Intelligent detection method for high-precision welding of an aircraft aluminum condenser

Abstract

The application relates to the technical field of image processing, and discloses an intelligent detection method for high-precision welding of an aircraft aluminum condenser, which comprises the following steps: acquiring an image of a welding area to be detected and preprocessing the image to obtain a welding area sub-image; calculating a gradient amplitude graph and a local fractal dimension graph; taking the product of a pixel gradient value and a corresponding pixel value as a composite density function, and calculating composite density weighted Zernike moments of various orders; adjusting the composite density weighted Zernike moments of various orders to generate direction-enhanced composite Zernike features; convolving the welding area sub-image to generate a group of Gabor energy feature graphs, and calculating multi-scale Gabor-Hu texture features; fusing the Zernike features and the multi-scale Gabor-Hu texture features to form a high-dimensional composite feature vector, and inputting the high-dimensional composite feature vector into a pre-trained classification model to output a quality detection category of the welding area. The application combines the gradient amplitude and the local fractal dimension, so that the Zernike moments can simultaneously describe the macroscopic contour and the microscopic details of a weld, and the accuracy of welding defect detection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to an intelligent detection method for high-precision welding of an aircraft aluminum condenser. BACKGROUND

[0002] The aircraft aluminum condenser is a heat exchange component in the aircraft environmental control system, and its quality is related to the stable operation of the entire system and even flight safety. Welding is a process in the manufacturing process of the condenser. Due to the characteristics of aluminum alloy materials and the complexity of the welding process, various defects such as pores, cracks, incomplete fusion, welding bumps and inconsistent surface topography are prone to occur in the welding area. These defects can seriously affect the sealing, pressure resistance and heat exchange efficiency of the condenser, and constitute a major safety hazard. The detection methods in the prior art mainly rely on X-ray, ultrasonic flaw detection and other means. Although the detection methods such as X-ray and ultrasonic flaw detection have high precision, they have problems such as expensive equipment, complex operation, long detection period, and insensitivity to certain types of surface or near-surface defects.

[0003] The intelligent detection method based on image processing has the advantages of non-contact, high efficiency and high objectivity in the field of welding quality detection, and can extract stable and high-discrimination features from weld images with complex texture and uneven reflection. In the existing machine vision-based welding detection technology, feature extraction is the key link to determine the detection performance. The existing machine vision-based welding detection technology has tried various feature descriptors, such as texture features such as gray level co-occurrence matrix and local binary pattern, and shape features such as Hough transform and edge contour. However, a single type of feature often cannot comprehensively describe the diversity and complexity of welding defects.

[0004] As a classic regional shape descriptor, the moment function is applied to image analysis due to its good geometric invariance. Among them, Zernike moments perform well in target recognition due to their orthogonality and rotational invariance, but standard Zernike moments treat the image as a uniform density distribution when calculating, ignoring important gradient and texture details inside the weld area, and are difficult to effectively represent the anisotropic features of the weld along the welding direction. Although Hu invariant moments have translation, scale and rotation invariance, as a global feature descriptor, they have limited ability to describe local information with rich texture details on the uneven surface of the weld. In addition, Gabor filters are very effective in texture analysis, but how to effectively integrate Gabor filters with shape features to improve the comprehensive ability to detect complex welding defects is still a current difficulty. Therefore, how to couple global features that can reflect the contour shape of the weld with local features that can depict surface texture and directionality is a technical problem that needs to be solved in the field of intelligent detection of aircraft condenser welding quality. SUMMARY

[0005] This invention provides an intelligent detection method for high-precision welding of aircraft aluminum condensers to solve the problem in the prior art that it is difficult to couple global features that reflect the weld contour morphology with local features that can characterize surface texture and directionality.

