Image fusion identification method and system based on flat metal wire surface defect detection

Abstract

The invention discloses an image fusion identification method and system based on flat metal wire surface defect detection, and relates to the technical field of visual image processing. Synchronously acquiring a multi-angle original image through the double-camera module and the multi-angle light source which are arranged on the upper side and the lower side of the metal flat wire; illumination suppression and reflection enhancement are carried out through the preprocessing unit, and noise and uneven illumination are eliminated; the fusion unit decomposes the image into a low-frequency sub-band and a high-frequency sub-band, the low-frequency sub-band calculates a definition weight based on gradient intensity to carry out weighted fusion, the high-frequency sub-band quantizes local complexity by calculating the sum of pixel neighborhood second derivative absolute values, the local complexity is used as an excitation input PCNN model, and a source image coefficient with the most pulse output times is dynamically selected to carry out fusion; and finally reconstructing and generating a comprehensive image and inputting the comprehensive image into a layered recognition unit for defect positioning and classification. And full-coverage online accurate detection of the surface defects of the moving metal flat wire is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to visual image recognition processing technology, and in particular to an image fusion recognition method and system based on metal flat wire surface defect detection. BACKGROUND

[0002] In the field of metal flat wire manufacturing, real-time detection of surface defects is directly related to product quality and production safety. Surface defects are generally scratches, cracks, pits, and oxidation spots.

[0003] Among them, the traditional detection method has the following problems:

[0004] Single light source angle is easy to form strong reflection area due to metal reflection, shielding small defects; conventional global imaging resolution is insufficient, and it is difficult to capture sub-millimeter level defect features.

[0005] Existing algorithms mostly rely on whole image analysis, with large redundant calculation, low recognition sensitivity for weak contrast defects, and high false detection and missed detection rates.

[0006] The prior art, such as Chinese patent application CN116664871A, proposes an intelligent control method and system based on deep learning. In this patent, two recognition units are used, each unit containing 5 cross-distributed cameras to obtain two groups of images of the product to be detected. The first group of images is used to detect the product, and if the results are not satisfactory, the product is directly rejected; otherwise, the product is flipped and enters the second stage. The second group of images is used for detection again, and if there are the same unqualified items in the two detection results, the product is determined to be unqualified and is rejected.

[0007] In this patent, it is mentioned that the detection of surface defects relies on image processing, but the detection capability for complex textures or small defects is still insufficient. Therefore, there is still a need for improvement in the existing technology. SUMMARY

[0008] In summary, the present application proposes an image fusion recognition method and system based on metal flat wire surface defect detection, which synchronously acquires multi-angle images of metal flat wire under multiple illumination conditions, and fuses the multi-angle images into a comprehensive image with high contrast and low noise, and then performs recognition, thereby significantly improving the accuracy and reliability of defect detection.

[0009] The technical solution of the present application is as follows:

[0010] An image fusion recognition method based on metal flat wire surface defect detection, characterized in that it comprises:

[0011] Step 1: A first camera module and a second camera module are arranged on the upper and lower sides of the metal flat wire to obtain multi-angle original images of the metal flat wire;

[0012] Step2: The image preprocessing unit performs standardization processing on the obtained multi-angle original image to obtain a multi-angle corrected image;

[0013] Step3: The multi-angle corrected image is input into the image fusion unit for fusion processing to finally obtain a comprehensive image;

[0014] Step4: The comprehensive image is input into the layered recognition unit for comparison,

[0015] In Step3, it further includes:

[0016] Step301: The multi-angle corrected image is decomposed into a low-frequency subband and a high-frequency subband;

[0017] Step302: The gradient intensity of the low-frequency subband is calculated as a definition weight; and the fusion is weighted according to the definition weight;

[0018] Step303: The sum of the absolute values of the second-order derivatives of the neighborhood of the pixel points of the high-frequency subband is calculated to quantify the local complexity;

[0019] Step304: The local complexity is input into the PCNN model as an excitation, and the number of pulse outputs is recorded;

[0020] Step305: The source image with the largest number of pulses is selected as the corresponding coefficient for fusion;

[0021] Step306: The fused low-frequency subband and high-frequency subband are reconstructed to generate a comprehensive image.

