Steel wire rope surface defect detection method and system based on YOLOv11 network optimization

Abstract

The invention provides a steel wire rope surface defect detection method and system based on YOLOv11 network optimization. The method comprises the following steps: constructing a convolution module based on two-dimensional Haar wavelet transform; a feature extraction backbone module and a detection head heat module in the YOLOv11 network adopt convolution modules based on two-dimensional wavelet transform; the method comprises the following steps: optimizing an EUCBv2 up-sampling module in a YOLOv11 network; constructing a data set containing surface defects of the steel wire rope, and training the optimized YOLOv11 network; and carrying out surface defect detection on the steel wire rope through the trained YOLOv11 network. According to the invention, efficient and accurate surface defect detection can be realized, the detection speed and accuracy are improved, and the method is suitable for real-time monitoring requirements in industrial production.

Description

Technical Field

[0001] This invention relates to machine vision and deep learning technologies, and in particular to a method and system for detecting surface defects in steel wire ropes based on YOLOv11 network optimization. Background Technology

[0002] With the development of infrastructure construction, the requirements for the safety status of wire ropes are becoming increasingly stringent. Traditional wire rope defect detection mainly relies on manual visual inspection or simple mechanical tools, which is not only time-consuming and labor-intensive but also difficult to guarantee accuracy. In recent years, although some image processing-based methods have been proposed, their robustness and real-time performance in complex environments still have shortcomings. Therefore, how to utilize advanced deep learning algorithms to improve the accuracy and efficiency of wire rope surface defect detection has become an urgent problem to be solved. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention aims to provide a method and system for detecting surface defects in steel wire ropes based on YOLOv11 network optimization. This method can achieve efficient and accurate surface defect detection, improve detection speed and accuracy, and is suitable for real-time monitoring needs in industrial production.

[0004] The technical solution to achieve the purpose of this invention is as follows:

[0005] A method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization includes:

[0006] Construct a convolution module based on two-dimensional Haar wavelet transform;

[0007] In the YOLOv11 network, the feature extraction backbone module and the detection head module adopt a convolutional module based on two-dimensional wavelet transform;

[0008] Optimize the EUCBv2 upsampling module in the YOLOv11 network;

[0009] A dataset containing surface defects of steel wire ropes was constructed, and an optimized YOLOv11 network was trained on it.

[0010] Surface defects in wire ropes are detected using a trained YOLOv11 network.

[0011] Furthermore, the construction of the wavelet transform-based convolutional module specifically includes:

[0012] One-dimensional Haar wavelet transform is implemented in two independent dimensions to form a depthwise convolution with four filter combinations and a stride of 2.

[0013] Each channel input is decomposed and convolved through four filters to form a continuous wavelet decomposition hierarchy for the low-frequency components.

[0014] Perform a k×k depthwise convolution on each layer component to form a convolution process involving k×k parameters;

[0015] The convolutional outputs of each level are subjected to inverse wavelet transform through transposed convolution, and then superimposed to synthesize the output results.

[0016] Furthermore, the one-dimensional Haar wavelet transform process is as follows: for the input image X, wavelet transform is implemented by applying convolution kernels of [1,1] / 2 and [1,-1] / 2 in a single spatial dimension, followed by processing through 2x downsampling.

[0017] Furthermore, the four filters are:

[0018]

[0019] Among them, f LL Represents a low-pass filter, f LH f HL f HH This forms a set of high-pass filters;

[0020] Each channel input is processed by four filters to progressively decompose the low-frequency components through convolution, forming a continuous wavelet decomposition hierarchy:

[0021]

[0022] in, i represents the current decomposition level, i = 1, 2, ..., l, where l is the total number of levels. The function is:

[0023]

[0024] Furthermore, the output of a k×k depthwise convolution is:

[0025]

[0026]

[0027] in, Let W be the wavelet transform function. i It is a depthwise convolution kernel weight tensor of size k×k.

[0028] Furthermore, performing an inverse wavelet transform yields:

[0029]

[0030] IWT(M) = Conv-transposed([f LL ,f LH ,f HL ,f HH ],[M LL M LH M HL M HH ])

[0031] Where Conv-transposed is the inverse wavelet transform function, Z (l+1) =0, Z i This represents the inverse wavelet transform result for each layer.

[0032] Furthermore, the use of wavelet transform-based convolutional modules in the YOLOv11 network's feature extraction backbone module and detection head module specifically includes: replacing the convolutional module after the bottleneck module in the feature extraction backbone module with a wavelet transform-based convolutional module, and replacing the convolutional module after the DWConv module in the detection head module with a wavelet transform-based convolutional module.