[0006] The intelligent detection method for high-precision welding of aircraft aluminum condensers of the present invention includes the following steps:

[0007] An image of the welding area of ​​the condenser to be inspected is acquired. The image is preprocessed to obtain a normalized welding area sub-image. The gradient magnitude map of the welding area sub-image is calculated, and the local fractal dimension map of the pixels in the gradient magnitude map is calculated using box counting. Taking the geometric center of the welding area sub-image as the origin, the product of the pixel gradient value in the gradient magnitude map and the corresponding pixel value in the local fractal dimension map is used as the composite density function. The composite density weighted Zernike moments of each order are calculated. A non-uniform angular frequency weight matrix matching the main welding direction is set, and the composite density weighted Zernike moments of each order are adjusted using this non-uniform angular frequency weight matrix to enhance the connection with the main welding direction. A directional enhancement composite Zernike feature is generated from the associated angular frequency components. A set of Gabor filter banks with different center frequencies and directional angles are used to convolve the sub-image of the welding area to generate a set of Gabor energy feature maps. Seven Hu invariant moments are calculated for each Gabor energy feature map. All Hu invariant moments are concatenated to form a multi-scale Gabor-Hu texture feature. The directional enhancement composite Zernike feature is then spliced ​​and fused with the multi-scale Gabor-Hu texture feature to form a high-dimensional composite feature vector. The high-dimensional composite feature vector is input into a pre-trained classification model, and the classification model outputs the quality detection category of the welding area.

[0008] Preferably, the preprocessing of the image to obtain a normalized welding area sub-image includes: converting the acquired image into a grayscale image; and using a... The median filter of the window is used to filter the grayscale image; the Canny operator is used to perform edge detection on the filtered grayscale image to identify the weld contour; based on the weld contour, a 256×256 pixel welding area sub-image is cropped from the grayscale image, and the pixel values ​​of the welding area sub-image are linearly normalized to the [0, 1] interval.

[0009] Preferably, the step of calculating the local fractal dimension map of pixels in the gradient magnitude map using box counting includes: for each pixel in the gradient magnitude map, taking a box as the center of the pixel. An analysis window of the pixel; in the analysis window, a set of boxes with sizes S of {2, 4, 8, 16} are set; for each box size S, the minimum number of boxes required to cover all non-zero gradient pixels in the analysis window is calculated ; in a double logarithmic coordinate system with as the abscissa and as the ordinate, the data points are linearly fitted by the least squares method, and the absolute value of the slope of the obtained straight line is taken as the local shape dimension of the center pixel.

[0010] Preferably, the adjusting the complex density weighted Zernike moments of each order by using the non-uniform angular frequency weight matrix comprises: performing Radon transform on the gradient amplitude map to determine the welding main direction ; multiplying each order of the complex density weighted Zernike moment by a rotation factor to achieve rotation normalization; and weighting the rotation-normalized Zernike moments by using the non-uniform angular frequency weight matrix.

[0011] Preferably, the Gabor filter bank is composed of 5 center frequencies and 8 direction angles, wherein the center frequencies are {1 / 4, 1 / 8, 1 / 16, 1 / 32, 1 / 64}, and the unit of the center frequency is cycle / pixel, and the direction angles are {0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°}.

[0012] Preferably, the constructing the multi-scale Gabor-Hu texture feature comprises: for each of the group of Gabor energy feature maps, first-order to seventh-order Hu invariant moments are calculated respectively; and all the calculated Hu invariant moments are concatenated into a one-dimensional feature vector in a predetermined order.

[0013] Preferably, the classification model is a random forest classification model.

[0014] Preferably, the classification model outputs the quality detection category of the welding area, comprising: the random forest classification model has 500 decision trees, and the maximum depth of each decision tree is 25, and the node splitting criterion is Gini impurity; and the quality detection category output by the random forest classification model is "qualified", "surface porosity", "unwelded" or "undercut".

[0015] Preferably, the acquiring the image of the welding area of the condenser to be detected comprises: acquiring an RGB image of the welding area of the condenser to be detected by using a camera with a coaxial light source.

[0016] Preferably, the pre-processing of the image to obtain a normalized welding area sub-image comprises: converting the RGB image into a gray image, smoothing and denoising by using a Gaussian filter, cutting and uniformly scaling the welding area in the gray image to 256*256 pixel size by using a bilinear interpolation algorithm to obtain a welding area sub-image, and linearly normalizing the pixel value of the welding area sub-image to the interval [0, 1].