[0022] In the image fusion method of the present application, Step2 further includes:

[0023] Step201: The multi-angle original image is decomposed into an illumination component and a reflection component, and then logarithmic conversion is performed;

[0024] Step202: The multi-angle original image after logarithmic conversion is converted from a spatial domain pixel position to a frequency domain spatial frequency distribution;

[0025] Step203: In the frequency domain, the high-frequency reflection is enhanced and the low-frequency illumination is suppressed by a filter;

[0026] Step204: The frequency domain data processed by the filter is converted back to a spatial domain image;

[0027] Step205: Exponential conversion is performed to restore the multi-angle original image.

[0028] In the image fusion method of the present application, the filter reduces its strength for the region close to the center of the frequency domain, and increases its strength for the region far from the center of the frequency domain.

[0029] In the image fusion method of the present application, the suppression or enhancement of the filter is controlled by two coefficients, one larger coefficient for enhancing high-frequency reflection details, and one smaller coefficient for suppressing low-frequency illumination.

[0030] In the image fusion method of the present application, the Step 302 specifically includes:

[0031] The low-frequency subband of the amplitude image is set as The definition of the sharpness weight of each image is the square of the gradient of the subband, and the calculation formula is as follows:

[0032]

[0033] Wherein, LL1, LL2,..., LL n n low-frequency subbands obtained after the decomposition of n input images, LLᵢ represents a low-frequency subband containing image information of the corresponding multi-angle original image,

[0034] The sharpness measure of the low-frequency subband LLᵢ of the i-th image accounts for the proportion of the sum of the sharpness measures of the low-frequency subbands of all n images,

[0035] Wherein, Between 0 and 1, and the sum of the n weights is 1,

[0036] Indicates the calculation of the gradient of the low-frequency subband image Iᵢ, which represents the intensity and direction of the change of the pixel value in the image,

[0037] Indicates the sum of the gradient modulus squares of the n low-frequency subband images,

[0038] The n low-frequency subbands are fused, and the calculation formula is:

[0039]

[0040] Wherein, Indicates the fused image of the low-frequency subband.

[0041] In the image fusion method of the present application, the Indicates the square of the gradient modulus, and for each pixel point (x, y) in an image, the gradient modulus square is calculated as: G x(x,y)²+Gᵧ(x,y)², summing up the value of all pixel points in the image, to obtain a scalar value representing the total intensity of the gradient of the entire LLᵢ image.

[0042] In the image fusion method of the application, the high-frequency sub-band is represented as H i , and the calculation formula is as follows:

[0043]

[0044] In the formula, represents the pixel value of the position (x,y) in the high-frequency sub-band at scale i, and represents the edge detail information of the point,

[0045] K is represented as a neighborhood radius, and d is represented as a displacement,

[0046] wherein, in the horizontal direction or the vertical direction, the absolute deviation between the sum of the eigenvalues of two symmetric points (x-d,y) and (x+d,y) with (x,y) as the center and a distance of d and twice the eigenvalue of the center point,

[0047] The high-frequency sub-band is fused by the PCNN model again.

[0048] In the image fusion method of the application, the PCNN model comprises the following:

[0049] External excitation term: , input is in the above formula;

[0050] Internal activity term: , wherein, is represented as the output feedback of the neighborhood neuron, is a connection coefficient, which controls the influence of the output feedback of the neighborhood neuron on the internal activity term,

[0051] Dynamic threshold term: , wherein, is represented as a threshold decay coefficient, which is used to control the exponential decay speed of the threshold; is represented as a threshold amplitude coefficient, which is used to control the updating strength of the threshold; is represented as the pulse output at the previous moment,

[0052] Pulse output term:

[0053]

[0054] When the internal activity term is greater than the dynamic threshold term, the output is 1, otherwise the output is 0.

[0055] In the image fusion method of the present application, a plurality of high-frequency subbands independently run the PCNN model and record the sum, as follows:

[0056]

[0057] wherein NI represents the number of iterations, represents the pulse output of all iterations,

[0058] the fused high-frequency subband coefficient select the source image with the most pulses as the corresponding coefficient, that is, wherein, ,

[0059] H f is output as LH f , HL f and HH f , and reconstruction is further performed to obtain a comprehensive image.

