Intelligent detection method for bending processing of aluminum alloy plate

Abstract

The application relates to the technical field of image processing, in particular to an intelligent detection method for aluminum alloy plate bending processing, which comprises the following steps: collecting a bending image sequence; separating each frame of image into a base image containing material texture information and a reflection image containing highlight information, to obtain a base image sequence; performing optical flow analysis and texture analysis on the base image sequence to identify processing defects generated in the bending process of the aluminum alloy plate; and generating a visual result image labeled with the positions of the processing defects and the positions of the highlights based on the processing defects and the reflection image, so as to realize detection of the bending defects of the aluminum alloy plate. According to the technical scheme, the light interference can be separated from the root cause, and finally, stable and reliable detection of the bending defects of the aluminum alloy plate is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an intelligent detection method for aluminum alloy plate bending processing. BACKGROUND

[0002] Bending forming is one of the key processing links of aluminum alloy plate, wherein the bending R angle, i.e. the inner radius of the bending die, usually affects the shape and size after bending, so the bending R angle plays a decisive role in the final assembly precision and quality of the product. However, in actual production, such as natural wear of the stamping die, uneven thickness of the plate itself, and internal residual stress generated in the cold working process of the material, etc., all of which may cause various surface defects in the R angle area during forming.

[0003] At present, the light flow method and the texture analysis method are generally used in the prior art to monitor the defects randomly generated in the dynamic processing process in real time. Among them, the light flow method constructs a motion vector field by tracking the motion trajectory of the pixel points in the image sequence, so as to quantitatively analyze the deformation of the object surface; the texture analysis method describes the gray pattern change of the local area of the image by mathematical method, to identify whether the state of the material surface changes.

[0004] However, the common physical assumption of the light flow method and the texture analysis method is that the brightness is constant, that is, the brightness of the corresponding pixels in the images taken at different times should remain unchanged, or at most undergo weak linear change, for the same physical point on the observed object. However, the aluminum alloy plate, especially the plate after surface treatment such as wire drawing, polishing or anodic oxidation, will show strong mirror or semi-mirror reflection characteristics, under the action of the bending machine, the surface curvature and normal direction of the R angle area will change continuously and dramatically to form a highlight area, resulting in a high false alarm rate of the prior art. Therefore, how to effectively eliminate or suppress this dynamic highlight reflection interference under the condition of conventional industrial lighting, so that the light flow and texture analysis technology can be stably and reliably applied to the defect detection of high-reflective plate, is a technical problem to be solved. SUMMARY

[0005] In order to solve the technical problem of high light reflection interference in the bending detection of aluminum alloy plate, the present application provides an intelligent detection method for aluminum alloy plate bending processing, which can effectively filter out false defect signals caused by high light movement, and improve the accuracy of bending detection.

[0006] The application provides an intelligent detection method for bending processing of an aluminum alloy plate, and the prevention and control method comprises the following steps: collecting an image of the aluminum alloy plate to be bent and performing pretreatment to obtain a bending image sequence; performing image decomposition on each frame of image in the bending image sequence, separating each frame of image into a base image containing material texture information and a reflection image containing highlight information, and obtaining a base image sequence; performing optical flow analysis and texture analysis on the base image sequence to identify processing defects generated in the bending process of the aluminum alloy plate; and based on the processing defects and the reflection image, generating a visual result image labeled with the positions of the processing defects and the positions of the highlights to realize detection of the bending defects of the aluminum alloy plate.

[0007] In one embodiment, the image decomposition on each frame of image in the bending image sequence comprises: calculating a base confidence weight of each pixel point in each frame of image in the bending image sequence, and the base confidence weight is used to represent the probability that the pixel point belongs to the base image; and based on the base confidence weight, a base image is obtained by iterative optimization.

[0008] By calculating a base confidence for each pixel, the possibility that the pixel belongs to a stable material texture can be quantified, the inherent texture of the material and the transient highlight reflection can be effectively distinguished, and the accuracy of subsequent separation is ensured.

[0009] In one embodiment, the base confidence weight satisfies the relationship: ; wherein, is the base confidence weight of the pixel (x, y), V(x, y) and D(x, y) represent the within-window variability and the total variability of the window at the pixel (x, y), respectively, V(x, y) and D(x, y) are standard components of the relative total variation filter algorithm, V(x, y) is used to measure the amplitude of the gradient within the window, and D(x, y) is used to comprehensively measure the amplitude and direction consistency of the gradient within the window, is a control parameter of a Gaussian function.