[0017] The present application has the advantages that: the present application combines the gradient amplitude reflecting the weld edge information with the local shape dimension representing the surface texture complexity by constructing the composite density weighted Zernike moment, so that the improved Zernike moment can describe the macro profile and micro details of the weld at the same time; in addition, the directionality of the Zernike moment is enhanced and adjusted by introducing the non-uniform angular frequency weight matrix matched with the main direction of welding, which improves the sensitivity of the feature to the directional defects such as cracks and incomplete fusion. A multi-scale Gabor-Hu texture feature is also constructed, the Hu invariant moment is calculated by a group of Gabor energy maps, and the surface roughness and weld texture details of the weld are obtained. The direction-enhanced shape feature and the multi-scale invariant texture feature are fused to form a composite feature with complementary information and strong discrimination, which reduces the influence of the limitations of single feature and improves the accuracy of the detection of complex welding defects of the aircraft aluminum condenser. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The present application provides an intelligent detection method for high-precision welding of an aircraft aluminum condenser. DETAILED DESCRIPTION

[0019] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0020] As Figure 1 shown, the intelligent detection method for high-precision welding of an aircraft aluminum condenser provided by the present application specifically comprises the following steps:

[0021] S1, an image of a welding area of a condenser to be detected is obtained, the image is pre-processed to obtain a normalized welding area sub-image; a gradient amplitude map of the welding area sub-image is calculated, and a local shape dimension map of the pixels in the gradient amplitude map is calculated by using box counting method.

[0022] Specifically, an RGB image of a welding area of a condenser to be detected is collected by using a camera matched with an axial light source, the RGB image is converted into a gray image, the gray image is smoothed and denoised by using a Gaussian filter, the welding area in the gray image is cropped and uniformly scaled to 256*256 pixels by using a bilinear interpolation algorithm, and a welding area sub-image is obtained, and pixel values of the welding area sub-image are linearly normalized to the interval [0, 1].

[0023] Gradients of the normalized welding area sub-image in x and y directions are calculated respectively, and a gradient amplitude graph is synthesized. With each pixel in the gradient amplitude graph as a center, a pixel neighborhood is used to cover the neighborhood, the number of required boxes is counted, for example, the number of boxes of 1*1, 2*2, 4*4, 8*8, 16*16, 32*32, 64*64, 128*128, 256*256, 512*512, 1024*1024, 2048*2048, and 4096*4096. The slope of a straight line obtained by linear regression fitting of the logarithm of the box size and the logarithm of the number of boxes is the local fractal dimension of the center pixel, and a local fractal dimension graph is obtained.

[0024] S2, with the geometric center of the welding area sub-image as an origin, a product of a pixel gradient value in the gradient amplitude graph and a corresponding pixel value in the local fractal dimension graph is taken as a compound density function, and compound density weighted Zernike moments of each order are calculated.

[0025] Specifically, corresponding pixel values in the gradient amplitude graph and the local fractal dimension graph are multiplied point by point to obtain a compound density function image. Coordinates of the welding area sub-image are mapped into a unit disc, according to a definition formula of Zernike moments, the compound density function is used to replace an image gray function in a traditional definition, and all Zernike moments from 0 order to a preset highest order are calculated.

[0026] S3, a non-uniform angular frequency weight matrix matched with a welding main direction is set, the non-uniform angular frequency weight matrix is used to adjust the compound density weighted Zernike moments of each order, an angular frequency component associated with the welding main direction is enhanced, and a direction-enhanced compound Zernike feature is generated; a group of Gabor filters with different center frequencies and direction angles are used to convolve the welding area sub-image, a group of Gabor energy feature graphs is generated, seven Hu invariant moments are calculated for each Gabor energy feature graph, all Hu invariant moments are cascaded, and a multi-scale Gabor-Hu texture feature is constructed.

[0027] ​​​​For example, assuming the main welding direction is horizontal, a non-uniform angular frequency weighting matrix is ​​constructed. The rows of the matrix correspond to the order n of the Zernike moment, and the columns correspond to the angular frequency m. Larger weighting coefficients are assigned to angular frequency components strongly correlated with the horizontal direction (i.e., columns with smaller m values), while smaller weighting coefficients are assigned to other angular frequency components. The magnitude of the composite density-weighted Zernike moment of each order is multiplied by the corresponding weighting coefficient in the non-uniform angular frequency weighting matrix to obtain the direction-enhanced composite Zernike eigenvector.

[0028] For example, a filter bank containing 5 center frequencies and 8 directional angles, totaling 40 Gabor filters, is designed. Each filter in the filter bank is convolved with a sub-image of the welding area, and the amplitude of the convolution result is taken to generate 40 Gabor energy feature maps. For each of these 40 Gabor energy feature maps, Hu invariant moments from the first to the seventh order are calculated. All the calculated Hu invariant moments are concatenated sequentially into a one-dimensional vector with a total dimension of 280, resulting in multi-scale Gabor-Hu texture features.

