A fusion algorithm suitable for railway vehicle brake block image recognition

Abstract

This invention discloses a fusion algorithm for image recognition of brake pads and shoes of railway locomotives and rolling stock. Existing image fusion methods struggle to simultaneously achieve both computational efficiency and accuracy. This invention calculates the average gradient magnitude and standard deviation of the acquired visible light and infrared images of railway locomotive and rolling stock brake pads and shoes, respectively, using these as indices of image texture edge intensity and thermal target intensity to calculate the weights of the grayscale mapping function. An adaptive parameter grayscale mapping method enhances the original image, and a structural decomposition algorithm improves the fusion quantization index, thereby enhancing the texture edge information and thermal target information of the visible light image. This invention enhances useful information, reduces information loss in the fused image, and utilizes structural decomposition to achieve rapid image fusion. The fused image carries a large amount of information, resulting in a clear image with good pixel distribution characteristics, while simultaneously balancing computational efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a fusion algorithm for image recognition of brake pads and brake shoes of railway locomotives and rolling stock. Background Technology

[0002] Brake pads and shoes are core components of the locomotive and rolling stock braking system. When they wear down to a certain extent, they fail, affecting the braking performance and operational safety of the locomotive and rolling stock; therefore, regular inspection of brake pads and shoes is necessary. It is recommended to describe the difficulty of image acquisition based on the location of the locomotive and rolling stock's brake pads and shoes. With the development of sensor technology, various sensors are used to acquire different data information from locomotive and rolling stock brake pads and shoes. Infrared images can effectively distinguish targets by capturing radiation information emitted from railway locomotive and rolling stock brake pads and shoes during the day or night, but the acquired data lacks spatial resolution, and the texture and detail performance of the brake pads and shoes are poor. In contrast, visible images can obtain rich texture appearance of brake pads and shoes, but they cannot distinguish thermal targets, especially in low-light environments. Image fusion technology plays an important role in enhancing information and reducing redundancy when integrating various data.

[0003] Existing image fusion methods struggle to balance computational efficiency and accuracy simultaneously. Basic methods are fast, but their fusion performance is limited; advanced methods can produce good visual results, but their computational efficiency is low. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a fusion algorithm for image recognition of brake pads and brake shoes of railway locomotives and rolling stock. It enhances the original image through an adaptive parameter grayscale mapping method and improves the fusion quantization index of the image through a structural decomposition algorithm. This enhances the texture edge information and thermal target information of the visible light image, ensuring both image calculation speed and quality.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A fusion algorithm for image recognition of brake pads and brake shoes of railway locomotives and rolling stock is characterized by the following steps:

[0007] Step S10: Calculate the average gradient magnitude G for the acquired visible light and infrared images of the brake pads and shoes of railway locomotives and rolling stock. Vis G IR With standard deviation S Vis S IR ;

[0008] Step S20: Using the results of step S10, the weights of the grayscale mapping function are calculated as the image texture edge intensity and thermal target intensity indices, respectively.

[0009]

[0010] In the formula, G Vis G is the average gradient modulus of a visible light image. IR S is the average gradient modulus of the infrared image. Vis S represents the standard deviation of the visible light image. IR The standard deviation of the infrared image;

[0011] c VIs For visible light image adaptive weights, c IR γ is the adaptive weight for the infrared image, which is calculated using the standard deviation of the two images to determine the first transformation weight of the infrared image.

[0012] Step S30:

[0013] Let the original image be X, and define the texture edge enhancement grayscale mapping function S(X) and the thermal target enhancement function gamma(X) as follows:

[0014]

[0015] Step S40:

[0016] Computational grayscale mapping to enhance visible and infrared images X′ Vis =S(X) Vis ), X′ IR =S(gamma(X) IR ));

[0017] Step S50:

[0018] For the two enhanced images obtained in step S40, slide the image window respectively, and decompose the image sliding window x into the image mean l, the image signal intensity ||xl|| and the signal structure (xl) / ||xl|| according to the structural decomposition formula x=l·1+||xl||·((xl) / ||xl||);

[0019] Step S60:

[0020] In step S50, parameters γ1 and p are set, and the fusion parameters α and β of each decomposed part are calculated. The weighted fused image sliding window is then calculated.

