Image enhancement method and system for titanium metal surface detection

Abstract

The present application relates to the technical field of image processing, and particularly relates to an image enhancement method and system for titanium metal surface detection. The method comprises the following steps: converting a titanium metal surface image to be processed from an RGB space to an HSI space, obtaining a luminance component, decomposing an initial reflection component into a plurality of high-frequency sub-band coefficients and a low-frequency sub-band coefficient, determining anisotropy degrees of each pixel point, generating an anisotropy saliency map, enhancing the high-frequency sub-band coefficients corresponding to the pixels with anisotropy degrees higher than a preset threshold, otherwise, inhibiting, reconstructing an enhanced reflection component by using the adjusted high-frequency sub-band coefficients and the low-frequency sub-band coefficient without processing, obtaining an enhanced luminance component, converting the enhanced luminance component back to the RGB space, and outputting a final enhanced image. That is, the scheme of the present application can obtain an enhanced image with a clear defect contour, few artifacts and preserved original color information, greatly improving the accuracy and reliability of subsequent automatic defect recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to an image enhancement method and system for titanium metal surface detection. BACKGROUND

[0002] The existing image enhancement techniques can be mainly divided into spatial domain methods, transform domain methods and methods based on Retinex theory.

[0003] The spatial domain methods such as histogram equalization (HE) and its variants (such as CLAHE) can improve the overall contrast of the image, but are prone to over-enhancement and amplify noise, and have limited improvement effect on local details.

[0004] The transform domain methods such as Fourier transform and traditional wavelet transform can process the image from the frequency domain, but the traditional wavelet transform lacks shift invariance and good direction selectivity, and is prone to produce ringing artifacts when processing linear defects with specific directionality.

[0005] The methods based on Retinex theory can effectively compensate for the effects of uneven illumination by decomposing the image into illumination component and reflection component, but in practice, how to accurately estimate the illumination component to avoid the halo effect, and how to effectively process the separated reflection component, are still the core challenges of this type of method. The existing methods based on Retinex theory usually uniformly process all high-frequency detail information when enhancing the reflection component, and cannot distinguish between real defects (such as scratches with strong anisotropy) and random noise (usually isotropic), resulting in the lack of pertinence of the enhancement of the reflection component, and the difficulty in selectively highlighting defects of specific morphology. SUMMARY

[0006] The purpose of the present application is to provide an image enhancement method and system for titanium metal surface detection to solve the problem of uneven illumination and difficulty in highlighting weak defects in the prior art. To this end, the present application provides solutions in the following two aspects.

[0007] In the first aspect, the present application provides an image enhancement method for titanium metal surface detection, comprising:

[0008] The image of the titanium metal surface to be processed is converted from RGB space to HSI space to obtain a luminance component I; based on the luminance component I, an initial illumination component is estimated by a guided filter weighted with a multi-scale Gaussian kernel function, and an initial reflection component is separated according to a Retinex model; a dual-tree complex wavelet transform is performed on the initial reflection component to decompose it into a plurality of high-frequency sub-band coefficients and a low-frequency sub-band coefficient in different scales and directions; a gradient structure tensor of the initial reflection component is calculated, and the anisotropy degree of each pixel point is determined according to the eigenvalue distribution of the gradient structure tensor to generate an anisotropy saliency map; a piecewise nonlinear mapping function related to the high-frequency sub-band coefficient amplitude and the anisotropy saliency map is constructed, the high-frequency sub-band coefficient corresponding to the pixel with an anisotropy degree higher than a preset threshold is enhanced by using a logarithmic function, and otherwise, the high-frequency sub-band coefficient is suppressed by using a gamma correction function to obtain an adjusted high-frequency sub-band coefficient; the dual-tree complex wavelet inverse transform is performed by using the adjusted high-frequency sub-band coefficient and the low-frequency sub-band coefficient without processing to reconstruct an enhanced reflection component; the enhanced reflection component is multiplied by the initial illumination component to obtain an enhanced luminance component I', and the enhanced luminance component I' is combined with the original hue component H and the saturation component S to convert back to the RGB space, and an enhanced image is output.