[0029] S4, the directional enhancement composite Zernike feature and the multi-scale Gabor-Hu texture feature are spliced ​​and fused to form a high-dimensional composite feature vector. The high-dimensional composite feature vector is then input into a pre-trained random forest classification model, which outputs the quality detection category of the welding area.

[0030] Specifically, the generated directional enhancement composite Zernike feature vector is concatenated end-to-end with the generated multi-scale Gabor-Hu texture feature vector to form a higher-dimensional composite feature vector. This higher-dimensional composite feature vector is then fed into a random forest classification model trained on a large dataset of weld samples with various labels including acceptable, porosity, and cracks. The random forest model consists of hundreds of decision trees. By combining the voting results from all decision trees, the quality inspection category of the weld area is output, such as acceptable or specific defect type.

[0031] In an optional embodiment, preprocessing the image to obtain a normalized welding area sub-image includes: converting the acquired image into a grayscale image; and using a... The median filter of the window filters the grayscale image; the Canny operator is used to perform edge detection on the filtered grayscale image to identify the weld contour; based on the weld contour, a 256×256 pixel welding area sub-image is cropped from the grayscale image, and the pixel values ​​of the welding area sub-image are linearly normalized to the [0, 1] interval.

[0032] The median filter takes each pixel and its 24 neighboring pixels, totally 25 pixels, and replaces the center pixel's value with the median value of these 25 pixels. It can effectively remove noise while preserving the clarity of the weld edge. After the image is denoised, the Canny edge detection algorithm is used to accurately locate the boundary of the weld. The Canny edge detection algorithm can identify the areas with sharp changes in brightness in the image and generate a clear weld contour. Based on this weld contour, the core position of the weld is determined, and a weld area sub-image is cropped around the core position. The gray values of all pixels in the weld area sub-image are linearly normalized to the interval [0, 1]. Normalization can eliminate the effects of uneven lighting and enhance the adaptability of the image under different shooting conditions.

[0033] In an optional embodiment, the method of calculating the local fractal dimension map of pixels in the gradient magnitude map using the box counting method comprises: for each pixel in the gradient magnitude map, taking a pixel analysis window centered on the pixel; in the analysis window, a set of boxes with sizes S = {2, 4, 8, 16} is set; for each box size S, the minimum number of boxes required to cover all non-zero gradient pixels in the analysis window is calculated. In a double logarithmic coordinate system with as the abscissa and as the ordinate, the data points are linearly fitted by the least squares method, and the absolute value of the slope of the resulting straight line is taken as the local fractal dimension of the center pixel.

[0034] For example, for any pixel point on the gradient magnitude map, an analysis window, for example, a pixel area, around the pixel point is obtained. First, the smallest pixel box is used to cover the window, and the number of such boxes required to cover all non-zero gradient value pixels is counted, which is denoted as N2. Then, a larger size box, for example, , is used, and the above covering and counting process is repeated to obtain a series of box numbers corresponding to different box sizes, such as N4, N8, N 16 . These data points with the reciprocal of the logarithmic value of the box size as the abscissa and the logarithmic value of the box number as the ordinate are approximately a straight line on the double logarithmic coordinate graph. By least squares fitting of the straight line, the absolute value of the slope of the straight line is calculated, which is the local fractal dimension of the center pixel, reflecting the texture fineness of the local area of the center pixel. Repeat the process for each pixel to generate a complete local fractal dimension map.

[0035] In an optional embodiment, the adjusting the complex density-weighted Zernike moments of each order by using the non-uniform angular frequency weight matrix comprises: performing Radon transform on the gradient magnitude map to determine the welding main direction ; multiplying each order of the complex density-weighted Zernike moment by a rotation factor to achieve rotation normalization; and weighting the rotation-normalized Zernike moment by using the non-uniform angular frequency weight matrix.

[0036] In order to make the extracted shape features not affected by the shooting angle of the weld in the image, rotation normalization processing is performed. By performing Radon transform on the gradient magnitude map, the most prominent direction of the linear feature in the image is detected, that is, the welding main direction is determined, for example, the angle thereof is calculated to be 30 degrees. Each order of the complex density-weighted Zernike moment is multiplied by a complex rotation factor , and the complex rotation factor is determined according to the angle of the welding main direction.