[0021]

[0022] β i =||x i -l i || p (i = 1, 2; p ≥ 0)

[0023]

[0024] Step S70:

[0025] Perform step S60 on the sliding windows of the entire image to obtain the corresponding weight α of the mean of each sliding window image. i The weights W of the entire image's base layer components, consisting of / (α1+α2), are... Bi (i = 1, 2), the corresponding weights for image signal intensity are max{||x1-l1||,||x2-l2||}×||x i -l i || p-1 The image detail layer weights W, composed of (β1+β2), are... Di (i = 1, 2);

[0026] Step S80:

[0027] Obtain the image base layer f(W) Bi ⊙f(X i )) and image detail layer f(W Di )⊙X i -f(W Di ⊙f(X i The fused image is obtained by summing the values ​​of each group, where f(·) is the mean filter and ⊙ is the pixel-wise product. for:

[0028]

[0029] Specifically, in step S10, the average gradient magnitude and standard deviation are calculated as follows:

[0030]

[0031] The image is X, i is the pixel value, N is the number of pixels, mean is the average, |·| is the modulus, and grad is the gradient.

[0032] The beneficial effects of this invention are:

[0033] 1) This invention uses grayscale mapping to enhance the texture and target information of the fused image, thereby enhancing useful information and reducing the loss of information in the fused image. It also utilizes structural decomposition to achieve rapid image fusion.

[0034] 2) The image fused by this invention carries a large amount of information, thus the fused image is clear and has good pixel distribution characteristics;

[0035] 3) This invention has a fast calculation speed, while taking into account both calculation efficiency and accuracy. Attached Figure Description

[0036] Figure 1 This is a flowchart of the process of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the principle of brake pad and brake shoe image detection.

[0038] Figure 3 Visible light image;

[0039] Figure 4 Infrared image;

[0040] Figure 5 This is the fusion result of the gradient transfer method;

[0041] Figure 6 The result is a fusion of anisotropic diffusion and KL transform methods;

[0042] Figure 7 This is the result of convolutional neural network fusion.

[0043] Figure 8 This is the fusion result of the present invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to specific embodiments.

[0045] like Figure 2 As shown, firstly, visible light and infrared images of brake pads and brake shoes of railway locomotives and rolling stock are acquired using infrared and visible light cameras; the present invention uses an enhanced structural decomposition method to extract features from thermal and visual images; then, based on the decomposition results, the two images are merged to synthesize a fused image; the fused image contains more information than a single image and can be used as input for an image-based brake pad detection algorithm.

[0046] like Figure 1 As shown, the present invention includes the following steps:

[0047] Step S10: Visible light image and infrared image as follows Figure 3 , Figure 4 As shown, the average gradient modulus G is calculated for the acquired visible light and infrared images of the brake pads and shoes of railway locomotives and rolling stock. Vis G IR With standard deviation S Vis S IR The calculation method is as follows:

[0048]

[0049] The image is X, i is the pixel value, N is the number of pixels, mean is the average, |·| is the modulus, and grad is the gradient.

[0050] Step S20: Using the results of step S10, the weights of the grayscale mapping function are calculated as the image texture edge intensity and thermal target intensity indices, respectively.

[0051]

[0052] In the formula, G Vis G is the average gradient modulus of a visible light image. IR S is the average gradient modulus of the infrared image. Vis S represents the standard deviation of the visible light image. IR The standard deviation of the infrared image;

[0053] c VIs For visible light image adaptive weights, c IR γ is the adaptive weight for infrared images, which is the adaptive weight used to calculate the degree of the first transformation of the infrared image by means of the standard deviation of the two images.

[0054] Step S30:

[0055] Let the original image be X, and define the texture edge enhancement grayscale mapping function S(X) and the thermal target enhancement function gamma(X) as follows:

[0056]

[0057] Step S40:

[0058] Computational grayscale mapping to enhance visible and infrared images X′ Vis =S(X) Vis ), X′ IR =S(gamma(X) IR )); to X-ray visible light image Vis Substituting S(X), the infrared image X IR After substituting gamma(X) into S(X), this step completes the effects of enhancing the grayscale texture of visible light images and enhancing the grayscale texture of infrared images and infrared targets based on adaptive weights.

[0059] Step S50:

[0060] For the two enhanced images obtained in step S40, slide the image window respectively, and decompose the image sliding window x into the image mean l, the image signal intensity ||xl|| and the signal structure (xl) / ||xl|| according to the structural decomposition formula x=l·1+||xl||·((xl) / ||xl||);

[0061] Step S60:

[0062] In step S50, parameters γ1 and p are set, and the fusion parameters α and β of each decomposed part are calculated. These parameters generate fusion weights. Different parameters will change the display bias of the fused image towards details or color blocks. A larger γ1 will make the fused image biased towards presenting information about thermal targets, and a larger p will make the fused image biased towards presenting more details rather than color blocks. The weighted fused image sliding window is then calculated.