[0060] An image fusion recognition system based on metal flat wire surface defect detection, characterized in that,

[0061] The image fusion recognition system comprises an image acquisition unit, an image preprocessing unit, an image fusion unit and a hierarchical recognition unit.

[0062] The image fusion recognition method and system based on metal flat wire surface defect detection according to the present application have the following beneficial effects:

[0063] 1. The traditional single light source angle is easy to form a strong reflection area on the metal surface, shielding small defects. The present application realizes synchronous / quasi-synchronous imaging through high-angle coaxial grazing light + low-angle lateral diffuse light and upper and lower bilateral cameras. It ensures that different types of defects present high contrast under at least one imaging angle or lighting condition, significantly improving the detectability of defects.

[0064] 2. The present application configures a high-resolution industrial camera and combines multi-angle complementary imaging. Preprocessing is performed before fusion, and PCNN is used to dynamically select high-frequency details in fusion, which significantly improves the feature extraction capability.

[0065] 3. The present application decomposes the image into low-frequency subbands and high-frequency subbands for separate processing. Low-frequency fusion is based on clarity weight weighted fusion, which retains the main structure information with the highest clarity. The sum of the absolute values of the second-order derivatives of the neighborhood is used as the external excitation for high-frequency fusion. Through the dynamic characteristics of PCNN pulse emission, the high-frequency coefficient with the most pulse emissions in multiple iterations is adaptively selected, and the low-frequency uneven illumination and the high-frequency reflection details are enhanced in the frequency domain. It effectively suppresses background noise and uneven illumination interference and highlights defect features. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 The structural connection block diagram of the image fusion recognition system of the present application is shown in the figure.

[0067] Figure 2 The structural schematic diagram of the image acquisition unit of the present application is shown in the figure.

[0068] Figure 3 The structural schematic diagram of the image acquisition unit of the present application is shown in the figure.

[0069] Figure 4 The flow chart of the image fusion recognition method of the present application is shown in the figure.

[0070] Figure 5 The logic block diagram of the image fusion recognition method of the present application is shown in the figure.

[0071] Figure 6 The logic block diagram of the image preprocessing unit of the present application is shown in the figure.

[0072] Figure 7 The logic block diagram of the image fusion unit of the present application is shown in the figure.

[0073] The reference signs are shown as follows: image acquisition unit 10, first camera module 101, second camera module 102, high-angle coaxial grazing light source 103, low-angle lateral diffuse light source 104, image preprocessing unit 20, image fusion unit 30, layered recognition unit 40. DETAILED DESCRIPTION

[0074] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0075] Embodiment one

[0076] Referring to Figure 1 The present embodiment proposes an image fusion recognition system 100 based on metal flat wire surface defect detection, which can detect the small defects on the edge and surface of the metal flat wire in a high-speed motion state, acquire complementary information through multi-angle imaging, and then perform layered recognition after image preprocessing and fusion enhancement, so as to improve the recognition accuracy.

[0077] The image fusion recognition system 100 includes an image acquisition unit 10, an image preprocessing unit 20, an image fusion unit 30, and a layered recognition unit 40.

[0078] The image acquisition unit 10 is the front end of information input, which is responsible for synchronously acquiring multi-view images of the edge region of the metal flat wire under different illumination conditions during the continuous motion of the metal flat wire.

[0079] The image preprocessing unit 20 performs standardization processing on the multi-angle images acquired by the image acquisition unit 10, eliminates noise and uneven illumination, and ensures the subsequent fusion quality.

[0080] The image fusion unit 30 fuses the complementary information of the processed multi-angle images into a comprehensive image with significantly enhanced defect features and suppressed background interference.

[0081] The layered recognition unit 40 further realizes accurate classification and positioning of defects.

[0082] Embodiment Two

[0083] Based on the above-mentioned embodiments, referring to Figures 2 to 7 The present embodiment further proposes an image fusion recognition method based on metal flat wire surface defect detection, specifically including the following steps:

[0084] Step 1: A first camera module 101 and a second camera module 102 are arranged on the upper and lower sides of the metal flat wire, for acquiring a plurality of multi-angle original images of the metal flat wire.

[0085] The image acquisition unit 10 includes a first camera module 101 and a second camera module 102, which are arranged on the upper and lower sides of the metal flat wire, respectively, for acquiring a plurality of multi-angle original images of the metal flat wire.