[0010] In one embodiment, the optical flow analysis on the base image sequence comprises: calculating an optical flow field based on the base image sequence; calculating a motion discontinuity index based on the optical flow field, the motion discontinuity index being used to represent the degree of abnormality of the optical flow field; and comparing the motion discontinuity index with a preset motion threshold to identify the processing defects.

[0011] The calculation of the optical flow field on the base image from which the highlight interference is removed can more accurately reflect the real physical deformation of the material, and the accuracy and sensitivity of the detection are improved.

[0012] In one embodiment, the motion discontinuity index satisfies the relationship: ; wherein, represents the motion discontinuity of a pixel point with coordinates at the t-th moment, and is a preset weight coefficient, represents the divergence of a pixel point at the t-th moment, represents the curl of a pixel point at the t-th moment, is the optical flow vector of a pixel point at the t-th moment, is a preset derivation operator.

[0013] In an embodiment, the texture analysis comprises: based on the sequence of base images, extracting texture features of each frame of base image to obtain a sequence of texture features; calculating the difference between texture features of two consecutive frames in the sequence of texture features to obtain a texture mutation index; and comparing the texture mutation index with a preset texture threshold to identify a processing defect.

[0014] Analyzing texture changes on the base image can more truly represent the surface state of the material and effectively avoid misjudging the moving highlight as texture mutation, thereby improving the accuracy of surface defect detection.

[0015] In an embodiment, the texture mutation index satisfies the relationship: ; wherein, is the texture mutation index at the t-th moment, represents a chi-square distance for measuring the difference between two histograms, and respectively represent local binary pattern feature maps of base images at the t-th moment and the (t-1)-th moment, represents a function for calculating a feature histogram.

[0016] In an embodiment, the processing defect generated in the bending process of the aluminum alloy plate is identified by: weighting and fusing the motion discontinuity index obtained by the optical flow analysis and the texture mutation index obtained by the texture analysis to generate a defect confidence map; and performing thresholding processing on the defect confidence map to obtain a pure defect mask map marking the position of the processing defect, wherein the pure defect mask map represents the processing defect.

[0017] By fusing the two indexes, it can not only be determined whether there is a defect, but also the type of defect can be preliminarily distinguished by analyzing which index dominates the occurrence of the defect, thereby providing a reference for subsequent process improvement and fault tracing.

[0018] In one embodiment, the generation process of the visualization result image is: taking the original image of the current frame as a background layer; taking the pure defect mask image corresponding to the processing defect as a defect layer and superimposing the defect layer on the background layer; performing thresholding processing on the reflection image to obtain a highlight interference mask image, and taking the highlight interference mask image as a highlight verification layer and superimposing the highlight verification layer on the background layer to obtain the visualization result image.

[0019] By separating and displaying the defects and the highlights on the same original image, it can be intuitively proved that the detected defects are not caused by the highlight artifacts, thereby improving the credibility of the detection result.

[0020] The technical scheme of the present application has the following beneficial technical effects:

[0021] By decomposing the reflection and the substrate image, the present application transfers the analysis object from the original image seriously disturbed by light to the substrate image reflecting only the information of the material itself, thereby ensuring the constant brightness premise for the subsequent optical flow and texture analysis algorithm, so that the false defect signals caused by the movement of the highlights can be effectively filtered out, and the stable and reliable detection of the bending defects of the aluminum alloy plate is realized. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of an intelligent detection method for aluminum alloy plate bending processing according to an embodiment of the present application.

[0023] Figure 2 is a schematic diagram of an aluminum alloy plate to be detected according to an embodiment of the present application.

[0024] Figure 3 is a schematic diagram of a visualization result image according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical scheme in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments.

[0026] Figure 1 is a flowchart of an intelligent detection method for aluminum alloy plate bending processing according to an embodiment of the present application. As shown in Figure 1 , the intelligent detection method for aluminum alloy plate bending processing 100 includes steps S101-S104, which will be described in detail below.

[0027] S101, collect the image of the bent aluminum alloy plate and perform preprocessing to obtain a bending image sequence.

[0028] In one embodiment, a deployable industrial camera can be used to take pictures of the bent aluminum alloy sheet. It is worth noting that in order to deal with the problem of large light intensity difference in the workshop environment, a high dynamic range (HDR) camera is preferably used, which can ensure that the details of the highlight area and the details of the dark area of the sheet can be clearly captured at the same time in a single exposure, avoiding information loss.