[0009] Preferably, the luminance component I is calculated from the average value of R, G and B channels.

[0010] Preferably, the initial illumination component is estimated by a guided filter weighted with a multi-scale Gaussian kernel function, including: filtering the luminance component I by using three Gaussian kernel functions with different standard deviations to obtain three images with different degrees of blurring , , ; performing weighted fusion on the three images , , to obtain a weighted Gaussian filter image :

[0011] ,

[0012] wherein , , is a weight coefficient, and , is the weighted Gaussian filter image, represents the first image, represents the second image, represents the third image; and the weighted Gaussian filter image As the guide image, the brightness component I is used as the input image, and the initial illumination component is calculated through the guided filter.

[0013] Preferably, the weighted fusion adopts equal weight fusion, and the weight coefficient , , Both .

[0014] Preferably, the dual-tree complex wavelet transform is implemented by two parallel real discrete wavelet transforms, one of which uses a q-shift filter bank and the other uses a king-shift filter bank.

[0015] Preferably, separating the initial reflection component according to the Retinex model includes: obtaining the initial reflection component by dividing the brightness component I by the initial illumination component.

[0016] Preferably, generating an anisotropic saliency map comprises: calculating a gradient structure tensor J of each pixel point in the initial reflection component, wherein the gradient structure tensor J is obtained by calculating the covariance matrix of its horizontal and vertical gradients in a neighborhood window centered on the pixel point; calculating two eigenvalues ​​of the gradient structure tensor J and ( ≥ ≥0); calculate the anisotropy metric value based on the eigenvalue ,in To prevent the smallest positive number with zero denominator, represents the first eigenvalue of the gradient tensor, Represents the second eigenvalue of the gradient tensor; the anisotropy measure of all pixels Normalized to the interval [0, 1] to generate the anisotropic saliency map.

[0017] Preferably, the step of constructing a piecewise nonlinear mapping function related to the amplitude of the high-frequency sub-band coefficient and the anisotropic saliency map comprises: comparing the pixel value in the anisotropic saliency map with a preset threshold value T; if the pixel value is greater than the threshold value T, then the corresponding high-frequency sub-band coefficient is Through the function Make adjustments, including is a symbolic function, and is the gain control parameter, is the high frequency subband coefficient, is the corrected output value; if the pixel value is not greater than the threshold T, then the corresponding high frequency sub-band coefficient Through the gamma correction function Make adjustments, including is the gamma correction parameter, and 0< <1, is the high frequency subband coefficient, is the output value after gamma correction.

[0018] Preferably, the dual-tree complex wavelet transform decomposes the initial reflection component into N scales, where N is an integer greater than or equal to 1, thereby obtaining 6N high-frequency sub-band coefficients and one low-frequency sub-band coefficient.

[0019] In a second aspect, an image enhancement system for titanium metal surface inspection includes:

[0020] A processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned image enhancement method for titanium metal surface detection.

[0021] The beneficial effects of the present invention are as follows: compared with the prior art, the titanium metal surface inspection image enhancement method proposed in the present invention can effectively solve the problem of uneven illumination and difficulty in highlighting subtle defects. The method can more accurately estimate and separate illumination components through a guided filter weighted fusion of multi-scale Gaussian kernel functions, effectively suppressing the halo artifacts commonly seen in traditional Retinex methods. By introducing gradient structure tensor analysis and calculating the degree of anisotropy of pixels, it can quantitatively distinguish linear defects with strong directionality (such as scratches) from isotropic background noise, and then construct a piecewise nonlinear mapping function to perform logarithmic enhancement on the high-frequency coefficients of the defect area and gamma suppression on the noise area. This processing method achieves selective highlighting of defect features and targeted suppression of noise. The enhanced image finally obtained not only has clear defect contours and high contrast, but also has a pure background, few artifacts, and retains the original color information, greatly improving the accuracy and reliability of subsequent automated defect recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The following is a schematic flow chart showing the steps of the image enhancement method for titanium metal surface detection in this embodiment;