[0086] It should be noted that the first camera module 101 and the second camera module 102 are configured with at least two high-resolution industrial cameras, preferably linear array or high-speed area array cameras, which are symmetrically arranged on the two sides of the metal flat wire.

[0087] The image acquisition unit 10 further includes a plurality of controllable LED light sources arranged on the two sides and above and below the metal flat wire according to a specific spatial angle.

[0088] The plurality of controllable LED light sources are respectively:

[0089] One or more high-angle coaxial grazing light sources 103 arranged on the upper and lower sides of the metal flat wire are used to suppress specular reflection, cooperate with the first camera module 101 and the second camera module 102, and acquire the overall gray scale distribution of the surface, which is used to identify surface impurities, oxidation spots, etc.

[0090] One or more low-angle lateral diffuse light sources 104 arranged on the left and right sides of the metal flat wire are used to highlight the visibility of edge profiles, small protrusions, and depressions.

[0091] Further, the metal flat wire is arranged along Figure 2The metal flat wire moves in the direction of the arrow, and the speed of the metal flat wire is obtained in real time through the speed encoder signal, and the first camera module 101 and the second camera module 102 are triggered for synchronous or quasi-synchronous image acquisition, so as to ensure that the same section of the metal flat wire is imaged at the same time under different illumination angles.

[0092] Each high-resolution industrial camera in the first camera module 101 and the second camera module 102 can independently acquire multiple images to form a set of multi-angle original image sets (such as I1, I2, I3, I4), which correspond to edge information under different illumination conditions.

[0093] All images are time-stamped and position-coded for subsequent spatial registration and sliding window positioning.

[0094] Multi-angle illumination breaks the limitation of a single perspective, and different types of defects are presented with high contrast in at least one imaging channel, so as to improve the accuracy of subsequent fusion recognition.

[0095] Step 2: The image preprocessing unit 20 performs standardization processing on the obtained multi-angle original image set, such as eliminating noise and uneven illumination, to ensure the quality of subsequent fusion and obtain multi-angle corrected images.

[0096] The multi-angle original image set is decomposed into an illumination component and a reflection component, wherein the illumination component represents slow-changing uneven illumination, and the reflection component represents the edge of the metal flat wire.

[0097] The two components are independently processed, and the formula is as follows:

[0098]

[0099] wherein, is used to convert to an additive form, so as to separate the components, and 1 is added to avoid undefined logarithm of pixel value 0.

[0100] represents converting the image from the spatial domain to the frequency domain;

[0101] represents enhancing reflection details and suppressing uneven illumination through a filter .

[0102] represents converting the processed image to the spatial domain;

[0103] represents reversing the logarithmic transformation to restore the image to the original intensity range.

[0104] Specifically, the filter is represented as:

[0105]

[0106] where, when close to the frequency center, , the light is suppressed; when far from the frequency center, , the reflection details are enhanced.

[0107] represents the Euclidean distance from the frequency point to the frequency midpoint, calculated as:

[0108]

[0109] M and N are the height and width of the image, respectively, in pixels. D0 is the cutoff frequency, generally D0=0.05×min(M,N), used to ensure that the filter is adaptive to the image size.

[0110] and are enhancement coefficients, where, and require , , generally takes a value of 2.0, takes a value of 0.5.

[0111] c represents the filter shape parameter, which controls the steepness of the transition band. Common values are c=1 or c=2, and the larger the value, the sharper the transition.

[0112] In this embodiment, first, the original image is subjected to mathematical operations, converting the image model originally composed of the multiplication of light and reflection components into the form of the addition of the two components, which is more convenient for subsequent separation processing, and avoids the calculation problem caused by the original pixel value being zero.

[0113] Then, the image after logarithmic conversion is converted from the spatial domain pixel position to the frequency domain spatial frequency distribution.

[0114] In the frequency domain, processing is performed through a filter, and the working principle of the filter is:

[0115] For the area close to the frequency domain center, representing the low-frequency component that changes slowly in the image, mainly corresponding to uneven lighting, the filter will be suppressed to reduce its intensity.

[0116] For the area far from the frequency domain center, representing the high-frequency component that changes sharply in the image, such as details or edges, mainly corresponding to reflection information, the filter will be enhanced to increase its intensity.