[0037] After rotation normalization, all the welds are aligned to the same reference. However, different Zernike moment components have different importance in describing the weld defects. For example, some high-order moments are more sensitive to the description of the undercut and other high-frequency defects. Further, the non-uniform angular frequency weight matrix is used to weight each Zernike moment component after rotation normalization. For example, a moment representing a key defect morphology is given a weight greater than 1, and a moment irrelevant to the defect is given a weight less than 1, highlighting the most valuable feature information for classification and improving the detection accuracy.

[0038] In an optional embodiment, the Gabor filter bank is composed of 5 center frequencies and 8 direction angles, wherein the center frequencies are {1 / 4, 1 / 8, 1 / 16, 1 / 32, 1 / 64}, the unit of the center frequency is cycle / pixel, and the direction angles are {0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°}.

[0039] The filter bank is used to capture the texture information of the weld from multiple dimensions, and the filter bank is composed of a total of 40 independent Gabor filters. The 5 center frequencies determine the sensitivity of the filter to the texture scale. For example, a high-frequency filter of 1 / 4 cycle / pixel can detect very fine and dense texture changes in the image, and a low-frequency filter of 1 / 64 cycle / pixel is more sensitive to large-scale and gentle texture patterns.

[0040] The eight direction angles enable the filter set to respond to texture features of different orientations, starting from 0 degree, i.e. the horizontal direction, with a step of 22.5 degrees, covering a range of 180 degrees, ensuring that the weld texture details in the horizontal, vertical or any diagonal direction can be effectively acquired. Convolution of the image with each of the 40 filters will obtain 40 energy feature maps, each of which highlights the texture response strength of the original image in a specific scale and direction, constituting a comprehensive weld texture feature.

[0041] In an optional embodiment, the classification model outputs a quality detection category of the welding area, including: the random forest classification model has 500 decision trees, and the maximum depth of each decision tree is 25, and the node splitting criterion is Gini impurity; the quality detection category output by the random forest classification model is "qualified", "surface porosity", "lack of penetration" or "undercut".

[0042] The random forest classification model contains 500 decision trees, which can analyze the input feature vector from 500 different perspectives, making the analysis result more accurate than a single decision tree. The maximum depth of each decision tree is 25 layers, which can prevent a single tree from becoming too complex and causing overfitting, ensuring the generalization ability of the random forest classification model to new samples.

[0043] During the training of each decision tree, when it is necessary to divide the data at a certain node, Gini impurity is used as an evaluation criterion to select which feature and threshold can bring the best classification effect. When the multi-scale Gabor-Hu texture feature vector is input into this trained model, each of the 500 trees will give an independent classification judgment, and the model will adopt the voting principle of minority obeying majority to count the prediction results of all decision trees, and the category with the most votes will be output as the final quality detection category.

[0044] The implementation principle of the intelligent detection method for high-precision welding of the aircraft aluminum condenser according to the embodiments of the present application is as follows: by constructing a composite density weighted Zernike moment, the gradient amplitude reflecting the weld edge information is combined with the local shape dimension representing the surface texture complexity, so that the improved Zernike moment can describe the macro profile and micro details of the weld at the same time. Moreover, by introducing a non-uniform angular frequency weight matrix matched with the main direction of welding, the directionality of the Zernike moment is enhanced and adjusted, and the sensitivity of the feature to the directional defects such as cracks and incomplete fusion is improved. In addition, a multi-scale Gabor-Hu texture feature is constructed, the Hu invariant moment is calculated by Gabor energy map, and the details of the weld surface roughness and weld wave texture are obtained. The direction-enhanced shape feature and the multi-scale invariant texture feature are fused to form a composite feature with strong complementary information and discrimination, and the accuracy of the detection of the complex welding defects of the aircraft aluminum condenser is improved.

