takes gradient size into account can preserve image details while making the image's pixel value distribution more uniform, improving image contrast and brightness, and reducing the impact of uneven illumination on the image, making the image more uniform and clearer. Adjustments are made based on the image's local features, allowing for targeted enhancement of images in different regions, resulting in greater adaptability. The standard deviation of this algorithm is higher than that of the highest-performing G2D and MSR algorithms, increasing by 16.10% compared to the G2D algorithm. The standard deviation objectively represents image contrast; the larger the standard deviation, the stronger the image contrast. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the overall flow chart of the cutterhead image correction method applicable to the shield machine under uneven lighting conditions;

[0050] Figure 2 It is a lighting-reflection model;

[0051] Figure 3 (a) is the original image of a cutterhead inside a shield machine, (b) is the illumination component extracted by continuous Gaussian convolution of the cutterhead image, (c) is the grayscale image of the result after algorithm enhancement, and (d) is the result image after algorithm enhancement;

[0052] Figure 4 is the histogram of the original image;

[0053] Figure 5 is the histogram of the rectified image;

[0054] Figure 6(a) is the result image after MSR algorithm enhancement, (b) is the result image after CAHLE algorithm enhancement, (c) is the result image after G2D algorithm enhancement, (d) is the result image after RetinexNet algorithm enhancement, (e) is the result image after GT algorithm enhancement, and (f) is the result image after the algorithm of the present invention enhancement. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present invention will be described clearly and completely below. The present invention relates to a cutterhead image correction method suitable for use in shield machines under uneven illumination.

[0056] Example: Figure 1 As shown, first, a low-light or unevenly illuminated image is input and the image is converted from RGB to HSV space. Second, a small-scale Gaussian convolution kernel is used to perform continuous convolution operations on the extracted illumination components to obtain the illumination component distribution of the scene, calculate the local contrast and gradient information, and combine the local contrast and gradient magnitude to determine the gamma value of each pixel position. The gamma value is dynamically adjusted according to the local features of the image, and the gamma-corrected image is converted from HSV to RGB space. Finally, a weighted adaptive histogram equalization algorithm is applied to the image, and guided filtering is performed to obtain an enhanced image, which is then output. Specifically, the following steps are included:

[0057] S1: Input image, such as Figure 3 As shown in (a), the scene illumination component is extracted;

[0058] S1.1: Convert the original image to HSV color space, where the light reflection model is as follows: Figure 2 As shown, the brightness channel is extracted as the illumination component of the scene. The brightness value F(x,y) is the result of the combined effect of the illumination component R(x,y) irradiating the scene and the reflection component S(x,y) of the object surface.

[0059] F(x,y)=R(x,y)×S(x,y) (10)

[0060] Where: R(x,y) is the component of light irradiating the scene, S(x,y) is the component reflected from the object surface;

[0061] S1.2: Use a small-scale Gaussian convolution kernel to perform continuous convolution operations on the extracted illumination components. With the help of the small-scale Gaussian convolution kernel, the distribution of illumination components of the scene is obtained, such as Figure 3 As shown in (b), the estimated value of the illumination component formula:

[0062]

[0063] Where: G(x,y) represents the Gaussian function, “*” represents the convolution operator;

[0064] S2: Calculate the local contrast and gradient information of the illumination component image after continuous convolution, where the grayscale image is as follows Figure 3 (c)

[0065] S2.1: Compute the horizontal and vertical gradient responses by convolving the image with the Sobel operator, which is a convolution kernel in the horizontal and vertical directions.

[0066] The Sobel operator in the horizontal direction is denoted as G x yes:

[0067]

[0068] The Sobel operator in the vertical direction is denoted as G y yes:

[0069]

[0070] Horizontal gradient response G x The calculation formula for (x,y) is:

[0071]

[0072] Where: I(x,y) is the grayscale value of the original image at position (x,y), G x (x,y) is the weight of the horizontal Sobel operator at position (i,j), where i and j are the row and column indices of the Sobel operator respectively;

[0073] Vertical gradient response G y The calculation formula for (x,y) is:

[0074]

[0075] Where: I(x,y) is the grayscale value of the original image at position (x,y), G x (x,y) is the weight of the Sobel operator at position (i,j) in the vertical direction, where i and j are the row and column indices of the Sobel operator respectively;

[0076] Combine the horizontal and vertical gradient responses to calculate the gradient magnitude at each pixel position:

[0077]

[0078] Where: G x (x,y) and G y(x, y) are the Sobel operator element values in the horizontal and vertical directions respectively;

[0079] S2.2: Illumination component image I after continuous convolution of original illumination component image I c Apply Gaussian blur to smooth the image and reduce the effect of noise. Applying Gaussian blur at position (x,y) means:

