Robot vision self-adaptive wavelet denoising method and system based on motion vector direction decoupling

Abstract

The application provides a robot vision adaptive wavelet denoising method and system based on motion vector direction decoupling, and belongs to the field of industrial robot vision image processing. The application projects the three-dimensional motion vector of the robot end to the image plane and decouples it into horizontal and vertical components, constructs a direction coupling matrix, calculates independent blur tendency factors and adaptive thresholds for horizontal, vertical and diagonal wavelet detail subbands respectively, performs improved semi-soft threshold shrinkage, combines the detail coefficients of each layer after threshold shrinkage processing with the unprocessed approximation subband, performs inverse discrete wavelet transform, and reconstructs the denoised workpiece image. The application introduces Sigmoid acceleration saturation mapping and online self-tuning closed loop based on positioning accuracy. The application realizes accurate matching of denoising strength and motion blur direction and strength, reduces the positioning deviation of the circular hole under dynamic working conditions, improves the vertical edge retention rate, and is suitable for high-speed high-precision robot vision guided operation.

Description

Technical Field

[0001] This invention belongs to the field of industrial robot visual image processing, and in particular relates to a robot vision adaptive wavelet denoising method and system based on motion vector direction decoupling. Background Technology

[0002] With the rapid development of automated assembly and welding operations by industrial robots, eye-to-hand vision guidance technology is being used more and more widely. However, when the robotic arm moves quickly, the end-effector camera and the workpiece are relatively displaced, resulting in motion blur in the image. The blurred area produces a large number of medium amplitude coefficients in the wavelet domain, which are mixed with the real weak edge coefficients. This makes it very easy for subsequent positioning algorithms to misjudge it as the real contour, causing grasping or locking failure.

[0003] Existing wavelet thresholding denoising methods suffer from the following drawbacks: First, the threshold is fixed or calculated solely based on noise standard deviation and decomposition scale, completely disregarding the real-time speed and acceleration of the robotic arm. This prevents automatic adjustment of denoising intensity as the positioning transitions between coarse and fine stages, leading to severe mismatches in denoising performance under dynamic conditions. Second, the threshold is isotropic, applying the same threshold to horizontal, vertical, and diagonal detail sub-bands. When the robotic arm moves at high speed in a certain direction, true edges in non-motion directions are excessively suppressed (falsely detected), while false edges in the motion direction remain (missed detection), highlighting a directional contradiction. Third, existing methods use general image quality metrics such as PSNR as evaluation criteria, which are not directly linked to the final positioning accuracy of engineering tasks such as bolt fastening and gripping. Furthermore, they lack an online closed-loop feedback mechanism, resulting in poor robustness to changes in workpiece material, lighting, or environment. Summary of the Invention

[0004] The main objective of this invention is to provide a robot vision adaptive wavelet denoising method based on motion vector direction decoupling. This method addresses the problems of existing wavelet denoising methods, which lack motion vector direction decoupling, have a disconnect between the threshold and the motion condition, and lack closed-loop adaptive capability. As a result, they cannot dynamically match the denoising threshold according to the real-time motion direction and intensity of the robotic arm, leading to the retention of motion blur pseudo-edges and the loss of real edges, resulting in low accuracy of visual positioning operations.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a robot vision adaptive wavelet denoising method based on motion vector direction decoupling, comprising the following steps: S1: Acquire workpiece images and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. Project the two onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component in the image plane. S2: Perform multi-scale wavelet decomposition on the acquired workpiece image to obtain approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands. S3: Obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation; S4: Calculate the velocity direction angle based on the horizontal motion component and the vertical motion component, calculate the acceleration direction angle based on the horizontal acceleration component and the vertical acceleration component, and construct the direction coupling matrix M from the velocity direction angle and the acceleration direction angle. The direction coupling matrix M contains the coupling coefficients of the horizontal, vertical and diagonal direction sub-bands. S5: Calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficients of the horizontal, vertical and diagonal directional sub-bands respectively, and calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation and decomposition scale. S6: An improved semi-soft threshold shrinkage operation is performed on the detail coefficients of each layer and each direction using independent adaptive thresholds. S7: Combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

[0006] In the preferred embodiment, the directional coupling matrix M is constructed from the velocity direction angle and the acceleration direction angle, as follows: Define the motion direction state vectors as follows: ; ; in, These represent the horizontal and vertical components of the three-dimensional acceleration vector at the end of the robotic arm projected onto the image plane. These represent the horizontal and vertical components of the three-dimensional velocity vector at the end effector of the robotic arm projected onto the image plane. These are the velocity direction angle and the acceleration direction angle, respectively. The coupling matrix components in the horizontal, vertical, and diagonal directions are as follows: ; in, , representing the coupling strength when the velocity direction is "parallel" and "perpendicular" to the wavelet subband direction, respectively; The global coupling coefficient for acceleration across all directional sub-band thresholds; The effective intensity value of the end-effector acceleration after Sigmoid nonlinear saturation mapping is used to characterize the actual impact of the current acceleration on motion fuzziness. These are the velocity direction angle and the acceleration direction angle, respectively.