[0117] The degree of filter suppression and enhancement is controlled by two coefficients: a larger coefficient is used to enhance high-frequency reflection details, and a smaller coefficient is used to suppress low-frequency illumination.

[0118] The filter shape parameter control can ensure a smooth transition between the low-frequency suppression area and the high-frequency enhancement area. At the same time, the cutoff frequency of the filter can be automatically adjusted according to the size of the image itself, ensuring that the processing effect adapts to images of different sizes.

[0119] The frequency domain data processed by the filter is converted back to a spatial domain image.

[0120] A reverse mathematical operation is performed on the data converted back to the spatial domain to offset the initial logarithmic conversion and restore the intensity range of the image to its original state.

[0121] The final output of the corrected image effectively suppresses the effects of uneven illumination while significantly enhancing the reflection components of the metal flat wire edges, improving the clarity and consistency of the image and providing high-quality input for subsequent multi-angle image fusion.

[0122] Step 3: Input the multi-angle corrected image into the image fusion unit 30 for fusion processing to obtain a comprehensive image.

[0123] Where each image is decomposed into a low-frequency subband and a high-frequency subband, and the input image is set to After decomposition, we get:

[0124]

[0125] In the formula, LL represents the low-frequency subband; LH, HL, and LL represent the high-frequency subband, representing the horizontal, vertical, and diagonal high-frequency details, respectively.

[0126] In this embodiment, the low-frequency subband is fused using the sharpness weight, and the low-frequency subband of the image is set to The sharpness weight of each image is defined as the gradient square of the subband, and the calculation formula is as follows:

[0127]

[0128] Where LL1, LL2,..., LL n n low-frequency subbands obtained after discrete wavelet transform of n input images. A certain LLᵢ contains the main outline, general structure, and average brightness information of the corresponding original image Iᵢ.

[0129] The sharpness measure of the low-frequency subband LLᵢ of the i-th image accounts for the proportion of the sum of the sharpness measures of all n low-frequency subbands. between 0 and 1, and the sum of n weights is 1. Reflects the clarity of the i-th image in the low-frequency part relative to other images. The larger, the more clear LLᵢ is, and the greater the contribution in the fusion result.

[0130] represents the gradient of the low-frequency sub-band image Iᵢ, which reflects the intensity and direction of the change of pixel values in the image, that is, where the change is fast and where the change is slow.

[0131] Usually, the Sobel operator is used to calculate the gradient. The Sobel operator will calculate the gradient approximation of the image in the horizontal direction G x and the vertical direction Gᵧ respectively.

[0132] represents the square of the gradient modulus. For each pixel point (x, y) in an image, the square of the gradient modulus is usually calculated as: G x (x, y)²+Gᵧ(x, y)². The sum of this value of all pixel points in the image is obtained. A scalar value representing the total intensity of the gradient of the entire LLᵢ image.

[0133] represents the sum of the squares of the gradients of the n low-frequency sub-band images.

[0134] In this embodiment, the n low-frequency sub-bands are fused, and the calculation formula is:

[0135]

[0136] wherein, represents the fused image of the low-frequency sub-band. According to the actual clarity contribution of each image in the low-frequency part, the weight of each image is adjusted. Specifically:

[0137] If the gradient of a certain region of image I1 in the low-frequency sub-band LL1 is large, large, then the pixel value of the region in LL1 in the corresponding region of dominates;

[0138] If the gradient of a certain region of image I2 in the low-frequency sub-band LL2 is small, large, then the pixel value of the region in LL2 in the corresponding region of contribution is very small.

[0139] In this embodiment, the local structure complexity of each pixel point in the high-frequency sub-band LH, HL and LL is measured by calculating the sum of the absolute values of the second-order derivatives of the pixel point and its neighborhood in the horizontal and vertical directions, so as to enhance the edge and detail information.

[0140] For high frequency subband, take any one direction of LH, HL and LL as an example, denoted as H i , the calculation formula is as follows:

[0141]

[0142] In the formula, denotes the pixel value of position (x, y) in high frequency subband at scale i, representing the edge detail information of the point.

[0143] K represents the neighborhood radius, usually taking 1 or 2; by calculating the symmetry destruction degree of the pixel point and its left / right / upper / lower symmetric neighborhood, the local edge and texture features are captured.