[0080]

[0081] Where: I b (x,y) is the pixel value at position (x,y) after Gaussian blurring, k is the radius of the Gaussian kernel, σ is the standard deviation of the Gaussian kernel, and I c (x+i,y+j) represents the illumination component image I after continuous convolution c The pixel value at position (x+i,y+j), represents the weight of the Gaussian kernel;

[0082] S2.3: For each pixel location (x, y), calculate the standard deviation of the pixel values in the surrounding area R as an estimate of the local contrast. R is an n×n square area centered at the pixel (x, y). The local contrast C(x, y) is given by the following formula:

[0083]

[0084] Where: μ is the average value of the pixel value in the R area, I b (x+i,y+j) is the pixel value at position (x+i,y+j) in the image;

[0085] S2.4: Determine the gamma value for each pixel position by combining local contrast and gradient magnitude. Considering local contrast and gradient magnitude as weights, dynamically adjust the gamma value based on the local features of the image, so that different regions have different enhancement levels, thereby better adapting to different parts of the image. The formula is as follows:

[0086]

[0087] r=α·C(x,y)+β·G(x,y) (21)

[0088] Where: I (x,y) is the input image, I′ (x,y) is the rectified image, r is the exponent value used for brightness enhancement, C(x,y) is the local contrast, G(x,y) is the gradient size at each pixel position, α and β are adjustment coefficients used to balance the effects of local contrast and gradient size;

[0089] S2.5: After the brightness is corrected, convert it back to RGB space;

[0090] S3: Perform weighted adaptive histogram equalization on the image;

[0091] S3.1: Gradient information calculation weight w(x,y):

[0092] w(x,y)=e c·G(x,y) (twenty two)

[0093] Where: G(x,y) is the gradient size at each pixel position, c is a constant used to control the degree of influence of the gradient size on the weight;

[0094] S3.2: Segment the image into local regions. The image is segmented into continuous, non-overlapping m×n sub-blocks, where m and n can be 4, 6, 8, etc. Each sub-block contains N pixels. The larger the sub-block, the more obvious the enhancement effect, but the more image detail information is lost.

[0095] S3.3: Perform local histogram equalization. For each pixel value i in the local area, calculate the number of pixels with pixel value i in the area to form a grayscale histogram H(i) of the local area.

[0096] S3.4: Calculate its cumulative distribution function CDF(i), which represents the sum of the number of pixels with pixel values less than or equal to i in the local area;

[0097] S3.5: For each pixel (x, y) in the local area, map its original grayscale value I(x, y) to the equalized grayscale value I e (x,y), the grayscale value formula after equalization is as follows:

[0098]

[0099] Where: CDF(I(x,y)) represents the value of the cumulative distribution function CDF of the original gray value I(x,y) in the local area, N is the total number of pixels in the local area, L is the gray level of the pixel, is the normalization factor, which is used to ensure the correct grayscale value range after equalization;

[0100] S3.6: For each pixel position (x, y), the grayscale value I after local histogram equalization is e (x,y) is weighted with the corresponding weight w(x,y);

[0101] I f (x,y)=w(x,y)×I e (x,y) (24)

[0102] Where: I e (x,y) is the grayscale value after local histogram equalization, and w(x,y) is the corresponding weight;

[0103] S3.7: Perform guided filtering on the image to obtain an enhanced image;

[0104] S3.8: Output the enhanced image, such as Figure 3 (d) shown.

[0105] Figure 4 is the histogram of the original image, Figure 5 is the histogram of the corrected image. The comparison of the histograms of the images before and after correction using the method of the present invention is obtained by Figure 5 It can be seen that after correction, the areas with too low brightness in the original image are enhanced as a whole, and the dynamic range of the image is compressed, resulting in a better correction effect.

[0106] like Figure 6 (a) shows the MSR effect, (b) shows the CALHE effect, (c) shows the G2D effect, (d) shows the RetinexNet effect, (e) shows the GT effect, and (f) shows the effect of our algorithm. To further compare the processing effects of different algorithms, simulation experiments were conducted using Matlab software, using objective metrics such as standard deviation, mean gradient, and entropy. Standard deviation reflects the contrast characteristics of an image, mean gradient is an important indicator of image clarity, and entropy measures the amount of information carried by the image. Table 1 shows a comparison of the images before and after processing using different methods.

[0107] Table 1 Experimental data of six algorithms

[0108]

[0109] As can be seen from the table, based on the given data comparison results, it can be seen that the algorithm of the present invention has achieved better results than other algorithms in terms of the three indicators of standard deviation, average gradient and entropy.