[0007] In the preferred embodiment, the adaptive threshold for directional heterogeneity is calculated using the following formula: ; in, Let be the total number of coefficients in a specific direction (horizontal / vertical / diagonal) of the j-th level decomposition scale. They represent the first Independent adaptive thresholds for detail subbands in the horizontal, vertical, and diagonal directions at the layer decomposition scale. This is an estimate of the standard deviation of the sensor's background noise, reflecting the intensity level of random noise in the image.

[0008] The fuzziness tendency factor for the horizontal, vertical, and diagonal sub-bands is calculated using the following formula: ; in, This is the global sensitivity coefficient. For velocity and acceleration weighting coefficients, These are the horizontal and vertical components of the three-dimensional velocity vector at the end of the robotic arm projected onto the image plane.

[0009] In the preferred scheme, the Sigmoid nonlinear saturation mapping function used is: ; in, For acceleration nonlinear coefficients, This is the normalized value of the total acceleration.

[0010] In the preferred embodiment, the improved semi-soft threshold shrinkage operation is performed according to the following threshold function, expressed as: ; Where λ is the smoothing coefficient. is the coefficient after threshold shrinkage, and T is the adaptive threshold of the corresponding sub-band.

[0011] The preferred scheme also includes online motion blur direction learning and sub-band weight self-tuning closed-loop fine-tuning steps: Based on the horizontal and vertical motion components of consecutive frames, the velocity direction angle is updated in real time, the dominant direction of the current motion blur is dynamically estimated, and the angle parameters in the direction coupling matrix M are adjusted accordingly. The visual positioning accuracy index of the denoised image is used as a feedback signal. When the positioning deviation exceeds the preset threshold or the number of false edges exceeds the preset threshold, at least one of the global sensitivity coefficient, velocity weight coefficient and acceleration weight coefficient is automatically fine-tuned. The fine-tuned parameters are then substituted back into the calculation of the ambiguity tendency factor in each direction until the positioning accuracy meets the requirements, at which point the parameters are locked.

[0012] In a preferred embodiment, the visual positioning accuracy indicators include the hole positioning deviation and the number of false edge detections.

[0013] In the preferred embodiment, the noise standard deviation is estimated using the median absolute deviation method, as shown in the formula: ; Where W is a one-dimensional vector consisting of all coefficients of the finest-scale horizontal detail subband.

[0014] In a preferred embodiment, the discrete wavelet transform employs an orthogonal wavelet basis with compact support and vanishing moment characteristics, including the Daubechies wavelet basis, the Symlet wavelet basis, or the Coiflet wavelet basis.

[0015] Secondly, the present invention provides a robot vision adaptive wavelet denoising system based on motion vector direction decoupling, used to implement the aforementioned robot vision adaptive wavelet denoising method based on motion vector direction decoupling, comprising: The image acquisition module is used to acquire images of the workpiece and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. It then projects these two vectors onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component within the image plane. The multi-scale wavelet decomposition module is used to perform multi-scale wavelet decomposition on the acquired workpiece image, and decompose it into approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands. The noise estimation module is used to obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation. The directional coupling matrix module is used to calculate the velocity direction angle based on the horizontal motion component and the vertical motion component, calculate the acceleration direction angle based on the horizontal acceleration component and the vertical acceleration component, and construct the directional coupling matrix M from the velocity direction angle and the acceleration direction angle. The directional coupling matrix M contains coupling coefficients for horizontal, vertical, and diagonal directional sub-bands. The adaptive threshold module is used to calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficient of the horizontal, vertical and diagonal directional sub-bands, and to calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation and decomposition scale. The threshold shrinkage module is used to perform an improved semi-soft threshold shrinkage operation on the detail coefficients of each layer in each direction using independent adaptive thresholds. The image reconstruction module is used to combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