[0144] d represents the offset, the summation is from d = -K to K, which represents the symmetric neighborhood considering the distance of d in horizontal and vertical directions.

[0145] In the embodiment, the above formula is composed of two parts, which respectively measure the local difference in horizontal and vertical directions.

[0146] Among them, the horizontal direction is: .

[0147] In the horizontal direction, the absolute deviation between the sum of the eigenvalues of the two symmetric points (x-d, y) and (x+d, y) with distance d from (x, y) as the center and twice the eigenvalue of the center point.

[0148] If H i In the flat area, then , the term tends to 0, if there is an edge or texture, the value of the term becomes large, which is used to detect local curvature.

[0149] Vertical direction: , similar to the horizontal direction.

[0150] It should be noted that the summation part is from -K to K, but the actual offset d can be negative, therefore, d in the above formula is the absolute value. And if d = 0, the term is meaningless, therefore, the absolute value range of the actual effective d is .

[0151] Subsequently, the PCNN model is further used to fuse the high frequency subband.

[0152] The PCNN model includes the following:

[0153] External excitation term: , the input is in the above formula, which directly reflects the complexity of local structure;

[0154] Internal activity term: where, denotes the output feedback of the neighborhood neuron, is the connection coefficient, which controls the influence of the output feedback of the neighborhood neuron on the internal activity item.

[0155] Dynamic threshold item: where, denotes the threshold decay coefficient, which is used to control the exponential decay rate of the threshold; denotes the threshold amplitude coefficient, which is used to control the update strength of the threshold; denotes the pulse output at the previous moment.

[0156] Pulse output item:

[0157]

[0158] When the internal activity item is greater than the dynamic threshold item, the output is 1, otherwise the output is 0.

[0159] Then, for each pixel position in all high-frequency subbands LH, HL and LL, the PCNN model is independently run and the sum is recorded, and the formula is as follows:

[0160]

[0161] where, NI represents the number of iterations, usually 3-5 times; denotes the cumulative pulse output of all iterations. The more the output times, the stronger the significant features contained in the corresponding source image.

[0162] Fused high-frequency subband coefficients Select the source image with the most pulse times as the corresponding coefficient, that is, where, . The adaptive decision retains the most significant local features, which improves the limitation compared with the traditional maximum value or weighted average.

[0163] Finally, H f is output as LH f , HL f and HH f , and then reconstructed to obtain the comprehensive image I 综合 , as follows:

[0164]

[0165] In this embodiment, the corrected image input fusion unit decomposes each image into a low-frequency subband and a high-frequency subband;

[0166] The gradient intensity of each image low-frequency subband is calculated as the sharpness weight; the most clear structure information is retained by weighted fusion according to the sharpness weight;

[0167] The sum of the second derivative absolute values of the neighborhood of each image high-frequency sub-band pixel point is calculated to quantify the local structure complexity;

[0168] The local complexity is taken as the excitation input for dynamic analysis by the PCNN model, the significant features are recognized according to the pulse output times, the high-frequency coefficients of the source image with more pulse output times are selected, and the most prominent edge and texture details are reserved;

[0169] The integrated image is generated by reconstructing the fused low-frequency sub-band and high-frequency sub-band, and the surface defect features are reserved.

[0170] Step 4: input the integrated image into the layered recognition unit 40 for positioning, defect classification and comparison.

[0171] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An image fusion recognition method based on metal flat wire surface defect detection, characterized in that, Comprise: Step1: set the first camera module and the second camera module on the upper and lower sides of the metal flat wire, and obtain multi-angle original images of the metal flat wire; Step2: the image preprocessing unit standardizes the obtained multi-angle original images to obtain multi-angle corrected images; Step3: input the multi-angle corrected image to the image fusion unit for fusion processing, and finally obtain a comprehensive image; Step4: input the comprehensive image into the hierarchical recognition unit for comparison, Wherein, Step3 further comprises: Step301: decompose the multi-angle corrected image into a low-frequency subband and a high-frequency subband; Step302: calculate the gradient intensity of the low-frequency subband as the definition weight; weighted fusion according to the definition weight; Step303: calculate the sum of the second derivative absolute values of the high-frequency subband pixel neighborhood to quantify the local complexity; Step304: input the local complexity into the PCNN model as the excitation, and record the pulse output times; Step305: select the source image with the most pulse times as the corresponding coefficient for fusion; Step306: reconstruct the fused low-frequency subband and high-frequency subband to generate a comprehensive image.