[0110] (1) The standard deviation of the proposed algorithm is 82.78, which is higher than the highest G2D algorithm (71.80) and MSR algorithm (71.30) among other algorithms. It is 16.10% higher than the G2D algorithm. The proposed algorithm can effectively improve the contrast information of the image, making the pixel value distribution of the image more extensive and the image clearer and brighter visually.

[0111] (2) The average gradient value of the algorithm of the present invention is 168.61, which is much higher than the average gradient values of other algorithms, especially the highest GT algorithm (116.22). In comparison, the algorithm of the present invention is improved by 45.08%, indicating that the algorithm of the present invention can significantly improve the clarity and edge information of the image, making the image sharper.

[0112] (3) The entropy value of the algorithm of the present invention is 7.19, which is significantly improved compared with other algorithms and 9.44% higher than the MSR algorithm. The images processed by the algorithm of the present invention also contain more information and can extract more details and features.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not limiting. Although the technical solutions of the present invention are described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the present invention, and all of them should be included in the scope of protection of the present invention.

Claims

1. A cutterhead image correction method suitable for use in shield machines under uneven illumination, comprising the following steps: S1: Input image and extract scene illumination components; S1.1: Convert the original image to the HSV color space and extract the brightness channel as the illumination component of the scene; S1.2: Use a small-scale Gaussian convolution kernel to perform continuous convolution operations on the extracted illumination components. With the help of small-scale Gaussian convolution kernel, the distribution of illumination components of the scene is obtained; S2: Calculate local contrast and gradient information for the illumination component image after continuous convolution; S2.1: Calculate the horizontal and vertical gradient responses by convolving the Sobel operator with the image, combine the horizontal and vertical gradient responses, and calculate the gradient magnitude at each pixel position: Where: G(x,y) is the gradient size of each pixel position, G x (x,y) and G y (x, y) are the Sobel operator element values in the horizontal and vertical directions respectively; S2.2: Illumination component image I after continuous convolution of original illumination component image I c Apply a Gaussian blur to smooth the image and reduce the effects of noise, expressed as a Gaussian blur at position (x,y): Where: I b (x,y) is the pixel value at position (x,y) after Gaussian blurring, k is the radius of the Gaussian kernel, σ is the standard deviation of the Gaussian kernel, and I c (x+i,y+j) represents the illumination component image I after continuous convolution c The pixel value at position (x+i,y+j), represents the weight of the Gaussian kernel; S2.3: For each pixel position (x, y), calculate the standard deviation of the pixel values in the surrounding n×n square area R as an estimate of the local contrast: Where: C(x,y) is the local contrast, μ is the average value of the pixel value in the R area, I b (x+i,y+j) is the pixel value at position (x+i,y+j) in the image; S2.4: Determine the gamma value for each pixel position based on the local contrast and gradient magnitude. The formula is as follows: r=α×C(x,y)+β×G(x,y) (5) Where: I (x,y) is the input image, I′ (x,y) is the rectified image, r is the exponent value for brightness enhancement, C(x,y) is the local contrast, G(x,y) is the gradient size at each pixel position, α and β are adjustment coefficients used to balance the effects of local contrast and gradient size; S2.5: After the brightness is corrected, convert it back to RGB space; S3: Perform weighted adaptive histogram equalization on the image; S3.1: Gradient information calculation weight w(x,y): w(x,y)=e c·G(x,y) (6) Where: G(x,y) is the gradient size at each pixel position, c is a constant used to control the degree of influence of the gradient size on the weight; S3.2: Segment the image, dividing the original image into local regions; S3.3: Perform local histogram equalization. For each pixel value i in the local area, calculate the number of pixels with pixel value i in the area to form a grayscale histogram H(i) of the local area. S3.4: Calculate its cumulative distribution function CDF(i), which represents the sum of the number of pixels with pixel values less than or equal to i in the local area; S3.5: For each pixel (x, y) in the local area, map its original grayscale value I(x, y) to the equalized grayscale value I e (x,y), the grayscale value formula after equalization is as follows: Where: CDF(I(x,y)) represents the value of the cumulative distribution function CDF of the original gray value I(x,y) in the local area, N is the total number of pixels in the local area, L is the gray level of the pixel, is the normalization factor, which is used to ensure the correct grayscale value range after equalization; S3.6: For each pixel position (x, y), the grayscale value I after local histogram equalization is e (x,y) is weighted with the corresponding weight w(x,y); I f (x,y)=w(x,y)×I e (x,y) (8) Where: I e (x,y) is the grayscale value after local histogram equalization, and w(x,y) is the corresponding weight; S3.7: Perform guided filtering on the image to obtain an enhanced image; S3.8: Output the enhanced image.

Images