[0016] This invention provides an adaptive wavelet denoising method for robot vision based on motion vector direction decoupling. To address the problems of false edge detection and loss of true edges caused by motion blur in robot vision systems under high-speed motion, and the shortcomings of existing wavelet denoising methods such as threshold isotropy, disconnect from motion conditions, and lack of closed-loop adaptation, this invention proposes to project the robot's end effector's 3D motion vector onto the image plane and decouple it into horizontal / vertical components, constructing a direction coupling matrix. Independent blur tendency factors and adaptive thresholds are calculated for the horizontal, vertical, and diagonal wavelet detail subbands, respectively. An improved semi-soft thresholding shrinkage is performed, and a Sigmoid acceleration saturation mapping and an online self-tuning closed loop based on positioning accuracy are introduced. This invention improves the accuracy of matching denoising intensity with the direction and intensity of motion blur, reduces the positioning deviation of circular holes under dynamic conditions, improves the vertical edge retention rate, and enhances image reconstruction accuracy, making it suitable for high-speed, high-precision robot vision guidance operations. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the noise reduction method of the present invention; Figure 2 The present invention addresses the issue that traditional fixed thresholds cannot distinguish between pseudo-edges and real edge maps under dynamic conditions. Figure 3 This is a diagram showing the precise suppression effect of the present invention on pseudo-edges based on the directional heterogeneity adaptive threshold of the directional coupling matrix. Detailed Implementation

[0018] Example 1 like Figure 1-3 As shown, a robot vision adaptive wavelet denoising method based on motion vector direction decoupling includes the following steps: S1: Acquire workpiece images and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. Project the two onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component within the image plane.

[0019] S2: Perform multi-scale wavelet decomposition on the acquired workpiece image to obtain approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands.

[0020] S3: Obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation.

[0021] S4: Calculate the velocity direction angle based on the horizontal and vertical motion components, calculate the acceleration direction angle based on the horizontal and vertical acceleration components, and construct the directional coupling matrix M from the velocity direction angle and the acceleration direction angle. The directional coupling matrix M contains the coupling coefficients of the horizontal, vertical, and diagonal directional sub-bands.

[0022] S5: Calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficients of the horizontal, vertical, and diagonal directional sub-bands, and calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation, and decomposition scale.

[0023] S6: An improved semi-soft threshold shrinkage operation is performed on the detail coefficients of each layer and each direction using independent adaptive thresholds.

[0024] S7: Combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

[0025] In this embodiment, after acquiring the workpiece image, the motion vector is projected onto the camera imaging plane. Through multi-scale wavelet decomposition of the motion vector direction, a directional coupling matrix is ​​constructed. Directional heterogeneity adaptive threshold calculation is performed, and after improved semi-soft threshold shrinkage, the denoised workpiece image is reconstructed. This reduces the center positioning deviation of the circular hole under dynamic working conditions, improves the vertical edge retention rate and signal-to-noise ratio, and improves the accuracy of the reconstructed image.

[0026] This embodiment illustrates the detailed steps of the method using a six-axis industrial robot end-effector vision system for bolt hole positioning on a metal workpiece. The workpiece is characterized by circular holes distributed on a metal panel, requiring a positioning accuracy within ±2mm. The robot's end-effector has a rated maximum speed. =2.0m / s², rated maximum acceleration =10m / s².

[0027] S1: Simultaneous acquisition of image and motion data.

[0028] An industrial camera is mounted on the end effector of a robotic arm and moves with the arm. When the robotic arm moves to a predetermined position in front of the workpiece, it triggers the camera to capture a grayscale image of the workpiece scene. The image resolution is 640×480 pixels. Simultaneously, the image processing unit reads the current end-effector velocity *v* and acceleration *a* from the robot controller via EtherCAT industrial Ethernet and aligns them with the image frame timestamp. At the same time, it transforms the three-dimensional motion vector of the end-effector output by the robot controller using a coordinate transformation matrix. With acceleration vector Projected onto the camera's imaging plane, the horizontal motion component within the image plane is obtained. Vertical motion components and the corresponding horizontal acceleration components Vertical acceleration component In this embodiment, the end-effector velocity v = 1.5 m / s and the acceleration a = 6 m / s² at a certain acquisition time, and the image contains slight blurring caused by end-effector motion and random noise from the sensor.