[0023] Figure 2 The structural block diagram of the image enhancement system for titanium metal surface detection in this embodiment is schematically shown. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] like Figure 1 As shown, an image enhancement method for titanium metal surface detection in this embodiment includes the following steps:

[0026] Step S1: Convert the titanium metal surface image to be processed from RGB space to HSI space to obtain its brightness component I; based on the brightness component I, estimate the initial illumination component through a guided filter with weighted fusion of multi-scale Gaussian kernel functions, and separate the initial reflection component according to the Retinex model.

[0027] Specifically, the input RGB image is converted to the HSI color space, where the luminance component I is calculated by averaging the R, G, and B channels. Subsequently, the luminance component I itself is used as the guide image and input image, and three Gaussian kernel functions of different scales are set, for example, with standard deviations of 15, 80, and 250, respectively. The luminance component I is filtered three times using the guide filter to obtain three filtering results with different degrees of smoothness. The three results are weighted averaged, and the weight coefficients can be set equal, that is, one-third each. The fused image is the initial illumination component L. According to the Retinex model, the initial reflection component R is obtained by dividing each pixel value of the original luminance component I by the pixel value of the initial illumination component L at the corresponding position.

[0028] In an optional embodiment, the initial illumination component is estimated by a guided filter weighted fusion of a multi-scale Gaussian kernel function, including:

[0029] S11: Three different standard deviations are used , , The brightness component I is filtered by the Gaussian kernel function to obtain three images with different blur levels. , , ;

[0030] S12: For three images , , Perform weighted fusion to obtain weighted Gaussian filtered image :

[0031] ,

[0032] in , , is the weight coefficient, and , is the weighted Gaussian filtered image, represents the first image, represents the second image, represents a third image;

[0033] S13: The weighted Gaussian filter image As the guide image, the brightness component I is used as the input image, and the initial illumination component is calculated through the guided filter.

[0034] Specifically, in order to accurately estimate the illumination component as the macroscopic structure of the image, this method combines the advantages of multi-scale analysis and edge-preserving filtering. The brightness image of the pixel I, the standard deviations of the three Gaussian kernels are set to be small scale Equal to 5, medium scale Equal to 20, large scale Equal to 80. Small-scale filtering can capture local illumination changes, large-scale filtering can reflect global illumination trends, and medium-scale filtering serves as a transition.

[0035] For three blurred images 、 and Perform weighted fusion. A simple implementation method is to use equal weight fusion, that is, , , Set them to one third and get the fused image This step integrates illumination information of different scales, but may produce halo effects at the edges of objects. To solve this problem, As the guide image, the original brightness image I is fed into the guide filter as the input image. The guide filter uses the edge information in I to guide the The filtering process of the output initial illumination component is smooth and clear, maintaining the important edges in the original image and avoiding halo artifacts.

[0036] Step S2: performing a dual-tree complex wavelet transform on the initial reflection component to decompose it into a plurality of high-frequency sub-band coefficients of different scales and directions and one low-frequency sub-band coefficient.

[0037] Specifically, the initial reflection component R obtained in step S1 is decomposed into three layers using a dual-tree complex wavelet transform. The transformation is implemented by two parallel real discrete wavelet transforms, one of which uses a q-shift filter group and the other uses a king-shift filter group. Each layer of decomposition will produce a low-frequency sub-band and six high-frequency sub-bands with strong directional selectivity. These six high-frequency sub-bands correspond to directions of plus or minus 15 degrees, plus or minus 45 degrees, and plus or minus 75 degrees, respectively. After three layers of decomposition, a total of 18 high-frequency sub-band coefficient matrices and a final low-frequency sub-band coefficient matrix are obtained.