[0029] Further, the camera can continuously collect the bending image of the aluminum alloy sheet at a constant frame rate, for example, 200 frames per second, as shown in Figure 2 The figure shows a schematic diagram of the aluminum alloy sheet to be detected collected by the embodiment of the present application.

[0030] Since the collected original image may contain random noise generated by the camera sensor, in order to reduce the interference of such noise on subsequent calculations, the image of the bent aluminum alloy sheet collected needs to be preprocessed. For example, the preprocessing method can be to use a 3x3 Gaussian filter to smooth each frame of image, which can effectively suppress Gaussian noise while better preserving the edge and detail information of the image. In this application, all preprocessed images are arranged in order of time according to the collected time sequence as a bending image sequence.

[0031] In this way, by deploying a suitable camera and preprocessing, a high-quality, low-noise bending image sequence can be obtained.

[0032] S102, image decomposition is performed on each frame of image in the bending image sequence, each frame of image is separated into a base image containing material texture information and a reflection image containing highlight information, and a base image sequence is obtained.

[0033] In one embodiment, a relative total variation (RTV) filtering technique can be used to perform image decomposition on each frame of image in the bending image sequence. The RTV is a structure-preserving image smoothing filter, which is fundamentally different from traditional Gaussian filtering, median filtering, etc. in principle. Traditional filters have difficulty in distinguishing which details in the image belong to the structure that should be preserved, for example, object edges or material textures, and which belong to non-structural details that should be filtered out, for example, noise or transient highlights. Due to the inherent texture of the material, although it is also high-frequency details locally, its local gradient is usually coherent; while the edge of the highlight reflection is also a high-frequency strong edge, but its appearance and spatial distribution are relatively isolated, and its gradient direction lacks consistency with the gradient direction of the surrounding texture. The RTV algorithm takes advantage of the difference in the statistical properties of the local gradient amplitude and direction to identify and separate these two different types of details.

[0034] Specifically, for each frame in the bent image sequence, the intrinsic variability V(x,y) and total variability D(x,y) of each pixel (x,y) within its neighborhood window are calculated. For example, the window can be 7x7. V(x,y) and D(x,y) are standard components of the RTV algorithm. V(x,y) primarily measures the magnitude of the gradient within the window, while D(x,y) comprehensively measures both the magnitude and direction consistency of the gradient.

[0035] Furthermore, a basis confidence weight can be constructed to guide the filter's operation. Its function is to assign a weight to each pixel, representing the probability that the pixel belongs to the stable basal layer texture that should be preserved. The basis confidence weight satisfies the following relationship:

[0036]

[0037] in, V(x,y) represents the basis confidence weight for pixel (x,y), with values ​​ranging from 0 to 1. The closer the value is to 1, the more likely the pixel is to belong to the basis layer. V(x,y) and D(x,y) represent the intrinsic variability and total variability of the window at pixel (x,y), respectively. These are the control parameters of the Gaussian function.

[0038] Furthermore, the gradient inconsistency of highlight edges is high, so they should receive low weights; the gradient consistency of material textures is high, so they should receive high weights. and The ratio is close to 1 at isolated strong edge locations, but much less than 1 in areas with coherent textures or smooth grayscale. is a hyperparameter used to control the width of the Gaussian function, which determines how sensitive the weight function is to gradient inconsistency. For example, it can be set to 0.3.

[0039] Next, after calculating the basis confidence weights for all pixels, a basis confidence weight map is obtained. This basis confidence weight map can be applied to an iterative optimization solver, such as weighted least squares. The goal of the solution process is to find a smooth image such that the weights... The difference between high-weighted regions and the original image is minimized, while larger differences are allowed in low-weighted regions, i.e., strong smoothing is performed. In this way, through iterative optimization, the regions of high-brightness reflection in the original image are effectively smoothed and suppressed, ultimately resulting in a base image containing only the stable texture information of the board material. Furthermore, by subtracting the base image from the original image, a reflection image containing all the separated high-brightness information is obtained, thus yielding the base image sequence.

[0040] Thus, by image decomposition based on relative total variation, dynamic highlights can be stripped from the original image, providing a pure data source for subsequent accurate defect analysis without the interference of light.

[0041] S103, optical flow analysis and texture analysis are performed on the substrate image sequence to identify processing defects generated in the bending process of the aluminum alloy plate.

[0042] In one embodiment, on the substrate image sequence, a dense optical flow algorithm can be applied to calculate the optical flow field (also known as velocity field) between adjacent frames. Since the substrate image sequence as input has met the basic assumption of constant brightness at this time, the calculated optical flow field can accurately reflect the true physical motion speed and direction of each point on the plate surface.