[0029] In the above embodiments, the image processing unit is an independent embedded industrial control computer. Alternatively: it can be deployed inside the robot controller, operating as part of the vision module within the controller; deployed in the cloud, where images and motion parameters are uploaded to the cloud via wireless networks such as 5G, processed, and then the denoised image is sent back to the local controller; or deployed inside the camera, where an image processing chip is integrated within the smart camera to directly output the denoised image.

[0030] The above embodiments use EtherCAT wired communication to acquire motion status. Alternatively, industrial Ethernet protocols such as Profinet and EtherNet / IP can be used, or wireless communication methods such as 5G and WiFi can be employed.

[0031] This embodiment reduces the alignment error between motion parameters and image frame timestamps, thus improving data accuracy.

[0032] S2: Multiscale wavelet decomposition.

[0033] A three-level (L=3) discrete wavelet transform is performed on the acquired workpiece image I using the Daubechies-4 (db4) orthogonal wavelet basis with tight support and fourth-order vanishing moments.

[0034] After decomposition, we obtain: an approximate subband at the coarsest scale. It reflects the overall outline and light and shadow distribution of the workpiece image; three sets of detail sub-bands, each containing horizontal... Vertical direction and diagonal direction The system consists of three sub-bands, where j=1, 2, and 3. j=1 corresponds to the finest scale (the largest sub-band size, 320×240), and j=3 corresponds to the coarsest scale. The coefficient distribution in the detail sub-bands exhibits the following characteristics: coefficients at the edges of the real workpiece have very large absolute values ​​(e.g., above 0.8) and are sparsely distributed; coefficients in flat background areas have very small absolute values ​​(e.g., 0.01~0.05), contributed by sensor noise; and coefficients in motion-blurred areas have moderate absolute values ​​(e.g., 0.15~0.30), which are the key areas to be identified and removed in this embodiment.

[0035] This embodiment can also use other orthogonal wavelet bases with compact support and vanishing moment characteristics, such as Symlet and Coiflet.

[0036] This embodiment reduces the wavelet transform calculation time for a single frame image, thereby improving the efficiency of the overall denoising process.

[0037] S3: Robust estimation of noise standard deviation.

[0038] Since the noise component accounts for the highest proportion and the signal component is the most fragmented in the finest-scale detail subband cH1, this embodiment extracts all coefficients from the cH1 subband to form a one-dimensional vector W (total number of elements N1 = 320 × 240 = 76800). Estimating the noise standard deviation using the median absolute deviation (MAD) method: ; The denominator constant 0.6745 is introduced to ensure that the MAD estimate under Gaussian noise is consistent with the true standard deviation.

[0039] In this embodiment, median(|W|) is calculated to be 0.067, then =0.067 / 0.6745≈0.0993. Because the median is used instead of the mean, the few large coefficients contributed by the margins have minimal impact on the estimation result. It can accurately reflect the background noise level of the sensor.

[0040] This embodiment improves the accuracy of calculating the sensor's background noise level and enhances the reliability of threshold calculation.

[0041] like Figure 2 As shown, the traditional fixed threshold cannot distinguish between pseudo-edges and real edges under dynamic conditions. The blue dashed line in the figure represents the traditional fixed threshold, and the gray curve represents the wavelet coefficient distribution under dynamic conditions. It can be seen that the fixed threshold cannot distinguish between motion-blurred pseudo-edges and real weak edges.

[0042] S4: Adaptive threshold calculation based on the motion state and orientation coupling matrix, including: S4.1: Motion component normalization and nonlinear mapping The horizontal and vertical motion components are normalized separately, and the velocity direction angle is calculated. and acceleration direction angle The expression is: ; The Sigmoid function is used to implement a nonlinear saturation mapping of the effects of acceleration, avoiding excessive increase of the threshold under high acceleration: ; in, For acceleration nonlinear coefficients, This is the normalized value of the total acceleration.

[0043] This embodiment takes This makes the mapping curve at the normalized value of the total acceleration. It is most sensitive around 0.5. It enters the saturation region at that time.

[0044] The above embodiments convert motion parameters into directional heterogeneous fuzzy tendency factors through Sigmoid saturation mapping. Alternatively, piecewise linear mapping, polynomial fitting, or end-to-end regression models based on deep neural networks can be used to directly predict the optimal threshold scaling factor for each sub-band from velocity / acceleration.

[0045] This embodiment simulates the physical characteristic of image blurring tending to saturate under high acceleration, avoiding excessive suppression of real edges under strong vibration, and improving the edge integrity of denoised images under high acceleration conditions (such as sudden stop or sudden start).