[0038] In an optional embodiment, the dual-tree complex wavelet transform decomposes the initial reflection component into N scales, where N is an integer greater than or equal to 1, thereby obtaining 6N high-frequency sub-band coefficients and one low-frequency sub-band coefficient.

[0039] Specifically, the dual-tree complex wavelet transform is an image decomposition tool that outperforms the traditional discrete wavelet transform. It uses two parallel real wavelet filter banks to approximate the analytic wavelet transform, resulting in near-translation invariance and excellent directional selectivity. This allows for more effective capture of image features and reduced aliasing distortion.

[0040] In one embodiment of this method, the number of decomposition levels is selected to be four. After the initial reflection component image undergoes four-level dual-tree complex wavelet decomposition, a low-frequency subband is generated, representing the overall image overview. Simultaneously, at each level of decomposition, six high-frequency subbands are generated, corresponding to directional information at ±15 degrees, ±45 degrees, and ±75 degrees, respectively. This results in a total of 24 high-frequency subband coefficient matrices, which depict edge and texture details of the image at different scales and directions.

[0041] Step S3, calculate the gradient structure tensor of the initial reflection component, and determine the degree of anisotropy of each pixel point according to the eigenvalue distribution of the gradient structure tensor, and generate an anisotropy saliency map; construct a piecewise nonlinear mapping function related to the amplitude of the high-frequency sub-band coefficient and the anisotropy saliency map, and enhance the high-frequency sub-band coefficients corresponding to pixels with an anisotropy degree higher than a preset threshold using a logarithmic function, otherwise suppress them using a gamma correction function to obtain adjusted high-frequency sub-band coefficients.

[0042] Specifically, the gradient structure tensor J of each pixel in the initial reflection component is calculated, and the gradient structure tensor J is obtained by calculating the covariance matrix of its horizontal and vertical gradients in a neighborhood window centered on the pixel point; the two eigenvalues ​​of the gradient structure tensor J are calculated. and ( ≥ ≥0); calculate the anisotropy metric value based on the eigenvalue:

[0043] ,in To prevent the smallest positive number with zero denominator, represents the first eigenvalue of the gradient tensor, represents the second eigenvalue of the gradient tensor;

[0044] The anisotropy value of all pixels Normalize to the interval [0, 1] to generate anisotropic saliency map.

[0045] In order to distinguish edge textures from flat areas in an image, this method calculates anisotropic saliency maps. The gradient of the initial reflection component image is calculated, for example, using the Sobel operator to obtain the horizontal gradient of each pixel. and vertical gradient . In a pixel centered In the neighborhood window, the gradient structure tensor J is calculated, whose elements are The mean of the squares, The mean of the squares, and and The mean of the products.

[0046] calculate Two eigenvalues ​​of matrix J and If a pixel lies on a strong vertical edge, its horizontal gradient Will be much larger than the vertical gradient ,lead to Much greater than ,For example For 5000 is 10. At this time, the anisotropy value pass minus The difference divided by 、 With a very small number The sum of is very close to 1. On the contrary, in the flat area, both eigenvalues ​​are close to zero. The value is also close to zero. Finally, all pixels The values ​​are linearly stretched to the range of 0 to 1, and the highlighted areas correspond to the linear structures in the image.

[0047] Constructing a piecewise nonlinear mapping function related to the amplitude of the high-frequency sub-band coefficient and the anisotropic saliency map, including: comparing the pixel value in the anisotropic saliency map with a preset threshold T; if the pixel value is greater than the threshold T, then the corresponding high-frequency sub-band coefficient Through the function Make adjustments, including is a symbolic function, and is the gain control parameter, is the high frequency subband coefficient, is the corrected output value; if the pixel value is not greater than the threshold T, then the corresponding high frequency sub-band coefficient Through the gamma correction function Make adjustments, including is the gamma correction parameter, and 0< <1, is the high frequency subband coefficient, is the output value after gamma correction.