[0045] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. An intelligent detection method for high-precision welding of an aircraft aluminum condenser, characterized in that, The method comprises the following steps: acquiring an image of a welding area of a condenser to be detected, pre-processing the image to obtain a normalized welding area sub-image; calculating a gradient amplitude map of the welding area sub-image, calculating a local fractal dimension map of pixels in the gradient amplitude map by using a box counting method; taking the geometric center of the welding area sub-image as the origin, and taking the product of the pixel gradient value in the gradient amplitude map and the corresponding pixel value in the local fractal dimension map as a compound density function, calculating compound density weighted Zernike moments of each order; setting a non-uniform angular frequency weight matrix matched with a main welding direction, adjusting the compound density weighted Zernike moments of each order by using the non-uniform angular frequency weight matrix, enhancing the angular frequency components associated with the main welding direction, and generating a direction-enhanced compound Zernike feature; convolving the welding area sub-image by using a group of Gabor filters with different center frequencies and direction angles to generate a group of Gabor energy feature maps, calculating seven Hu invariant moments for each Gabor energy feature map, concatenating all the Hu invariant moments to form a multi-scale Gabor-Hu texture feature; splicing and fusing the direction-enhanced compound Zernike feature and the multi-scale Gabor-Hu texture feature to form a high-dimensional compound feature vector, and inputting the high-dimensional compound feature vector into a pre-trained classification model to output a quality detection category of the welding area.

2. The intelligent detection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The pre-processing of the image to obtain a normalized welding area sub-image comprises: converting the acquired image into a gray-scale image; filtering the gray-scale image using a median filter of a window; performing edge detection on the filtered gray-scale image using a Canny operator to identify a weld contour; cropping a 256x256 pixel welding area sub-image from the gray-scale image according to the weld contour; and linearly normalizing pixel values of the welding area sub-image to the interval [0, 1]. The pre-processing of the image to obtain a normalized welding area sub-image comprises: converting the acquired image into a gray-scale image; filtering the gray-scale image using a median filter of a window; performing edge detection on the filtered gray-scale image using a Canny operator to identify a weld contour; cropping a 256x256 pixel welding area sub-image from the gray-scale image according to the weld contour; and linearly normalizing pixel values of the welding area sub-image to the interval [0, 1].

3. The intelligent detection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The method comprises: for each pixel in the gradient amplitude graph, taking an analysis window centered on the pixel; in the analysis window, setting a group of boxes with sizes S being {2, 4, 8, 16}; for each box size S, calculating the minimum number of boxes required to cover all non-zero gradient pixels in the analysis window ; in a double logarithmic coordinate system with the number of boxes as the horizontal coordinate and the logarithm of the minimum number of boxes as the vertical coordinate, performing a least square linear fitting on the data points, and taking the absolute value of the slope of the obtained straight line as the local fractal dimension of the center pixel. ​​​ 4. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The adjusting of the compound density weighted Zernike moments of each order by using the non-uniform angular frequency weight matrix comprises: Radon transform of the gradient magnitude map to determine the main direction of the weld ; weighting the complex density of each order of zernike moments multiplying by a rotation factor to achieve rotation normalization; The non-uniform angular frequency weight matrix is used to weight the rotated and normalized Zernike moments.

5. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The Gabor filter group is composed of five center frequencies and eight direction angles, wherein the center frequencies are {1 / 4, 1 / 8, 1 / 16, 1 / 32, 1 / 64}, and the unit of the center frequencies is cycle / pixel; and the direction angles are {0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, 157.5°}.

6. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The forming of the multi-scale Gabor-Hu texture feature comprises: The first-order to seventh-order Hu invariant moments are calculated for each Gabor energy feature map in the group of Gabor energy feature maps respectively; All the calculated Hu invariant moments are concatenated into a one-dimensional feature vector in a predetermined order.

7. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The classification model is a random forest classification model.

8. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 7, characterized in that, The classification model outputs the quality detection category of the welding area, which comprises: The random forest classification model has 500 decision trees, and the maximum depth of each decision tree is 25, and the node splitting criterion is Gini impurity. The quality detection category output by the random forest classification model is "qualified", "surface porosity", "unwelded", or "undercut".

9. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 1, characterized in that, The acquisition of the image of the welding area of the condenser to be detected comprises: acquiring an RGB image of the welding area of the condenser to be detected by using a camera matched with an axial light source.

10. The intelligent inspection method for high-precision welding of an aircraft aluminum condenser according to claim 9, characterized in that, The pre-processing of the image to obtain the normalized welding area sub-image comprises: The RGB image is converted into a gray image, a Gaussian filter is used for smoothing and denoising, the welding area in the gray image is cropped and uniformly scaled to 256*256 pixels in size by using a bilinear interpolation algorithm, and a welding area sub-image is obtained. The pixel values of the welding area sub-image are linearly normalized to the interval [0, 1].

Images