[0046] 4.2: Construction of Directional Coupling Matrix and Calculation of Directional Heterogeneity Fuzzy Tendency Factor (1) Constructing the directional coupling matrix

[0047] Define the motion direction state vectors as follows: ; ; in, These represent the horizontal and vertical components of the three-dimensional acceleration vector at the end of the robotic arm projected onto the image plane. These represent the horizontal and vertical components of the three-dimensional velocity vector at the end effector of the robotic arm projected onto the image plane. These are the velocity direction angle and the acceleration direction angle, respectively. The coupling matrix components in the horizontal, vertical, and diagonal directions are as follows: ; in, The above matrix establishes a nonlinear coupling relationship between the direction of motion and the wavelet subband direction.

[0048] Calculation of directional heterogeneity fuzzy tendency factor : ; Where k is the global sensitivity coefficient. These are the weighting coefficients.

[0049] In this embodiment, k=0.5 is the global sensitivity coefficient. These are the weighting coefficients; All are restricted to the closed interval [0.8, 2.0].

[0050] This example uses the substituted data for calculation:

[0051] Calculate the directional coupling matrix: .

[0052] Calculate the fuzzy tendency factor (take) ): .

[0053] Therefore, the fuzzy tendency factor actually used in this embodiment is: ; in, This indicates that there is almost no vertical movement, the threshold remains at the classic value, and the true vertical edge is fully protected. Significantly increased to suppress horizontal motion blur false edges; By integrating horizontal and vertical motion and modulating it with a directional coupling matrix, precise suppression of mixed blur in diagonal directions is achieved, improving the integrity of true edges in non-motion directions.

[0054] Directional coupling matrix in the above embodiments From the direction angle of velocity and acceleration direction angle Calculated via nonlinear combination. Alternatively, it can be simplified to calculation based solely on the velocity direction angle. The decision is made to either ignore the influence of the acceleration direction or adopt a linear weighted model. Alternatively, a fuzzy kernel-based direction prior lookup table approach can be used to pre-calibrate the coupling coefficients corresponding to different motion directions.

[0055] The closed-loop fine-tuning of the global sensitivity coefficient k in the above embodiments is applicable to finely adjustable velocity weights. Acceleration weight or nonlinear coefficient It can also preset multiple sets of parameters according to different workpiece types and automatically switch between them.

[0056] S4.3: Calculate the independent adaptive threshold for each subband. For each detail subband at scale j (j=1,2,3), calculate its specific threshold: ; in, The total number of coefficients in the sub-band corresponding to the j-th layer; the first layer , , .

[0057] Threshold calculation for the first layer (j=1) in this embodiment: =1.416*0.0993*4.97≈0.698; =1.027*0.0993*4.97≈0.506; =1.447*0.0993*4.97≈0.713.

[0058] Compared with the original global threshold (T1=0.664): the horizontal sub-band threshold is slightly improved, which more thoroughly removes horizontal false edges; the vertical sub-band threshold is significantly reduced, which improves the integrity of weak edges in the vertical direction; the diagonal sub-band threshold is the highest, which effectively suppresses the mixed blur in the diagonal direction.

[0059] like Figure 3 As shown, the directional heterogeneity adaptive threshold of the directional coupling matrix effectively suppresses pseudo-edges. The red solid line in the figure represents the adaptive threshold for directional heterogeneity calculated based on the directional coupling matrix and robot kinematics. =1.416, T=0.698), the gray curve represents the distribution of wavelet subband coefficients under the same dynamic condition. It can be seen that the adaptive threshold of the present invention, through the directional heterogeneity mechanism, accurately includes all coefficients of the left motion-blurred pseudo-edge into the directional heterogeneity adaptive suppression region (red shaded area) for filtering; at the same time, relying on wavelet directional sensitivity optimization, it completely preserves the high-amplitude coefficients of the right real edge, and completely solves the inherent contradiction of "false edge omission suppression" and "real edge deletion" caused by the traditional fixed threshold (blue dashed line, T=0.664) being unable to distinguish directional characteristics and only able to process in one cut, which leads to "false edge omission suppression" and "real edge deletion".

[0060] S5: Improved Semi-Soft Thresholding Shrinkage and Image Reconstruction For the coarsest approximate subband All detail coefficients, except those specified, undergo an improved semi-soft thresholding shrinkage operation. This threshold function is coupled with the aforementioned directional heterogeneous adaptive threshold T (already linked by the directional coupling matrix). (And obtained through joint optimization calculations of kinematics and wavelet sensitivity) This is used in conjunction with [other methods]. For each detail coefficient c, its processed coefficients [are used]. for: ; Where λ is the smoothing coefficient. is the coefficient after threshold shrinkage, and T is the adaptive threshold of the corresponding sub-band.