[0048] Specifically, an adaptive strategy is used to enhance high-frequency details, which relies on the guidance of the anisotropic saliency map. A threshold T is set, for example, T is equal to 0.7. For each high-frequency subband coefficient obtained by the dual-tree complex wavelet transform , check the pixel value of the corresponding position in the anisotropic saliency map.

[0049] If the pixel value is greater than 0.7, it means that the coefficient corresponds to a significant edge or contour in the image. In this case, a logarithmic function is used to adjust the image to avoid over-sharpening. For example, the gain parameter is 25, is 0.05, for a coefficient of 60 , the adjusted value equal Multiply the sign of Multiply by 1 and add and The logarithm of the product of absolute values ​​can moderately enhance edge contrast while suppressing ringing effects.

[0050] If the pixel value is not greater than 0.7, it means that the coefficient corresponds to a texture area or a flat area. In this case, the gamma correction function is used for adjustment. For example, the gamma parameter is 0.85, for a coefficient of 15 , the adjusted value equal The sign of the multiplied by its absolute value This method can effectively improve the visibility of weak textures and make dark details clearer, while avoiding significantly amplifying potential noise.

[0051] Step S4: performing an inverse dual-tree complex wavelet transform using the adjusted high-frequency sub-band coefficients and the unprocessed low-frequency sub-band coefficients to reconstruct an enhanced reflection component.

[0052] Specifically, the 18 high-frequency subband coefficients adjusted by the piecewise nonlinear mapping function in step S3 are combined with the original, unprocessed low-frequency subband coefficients decomposed in step S2 as input to the dual-tree complex wavelet inverse transform. A standard dual-tree complex wavelet inverse transform reconstruction algorithm is then executed, synthesizing all subband coefficients from the coarsest scale to the finest scale. Ultimately, a single image with the same size as the original reflection component is obtained, which is the enhanced reflection component R'.

[0053] Step S5: multiply the enhanced reflection component by the initial illumination component to obtain an enhanced brightness component I', and combine the enhanced brightness component I' with the original hue component H and saturation component S, converting them back to RGB space, and outputting the final enhanced image.

[0054] Specifically, the enhanced reflection component R' obtained in step S4 is pixel-wise multiplied with the initial illumination component L estimated in step S1 to obtain an enhanced intensity component I'. Then, the enhanced intensity component I' is recombined with the hue component H and the saturation component S preserved from the original image when converted in step S1 to form a new HSI image. A standard HSI to RGB color space conversion algorithm is invoked to convert the new HSI image back to the RGB color space, thus obtaining the final color-fidelity and defect-detailed enhanced image.

[0055] The present application also provides an image enhancement system for titanium metal surface detection. As shown in Figure 2 the system comprises a processor and a memory, the memory stores computer program instructions which, when executed by the processor, implement the image enhancement method for titanium metal surface detection according to the present application.

[0056] The system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and thus will not be described here.

[0057] In the present application, the aforementioned memory can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or apparatus. For example, the computer readable storage medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random-Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC) and the like, or any other medium that can be used to store desired information and can be accessed by an application, module or both. Any such computer storage medium can be part of the device or accessible or connectable to the device. Any application or module described in the present application can be implemented by computer readable / executable instructions stored or otherwise held by such computer readable medium.

[0058] In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three or more, etc., unless otherwise explicitly specified.

[0059] Although this specification has shown and described several embodiments of the present invention, it is obvious to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and substitutions without departing from the idea and spirit of the present invention.