[0043] Further, in order to quantify defects from the obtained optical flow field, a motion discontinuity index can be constructed to capture abnormal patterns in the optical flow field, which are usually directly related to physical discontinuities of the material such as cracks, tears, etc. The motion discontinuity index satisfies the relationship:

[0044]

[0045] wherein, represents the motion discontinuity of the pixel point with coordinates at the t-th moment, and are preset weight coefficients, represents the divergence of the pixel point at the t-th moment, represents the curl of the pixel point at the t-th moment, is the optical flow vector of the pixel point at the t-th moment, is a preset derivative operator.

[0046] Further, the Nabla operator (Nabla operator) or the Del operator, as a mathematical operator, represents the derivative operation in each direction of space. For example, in a two-dimensional space, it can be represented as a vector form:

[0047]

[0048] wherein, represents the rate of change in the x direction, represents the rate of change in the y direction.

[0049] It can be understood that the divergence describes the expansion or compression intensity of the velocity field, the opening of a micro crack will produce a large positive divergence value, the absolute value is used in the present scheme, only the intensity of the expansion or compression of the velocity field is concerned, and the direction is not considered; the curl measures the local rotation intensity of a point, the non-zero curl indicates that the flow is rotating, corresponding to the occurrence of shear or tearing, that is, the adjacent material layers are moving relative to each other, and the zero curl indicates that the flow is irrotational, corresponding to the smooth, non-tearing stretching or compression of the workpiece.

[0050] Further, And The weight coefficients of and can be set to 0.5. In the normal and uniform material deformation process, the divergence and curl of the optical flow field are very close to zero, and any processing defect causing physical discontinuity of the material will cause the divergence or curl to locally sharply increase at the defect position, therefore, when the MDI value is greater than the preset motion threshold, it can be considered that the point has a processing defect. Exemplarily, the motion threshold can be set to 0.05.

[0051] In an optional embodiment, on the same sequence of base images, the local binary pattern (LBP) feature of each pixel can be calculated to obtain the texture feature of each frame of base image, wherein the LBP is a high-efficiency texture description operator, which encodes the local texture pattern by comparing the gray level relationship between a pixel and its neighborhood pixels.

[0052] In order to characterize the abnormal change of the surface state, such as wrinkles and other defects, the difference between the texture features of two consecutive frames in the texture feature sequence can be calculated to obtain a texture mutation index, which is used to measure the change intensity of the texture of the entire R-angle region between consecutive time frames, and the texture mutation index satisfies the relationship:

[0053]

[0054] Wherein, is the texture mutation index at time t, represents the chi-square distance for measuring the difference between two histograms, and is used to measure the difference between two histograms, and respectively represent the local binary pattern feature maps of the base images at time t and t-1, represents a function for calculating the feature histogram.

[0055] In the normal bending process, the surface texture changes gradually, and the LBP histograms of two consecutive frames are very similar, and the TAI value is small. When the surface suddenly has wrinkles and other processing defects, new LBP patterns are introduced, causing the histogram to change dramatically, and the TAI value produces a sharp peak. Therefore, when the TAI value is greater than the preset texture threshold, it is considered that a processing defect has occurred at this moment. For example, the texture threshold can be 2.0.

[0056] Further, by fusing the analysis results of MDI and TAI, such as taking the union after thresholding and normalizing the two respectively, or weighting fusion after normalization and then thresholding, a pure defect mask map can be obtained which comprehensively marks all the identified processing defect positions.

[0057] In this way, by calculating the two indicators of MDI and TAI in parallel, the deformation state of the plate can be comprehensively monitored from two orthogonal dimensions of motion and texture, so that various processing defects can be reliably identified.

[0058] S104, based on the processing defect and the reflection image, a visual result image which simultaneously marks the processing defect position and the highlight position is generated to realize the detection of the bending defect of the aluminum alloy plate.

[0059] In one embodiment, the pixel points with a value of 1 in the pure defect mask map are the identified processing defect positions. Since the reflection image captures all the dynamic highlight information in the original image, a binary highlight interference mask map can be generated by performing a brightness thresholding on the reflection image, such as extracting the pixel region whose brightness value is in the top 5%. The pixel points with a value of 1 in the map correspond to the positions of the dynamic highlights in the original image.