2. The image fusion recognition method of claim 1, wherein, The Step2 further comprises: Step201: decompose the multi-angle original image into illumination component and reflection component, and then perform logarithmic conversion; Step202: convert the multi-angle original image after logarithmic conversion from spatial domain pixel position to frequency domain spatial frequency distribution; Step203: in the frequency domain, enhance the high-frequency reflection and suppress the low-frequency illumination through the filter; Step204: convert the frequency domain data processed by the filter back to the spatial domain image; Step205: perform exponential conversion to restore the multi-angle original image.

3. The image fusion recognition method of claim 2, wherein, The filter reduces the intensity of the area close to the center of the frequency domain, and increases the intensity of the area far from the center of the frequency domain.

4. The image fusion recognition method of claim 2, wherein, The suppression or enhancement of the filter is controlled by two coefficients, one larger coefficient is used to enhance the high-frequency reflection details, and one smaller coefficient is used to suppress the low-frequency illumination.

5. The image fusion recognition method of claim 1, wherein, The Step302 specifically comprises: The low frequency subband of the set of images is set as The definition of the sharpness weight of each image is the square of the gradient of the subband, and the calculation formula is as follows: wherein LL1, LL2,..., LL n denote n low-frequency subbands obtained after decomposition of n input images, LLᵢ denotes a certain one containing image information of the corresponding multi-angle original image, a ratio of a sharpness measure of a low frequency subband LLi of the i-th image to a sum of all n image low frequency subband sharpness measures, wherein, between 0 and 1, and the sum of the n weights is 1, represents computing for the low frequency sub-band image Iᵢ its gradient, which represents the intensity and direction of the change of pixel values in the image, denotes the sum of the gradient modulus squared of the n low frequency subband images, n low-frequency subbands are fused, and the calculation formula is: wherein, represents a fusion image of the low-frequency sub-band.

6. The image fusion recognition method of claim 5, wherein, The The square of the gradient modulus is denoted by G, which is calculated for each pixel (x, y) in an image as follows: G x (x,y)²+Gᵧ(x,y)², and the sum of this value for all pixels in the image gives a scalar value representing the total gradient intensity of the entire LLᵢ image.

7. The image fusion recognition method of claim 1, wherein, The high frequency subband representation is H i with the following formula: wherein represents the pixel value at position (x, y) in the high frequency subband at scale i, representing the edge detail information at that point, K represents the neighborhood radius, and d represents the offset, Wherein, in the horizontal direction or the vertical direction, the absolute deviation between the sum of the eigenvalues of two symmetric points (x−d, y) and (x+d, y) with (x, y) as the center and the distance d, and twice the eigenvalue of the center point, And fuse the high-frequency subband by the PCNN model.

8. The image fusion recognition method of claim 7, wherein, The PCNN model comprises the following: External excitation term: , input is the above formula in ; internal activity item: wherein, is an output feedback of the neighborhood neuron, is a connection coefficient controlling the influence of the output feedback of the neighborhood neuron on the internal activity item, Dynamic threshold item: Wherein, Threshold attenuation coefficient, used to control the exponential decay rate of the threshold; Threshold amplitude coefficient, used to control the update strength of the threshold; The pulse output at the previous moment, Pulse output term: When the internal activity term is greater than the dynamic threshold term, the output is 1, otherwise the output is 0.

9. The image fusion recognition method of claim 8, characterized in that, A plurality of high-frequency subbands independently run the PCNN model and record the sum, and the formula is as follows: where NI indicates the number of iterations, represents the cumulative pulse output of all iterations, Fused high frequency subband coefficients The source image with the most number of pulses is selected as the corresponding coefficient, i.e. wherein, , H f output as LH f , HL f and HH f , and then reconstructing to obtain a final integrated image.

10. An image fusion recognition system based on metal flat wire surface defect detection, comprising the image fusion recognition method of claim 1, characterized in that The image fusion recognition system comprises an image acquisition unit, an image preprocessing unit, an image fusion unit and a hierarchical recognition unit.

Images