[0061] In this embodiment, λ=10 is chosen to make the transition region curve continuous and smooth.

[0062] The physical meaning of the above formula is: | c |≤T: Set directly to zero to completely remove sensor noise and motion blur pseudo-edges (pseudo-edge coefficients have been modulated by the directional coupling matrix). (effectively identified) Exponential smooth transition avoids Gibbs oscillations at hard thresholds while preserving the continuity of weak edges; |c|≥2T: Zero-loss preservation, fully maintaining the energy and contrast of strong edges. This is thanks to the joint optimization link of robot kinematics → wavelet direction sensitivity → threshold, which ensures that strong edge coefficients are not mistakenly deleted.

[0063] In this embodiment, the horizontal detail sub-band The inner edge coefficient value is 0.82 ( and ), output after processing Vertical detail sub-band The coefficient value of a certain vertical weak edge is: 0.55 ( and ), output after processing .

[0064] The directional heterogeneity fuzziness tendency factor of this invention makes the vertical subband threshold... By setting the threshold below the global threshold, Gibbs oscillations in the hard thresholding function are avoided, ensuring the continuity of weak edges. After soft thresholding of all detail coefficients, it is compared with the unprocessed approximate subband. Combine the images, perform inverse discrete wavelet transform (IDWT), and reconstruct the denoised workpiece image. The entire processing flow does not require fuzzy kernel estimation, and the threshold parameters can be further optimized by utilizing online motion fuzzy direction learning and subband weight self-tuning closed-loop mechanism to achieve adaptive denoising under dynamic conditions.

[0065] In this embodiment, other continuous smooth threshold functions such as the Garrote threshold function and the hard-soft threshold trade-off function can be used.

[0066] S6: Denoising Image Output and Visual Localization. This involves outputting the denoised image... The data is then fed into a subsequent visual positioning module, using algorithms such as Canny edge detection and Hough circle detection, to extract the center coordinates of the workpiece's circular hole, guiding the robotic arm to complete the bolt fastening operation. All computational steps in this method are performed in the wavelet domain, utilizing the directional coupling matrix. Differential threshold scaling is applied to horizontal, vertical, and diagonal sub-bands, and the thresholds of each sub-band are adjusted in real time through a joint optimization link of robot kinematics → wavelet direction sensitivity → threshold, without the need for additional fuzzy kernel estimation and iterative calculations, in a single frame. The image processing time is less than 10ms, which fully meets the real-time requirement of 30fps for industrial robots. Comparative experimental conditions: industrial camera resolution 640×480, frame rate 30fps, exposure time 10ms; robotic arm end effector speed 1.5m / s, acceleration 6m / s², moving along the horizontal direction of the image; workpiece is a metal panel with 8 circular holes with a diameter of 5mm on the surface.

[0067] In the comparative test of this embodiment, the noisy original image, the Gaussian filtering result, and the traditional fixed threshold wavelet denoising were compared respectively. Setting it to 1 means eliminating the effect of motion fuzziness tendency factors, and using only an adaptive method based on global motion information. (Determined by velocity amplitude but without directional decoupling) and the complete method of this invention (including directional coupling matrix) and directional isomerization Canny edge detection and Hough circle detection were performed, and the measured data are shown in Table 1.

[0068] Table 1 Measured Data

[0069] As shown in the table above, the complete method of this invention achieves a circular hole positioning deviation of only 0.05mm under high-speed dynamic conditions, far exceeding the requirement of a working tolerance of ±2mm. Furthermore, the number of false circles is reduced from 2 in the fixed threshold method to 0, and the vertical edge retention rate is increased from 78% to 97%. This demonstrates the effectiveness of the directional coupling matrix and the directional heterogeneity fuzziness tendency factor. It can effectively suppress false edges generated during dynamic photography, while perfectly preserving the real workpiece edges, especially significantly protecting weak edges in non-motion directions.

[0070] In another set of static operating conditions ( In the verification test, The results of the method of the present invention are completely consistent with those of the fixed threshold wavelet method, without introducing any additional side effects, proving that the method of the present invention has good compatibility with static working conditions.