Claims

1. An image enhancement method for titanium metal surface detection, characterized in that: The following steps are involved: Convert the titanium metal surface image to be processed from RGB space to HSI space to obtain its brightness component I; Based on the brightness component I, the initial illumination component is estimated through a guided filter weighted fusion of multi-scale Gaussian kernel functions, and the initial reflection component is separated according to the Retinex model; Performing dual-tree complex wavelet transform on the initial reflection component to decompose it into multiple high-frequency sub-band coefficients of different scales and directions and one low-frequency sub-band coefficient; Calculate the gradient structure tensor J of each pixel in the initial reflection component. The gradient structure tensor J is obtained by calculating the covariance matrix of its horizontal and vertical gradients in a neighborhood window centered on the pixel. Calculate the two eigenvalues ​​of the gradient structure tensor J and ( ≥ ≥0); calculate the anisotropy metric value based on the eigenvalue , To prevent the smallest positive number with zero denominator, 、 Represent the first eigenvalue and second eigenvalue of the gradient tensor respectively; The anisotropy value of all pixels Normalize to the interval [0, 1] to generate anisotropic saliency map; Comparing the pixel value in the anisotropic saliency map with a preset threshold T; If the pixel value is greater than the threshold T, the corresponding high-frequency sub-band coefficient Through the function Make adjustments, is a symbolic function, and is the gain control parameter, is the high frequency subband coefficient, is the output value after correction; If the pixel value is not greater than the threshold T, then the corresponding high-frequency sub-band coefficient Through the gamma correction function Make adjustments, is the gamma correction parameter, and 0< <1, is the output value after gamma correction; The high-frequency sub-band coefficients corresponding to pixels with anisotropy higher than a preset threshold are enhanced using a logarithmic function, and vice versa, they are suppressed using a gamma correction function to obtain the adjusted high-frequency sub-band coefficients; Using the adjusted high-frequency sub-band coefficients and the unprocessed low-frequency sub-band coefficients, an inverse double-tree complex wavelet transform is performed to reconstruct the enhanced reflection component. The enhanced reflection component is multiplied by the initial illumination component to obtain the enhanced brightness component I', and the enhanced brightness component I' is combined with the original hue component H and saturation component S, converted back to RGB space, and the final enhanced image is output.

2. The image enhancement method for titanium metal surface detection according to claim 1, characterized in that: The brightness component I is calculated by averaging the values ​​of the three channels R, G, and B.

3. The image enhancement method for titanium metal surface detection according to claim 1, characterized in that: The initial illumination component is estimated by a guided filter weighted fusion of a multi-scale Gaussian kernel function, including: Three Gaussian kernel functions with different standard deviations are used to filter the brightness component I to obtain three images with different blur levels. , , ; For three images , , Perform weighted fusion to obtain weighted Gaussian filtered image : ， in , , is the weight coefficient, and , is the weighted Gaussian filtered image, represents the first image, represents the second image, represents a third image; The weighted Gaussian filtered image As the guide image, the brightness component I is used as the input image, and the initial illumination component is calculated through the guided filter.

4. The image enhancement method for titanium metal surface detection according to claim 3, characterized in that: The weighted fusion adopts equal weight fusion, and the weight coefficient , , Both .

5. The image enhancement method for titanium metal surface detection according to claim 1, characterized in that: The dual-tree complex wavelet transform is implemented by two parallel real discrete wavelet transforms, one of which uses a q-shift filter bank and the other uses a king-shift filter bank.

6. The image enhancement method for titanium metal surface detection according to claim 1, characterized in that: The separating the initial reflection component according to the Retinex model includes: obtaining the initial reflection component by dividing the brightness component I by the initial illumination component.

7. The image enhancement method for titanium metal surface detection according to claim 1, characterized in that: The dual-tree complex wavelet transform decomposes the initial reflection component into N scales, where N is an integer greater than or equal to 1, thereby obtaining 6N high-frequency sub-band coefficients and one low-frequency sub-band coefficient.

8. An image enhancement system for titanium metal surface detection, characterized in that: include: A processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the image enhancement method for titanium metal surface detection according to any one of claims 1 to 7.

Images