[0060] Finally, a visual result image which simultaneously marks the processing defect position and the highlight position can be generated to integrate all the key information together. The generation process is as follows: taking the original image of the current frame as the background layer; taking the pure defect mask map corresponding to the processing defect as the defect layer and superimposing it on the background layer; taking the highlight interference mask map as the highlight verification layer and superimposing it on the background layer to obtain the visual result image. Figure 3 As shown in FIG. 8, which is a schematic diagram of the visual result image according to an embodiment of the present application. The highlight region is light gray, the shadow region is dark gray, and the potential defect region is displayed as pure white.

[0061] In this way, by generating the visual result, not only is the clear defect positioning provided for the staff, but also the reliability of the detection system is effectively enhanced.

[0062] It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An intelligent detection method for bending processing of an aluminum alloy plate, characterized by, The method comprises the following steps: An image of a bent aluminum alloy plate is collected and preprocessed to obtain a bending image sequence; Each frame of image in the bending image sequence is decomposed to separate each frame of image into a base image containing material texture information and a reflection image containing highlight information to obtain a base image sequence; the decomposition of each frame of image in the bending image sequence comprises: calculating a base confidence weight of each pixel point in each frame of image in the bending image sequence, the base confidence weight being used to represent the probability of the pixel point belonging to the base image; based on the base confidence weight, a base image is obtained by iterative optimization; the base confidence weight satisfies the following relationship: wherein, is the base confidence weight for pixel (x, y), V(x, y) and D(x, y) represent the within-window variability and the total window variability at pixel (x, y), respectively, V(x, y) and D(x, y) are standard components of the relative total variation filtering algorithm, V(x, y) is used to measure the magnitude of the gradients within the window, and D(x, y) is used to measure the consistency of the magnitude and direction of the gradients within the window, is a control parameter of the Gaussian function; Optical flow analysis and texture analysis are performed on the base image sequence to identify processing defects generated in the bending process of the aluminum alloy plate; Based on the processing defects and the reflection image, a visual result image labeled with the positions of the processing defects and the positions of the highlights is generated to realize the detection of the bending defects of the aluminum alloy plate.

2. The intelligent detection method for bending processing of an aluminum alloy plate according to claim 1, characterized in that, The optical flow analysis of the base image sequence comprises: An optical flow field is calculated based on the base image sequence; A motion discontinuity index is calculated based on the optical flow field, the motion discontinuity index being used to represent the abnormality degree of the optical flow field; The motion discontinuity index is compared with a preset motion threshold to identify the processing defects.

3. The intelligent detection method for bending processing of an aluminum alloy plate according to claim 2, characterized in that, The motion discontinuity index satisfies the following relationship: wherein, represents motion discontinuity of a pixel point with coordinates at the tth moment, and are preset weight coefficients, represents divergence of the pixel point at the tth moment, represents curl of the pixel point at the tth moment, is an optical flow vector of the pixel point at the tth moment, is a preset derivation operator.

4. The intelligent detection method for bending processing of an aluminum alloy plate according to claim 1, characterized in that, The texture analysis comprises: Texture features of each frame of base image are extracted based on the base image sequence to obtain a texture feature sequence; A texture mutation index is obtained by calculating the difference between the texture features of two consecutive frames in the texture feature sequence; The texture mutation index is compared with a preset texture threshold to identify the processing defects.

5. The intelligent detection method for bending processing of an aluminum alloy sheet according to claim 4, characterized in that, The texture mutation index satisfies the following relationship: wherein, is the texture abruptness indicator at time t, denotes the chi-square distance for measuring the difference between two histograms, and denote the local binary pattern feature maps of the base image at time t and t-1, respectively, denotes a function for computing the feature histogram.

6. The intelligent detection method for bending processing of an aluminum alloy plate according to claim 1, characterized in that, The identification of the processing defects generated in the bending process of the aluminum alloy plate comprises: The motion discontinuity index obtained by the optical flow analysis and the texture mutation index obtained by the texture analysis are weighted and fused to generate a defect confidence map; The defect confidence map is thresholded to obtain a pure defect mask image marking the positions of the processing defects, the pure defect mask image representing the processing defects.

7. The intelligent detection method for bending processing of an aluminum alloy plate according to claim 1, characterized in that, The generation process of the visual result image is as follows: An original image of a current frame is taken as a background layer; A pure defect mask image corresponding to the processing defects is taken as a defect layer and superimposed on the background layer; The reflection image is thresholded to obtain a highlight interference mask image, and the highlight interference mask image is taken as a highlight verification layer and superimposed on the background layer to obtain the visual result image.

Images