[0063]

[0064] β i =||x i -l i || p (i = 1, 2; p ≥ 0)

[0065]

[0066] Step S70:

[0067] Perform step S60 on the sliding windows of the entire image to obtain the corresponding weight α of the mean of each sliding window image. i The weights W of the entire image's base layer components, consisting of / (α1+α2), are... Bi (i = 1, 2), the corresponding weights for image signal intensity are max{||x1-l1||,||x2-l2||}×||x i -l i || p-1 The image detail layer weights W, composed of (β1+β2), are... Di (i = 1, 2);

[0068] Step S80:

[0069] Obtain the image base layer f(W) Bi ⊙f(X i )) and image detail layer f(W Di )⊙X i -f(W Di ⊙f(X i The fused image is obtained by summing the values ​​of each group, where f(·) is the mean filter and ⊙ is the pixel-wise product. for:

[0070]

[0071] The following table shows a comparative analysis of the fusion results obtained by the present invention and existing methods using the average quality evaluation factors of multiple traditional images.

[0072] EN (Information Entropy) reflects the amount of information in the fused image, MI (Mutual Information) reflects the amount of information the fused image contains from the source images, and Q... G(Edge Preservation) evaluates the gradient information inherited by the fused image from the original infrared and visible images. SD (Standard Deviation) is the main measure of the contrast and sharpness of the fused image. Among all four metrics, the larger the value, the better the fusion result of the corresponding algorithm.

[0073] Table 1 Objective fusion indicators for each method

[0074]

[0075]

[0076] As shown in Table 1, the edge preservation degree Q of this invention among the four parameters is... TV / F At a mid-level, with similar results to the structural patch decomposition method, the other three metrics are significantly better than other contrast fusion methods, thus exhibiting the clearest image, the largest information content, and the best pixel distribution characteristics; from Figure 5-8 The results also show that the method obtains significant thermal target information while preserving effective information such as visible light image texture and edges, and it takes the shortest time.

[0077] The content of this invention is not limited to the embodiments listed. Any equivalent modifications made by those skilled in the art to the technical solutions of this invention by reading this specification are covered by the claims of this invention.

Claims

1. A fusion algorithm for image recognition of brake pads and brake shoes of railway locomotives and rolling stock, characterized in that: Includes the following steps: Step S10: Calculate the average gradient magnitude G for the acquired visible light and infrared images of the brake pads and shoes of railway locomotives and rolling stock. Vis G IR With standard deviation S Vis S IR ; Step S20: Using the results of step S10, the weights of the grayscale mapping function are calculated as the image texture edge intensity and thermal target intensity indices, respectively. In the formula, G Vis G is the average gradient modulus of a visible light image. IR S is the average gradient modulus of the infrared image. Vis S represents the standard deviation of the visible light image. IR The standard deviation of the infrared image; c Vis For visible light images, adaptive weights, c IR For infrared image adaptive weights, γ To calculate the adaptive weights for the first transformation of an infrared image using the standard deviation of two images; Step S30: Let the original image be X, and define the texture edge enhancement grayscale mapping function S(X) and the thermal target enhancement function gamma(X) as follows: Step S40: Computational grayscale mapping to enhance visible and infrared images ； Step S50: For the two enhanced images obtained in step S40, take a sliding window of the image and apply it according to the structural decomposition formula x= l ·1+||x- l ||·((x- l ) / ||x- l ||) Decompose the image sliding window x into the image mean. l Image signal intensity ||x- l ||and signal structure (x- l ) / ||x- l ||; Step S60: In step S50, parameters γ1 and p are set, and the fusion parameters of image mean, image signal intensity, and signal structure are calculated. α、β， Calculate the weighted fused image sliding window : b i =||x i -l i || p ,(i=1,2;p≥0) ； Step S70: Perform step S60 on the sliding window of the entire image to obtain the corresponding weights of the image mean of each sliding window. α i / ( α 1 + α 2 The weights W of the base layer components of the entire image are composed of... Bi For i=1,2, the corresponding weights for image signal strength are max{||x1- l 1||,||x2- l 2||}×||x i - l i || p-1 / ( β 1 +β 2 The image detail layer weights W are composed of... Di ; Step S80: Obtain the image base layer f (W Bi ⊙f(X i Image detail layer f (W Di )⊙X i - f (W Di ⊙ f (X i )),in f (·) represents mean filtering, and ⊙ represents pixel-wise multiplication. The fused image is obtained by summing the visible light image and the infrared image. for: 。 2. The fusion algorithm for image recognition of brake pads and brake shoes of railway locomotives and rolling stock according to claim 1, characterized in that: In step S10, the average gradient magnitude and standard deviation are calculated as follows: In the formula, j For pixel values, N For the number of pixels, mean To take the average, To take the mold, grad To calculate the gradient.

Images