[0071] S7: Fine-tuning of positioning accuracy closed-loop parameters. (Online motion fuzzy direction learning and sub-band weight self-tuning closed-loop mechanism) The system establishes a direct correlation between denoising performance and final positioning accuracy, realizing a closed-loop mechanism for online motion blur direction learning and sub-band weight self-tuning. 1) Online learning: Based on motion parameters of consecutive frames Real-time update of motion direction angle The dominant direction of the current motion blur is dynamically estimated, and the directional coupling matrix is ​​adjusted accordingly. The angle parameters in the middle.

[0072] 2) Self-tuning: If the positioning deviation of the circular hole in three consecutive frames exceeds 0.1mm, or the number of false edges is ≥1, the global sensitivity coefficient will be automatically fine-tuned. (Adjust step size ±0.05), or you can choose to fine-tune the speed weight. Acceleration weight or Sigmoid nonlinear coefficient .

[0073] 3) Closed-loop optimization: Substitute the fine-tuned parameters back into sub-step 4.2 for calculation. The parameters will be automatically locked until the positioning accuracy meets the requirements.

[0074] In this embodiment, the global sensitivity coefficient is adjusted through closed-loop fine-tuning. The value eventually stabilized at 0.48 (initially 0.5), and the positioning deviation was further reduced from 0.05 mm to 0.03 mm. This mechanism enables the system to adapt to different workpiece materials, lighting conditions, and operating environments, significantly improving robustness and versatility.

[0075] Example 2 Further illustrating with reference to Embodiment 1, a robot vision adaptive wavelet denoising system based on motion vector direction decoupling includes: a method for implementing the aforementioned robot vision adaptive wavelet denoising method based on motion vector direction decoupling, comprising: The image acquisition module is used to acquire images of the workpiece and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. It then projects these two vectors onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component within the image plane. The multi-scale wavelet decomposition module is used to perform multi-scale wavelet decomposition on the acquired workpiece image, and decompose it into approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands. The noise estimation module is used to obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation. The directional coupling matrix module is used to calculate the velocity direction angle based on the horizontal motion component and the vertical motion component, calculate the acceleration direction angle based on the horizontal acceleration component and the vertical acceleration component, and construct the directional coupling matrix M from the velocity direction angle and the acceleration direction angle. The directional coupling matrix M contains coupling coefficients for horizontal, vertical, and diagonal directional sub-bands. The adaptive threshold module is used to calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficient of the horizontal, vertical and diagonal directional sub-bands, and to calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation and decomposition scale. The threshold shrinkage module is used to perform an improved semi-soft threshold shrinkage operation on the detail coefficients of each layer in each direction using independent adaptive thresholds. The image reconstruction module is used to combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

[0076] This embodiment provides the working process, working details and technical effects of a robot vision adaptive wavelet denoising method based on motion vector direction decoupling. Please refer to Embodiment 1 for details, which will not be repeated here.

[0077] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A robot vision adaptive wavelet denoising method based on motion vector direction decoupling, characterized in that, Includes the following steps: S1: Acquire workpiece images and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. Project the two onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component in the image plane. S2: Perform multi-scale wavelet decomposition on the acquired workpiece image to obtain approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands. S3: Obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation; S4: Calculate the velocity direction angle based on the horizontal motion component and the vertical motion component, calculate the acceleration direction angle based on the horizontal acceleration component and the vertical acceleration component, and construct the direction coupling matrix M from the velocity direction angle and the acceleration direction angle. The direction coupling matrix M contains the coupling coefficients of the horizontal, vertical and diagonal direction sub-bands. S5: Calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficients of the horizontal, vertical and diagonal directional sub-bands respectively, and calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation and decomposition scale. S6: An improved semi-soft threshold shrinkage operation is performed on the detail coefficients of each layer and each direction using independent adaptive thresholds. S7: Combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

2. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, The directional coupling matrix M is constructed from the velocity direction angle and the acceleration direction angle, as follows: Define the motion direction state vectors as follows: ； ； in, These represent the horizontal and vertical components of the three-dimensional acceleration vector at the end of the robotic arm projected onto the image plane. These represent the horizontal and vertical components of the three-dimensional velocity vector at the end effector of the robotic arm projected onto the image plane. These are the velocity direction angle and the acceleration direction angle, respectively. The coupling matrix components in the horizontal, vertical, and diagonal directions are as follows: ； in, These represent the coupling strength when the velocity direction is "parallel" and "perpendicular" to the wavelet subband direction, respectively. The global coupling coefficient for acceleration across all directional sub-band thresholds; This represents the effective intensity value of the robotic arm's end-effector acceleration after Sigmoid nonlinear saturation mapping. These are the velocity direction angle and the acceleration direction angle, respectively.

3. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 2, characterized in that, The adaptive threshold for directional heterogeneity is calculated using the following formula: ； in, Let be the total number of coefficients in the sub-band corresponding to the j-th layer. They represent the first Independent adaptive thresholds for detail subbands in the horizontal, vertical, and diagonal directions at the layer decomposition scale. This is an estimate of the standard deviation of the sensor's background noise, reflecting the intensity level of random noise in the image; The fuzziness tendency factor for the horizontal, vertical, and diagonal sub-bands is calculated using the following formula: ； in, This is the global sensitivity coefficient. These are the weighting coefficients for velocity and acceleration.

4. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 3, characterized in that, The Sigmoid nonlinear saturation mapping function used is: ； in, For acceleration nonlinear coefficients, This is the normalized value of the total acceleration.

5. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, The improved semi-soft threshold shrinkage operation is performed according to the following threshold function, expressed as: ； Where λ is the smoothing coefficient. is the coefficient after threshold shrinkage, and T is the adaptive threshold of the corresponding sub-band.

6. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, It also includes online motion blur direction learning and subband weight self-tuning closed-loop fine-tuning steps: Based on the horizontal and vertical motion components of consecutive frames, the velocity direction angle is updated in real time, the dominant direction of the current motion blur is dynamically estimated, and the angle parameters in the direction coupling matrix M are adjusted accordingly. The visual positioning accuracy index of the denoised image is used as a feedback signal. When the positioning deviation exceeds the preset threshold or the number of false edges exceeds the preset threshold, at least one of the global sensitivity coefficient, velocity weight coefficient and acceleration weight coefficient is automatically fine-tuned. The fine-tuned parameters are then substituted back into the calculation of the ambiguity tendency factor in each direction until the positioning accuracy meets the requirements, at which point the parameters are locked.

7. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, The visual positioning accuracy indicators include the hole positioning deviation and the number of false edge detections.

8. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, The noise standard deviation is estimated using the median absolute deviation method, with the following formula: ； Where W is a one-dimensional vector consisting of all coefficients of the finest-scale horizontal detail subband.

9. The robot vision adaptive wavelet denoising method based on motion vector direction decoupling according to claim 1, characterized in that, The discrete wavelet transform employs an orthogonal wavelet basis with compact support and vanishing moment characteristics, including the Daubechies wavelet basis, the Symlet wavelet basis, or the Coiflet wavelet basis.

10. A robot vision adaptive wavelet denoising system based on motion vector direction decoupling, characterized in that, To implement the robot vision adaptive wavelet denoising method based on motion vector direction decoupling as described in any one of claims 1-9, the method includes: The image acquisition module is used to acquire images of the workpiece and obtain the robot's three-dimensional motion velocity vector and acceleration vector at the current moment. It then projects these two vectors onto the camera imaging plane to obtain the horizontal motion component, vertical motion component, horizontal acceleration component, and vertical acceleration component within the image plane. The multi-scale wavelet decomposition module is used to perform multi-scale wavelet decomposition on the acquired workpiece image, and decompose it into approximate subbands and detail subbands in each layer and direction. The detail subbands include horizontal detail subbands, vertical detail subbands and diagonal detail subbands. The noise estimation module is used to obtain the finest scale detail subbands, extract all detail subband coefficients, and estimate the noise standard deviation. The directional coupling matrix module is used to calculate the velocity direction angle based on the horizontal motion component and the vertical motion component, calculate the acceleration direction angle based on the horizontal acceleration component and the vertical acceleration component, and construct the directional coupling matrix M from the velocity direction angle and the acceleration direction angle. The directional coupling matrix M contains coupling coefficients for horizontal, vertical, and diagonal directional sub-bands. The adaptive threshold module is used to calculate the fuzziness tendency factor of the corresponding directional detail sub-band based on the coupling coefficient of the horizontal, vertical and diagonal directional sub-bands, and to calculate the independent adaptive threshold of each layer and each directional detail sub-band according to the fuzziness tendency factor, noise standard deviation and decomposition scale. The threshold shrinkage module is used to perform an improved semi-soft threshold shrinkage operation on the detail coefficients of each layer in each direction using independent adaptive thresholds. The image reconstruction module is used to combine the detail coefficients of each layer after threshold shrinkage with the unprocessed approximate subbands, perform inverse discrete wavelet transform, and reconstruct the denoised workpiece image.

Images