[0033] Furthermore, the optimization of the EUCBv2 upsampling module in the YOLOv11 network specifically includes: first, using an upsampling operation Up(·) with an amplification factor of 2 to expand the feature map; then, using a 3×3 depthwise separable convolution DWC(·), followed by batch normalization BN(·) and h-Swish(·) activation function operations to enhance the representational ability of the upsampled feature map; finally, using a 1×1 convolution to reduce the number of channels to match the feature map in the next stage.

[0034] Furthermore, the h-Swish activation function is:

[0035]

[0036] ReLU(x) = max(0,x)

[0037] Here, ReLU(x) is the activation function.

[0038] A steel wire rope surface defect detection system, comprising:

[0039] The YOLOv11 network optimization unit constructs a convolutional module based on two-dimensional Haar wavelet transform; the feature extraction backbone module and the detection head module in the YOLOv11 network adopt convolutional modules based on two-dimensional wavelet transform; the EUCBv2 upsampling module in the YOLOv11 network is optimized.

[0040] The training unit constructs a dataset containing surface defects of steel wire ropes and trains the optimized YOLOv11 network.

[0041] The detection unit uses a trained YOLOv11 network to detect surface defects in the wire rope.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention optimizes the YOLOv11 network by integrating wavelet transform convolutional layers and EUCBv2 upsampling modules to enhance the ability to extract frequency domain features and improve model accuracy; the present invention improves mAP to 0.889, which is 7.7% higher than the basic YOLOv11 model; and it realizes simultaneous identification of multiple types of defects, enhancing the robustness and adaptability of the system. Attached Figure Description

[0043] Figure 1 This diagram illustrates the receptive field effect resulting from convolution in the wavelet domain.

[0044] Figure 2 This is a schematic diagram of the EUCBv2 module.

[0045] Figure 3 A model architecture diagram for adding EUCBv2 to YOLOv11. Detailed Implementation

[0046] This invention relates to a method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization, comprising:

[0047] 1. YOLOv11 network optimization based on wavelet transform convolution: Introducing wavelet transform convolution (WTConv) to enhance the ability to extract frequency domain features.

[0048] In this invention, the Haar wavelet transform (HaarWT) is adopted as the analysis tool. For the input image X, the one-dimensional Haar wavelet transform is achieved by applying convolution kernels of [1,1] / 2 and [1,-1] / 2 in a single spatial dimension, followed by downsampling by a factor of 2. To perform the two-dimensional Haar wavelet transform, this invention performs this process in two independent dimensions, forming a depthwise convolution of four filters that are active when the downsampling factor is 2:

[0049]

[0050] Here, f LL This represents a low-pass filter, while f LH f HL f HH This forms a set of high-pass filters. For each input channel, the output of the convolution is:

[0051] [XLL ,X LH ,X HL ,X HH ] = Conv([f LL ,f LH ,f HL ,f HH ],X) (2)

[0052] It consists of four channels, each with a resolution of half that of the input image X in each spatial dimension. X LL It is the low-frequency part of the image, X LH ,X HL ,X HH These represent the horizontal, vertical, and diagonal high-frequency components of the image, respectively.

[0053] Since the kernel in Equation (1) forms an orthogonal basis, the inverse wavelet transform (IWT) can be implemented by transpose convolution:

[0054] X = Conv-transposed([f LL ,f LH ,f HL ,f HH ],[X LL ,X LH ,X HL ,X HH (3) The stepwise decomposition of low-frequency components forms a continuous wavelet decomposition hierarchy. Each level of decomposition follows the following pattern:

[0055]

[0056] in Here, 'i' represents the current decomposition level. This process improves the frequency resolution while relatively reducing the spatial resolution of lower frequencies.

[0057] like Figure 1 For the low-frequency components of the second-order wavelet Applying a 3×3 convolution operation results in a convolution process involving 9 parameters, which actually affects the low-frequency response of the input signal X in the 12×12 region.

[0058] Increasing the size of the convolution kernel leads to an exponential increase in the number of parameters (and corresponding degrees of freedom). To effectively mitigate this problem, firstly, wavelet transform is applied to the input data to separate low-frequency and high-frequency information, followed by filtering and downsampling. Next, small-sized depthwise convolutions are applied to images of different frequencies, and finally, the output is synthesized through inverse wavelet transform. In other words, this process can be expressed as:

[0059] Y = IWT(Conv(W,WT(X))) (5)

[0060] Here, X represents the input tensor, and W is a k×k depthwise convolutional kernel weight tensor with four times the number of input channels as X. This method not only achieves convolutional separation between frequency components but also allows small kernels to operate over a wider region of the original input, i.e., expanding their receptive field relative to the input. See the detailed illustration. Figure 1 .

[0061] This invention employs this primary combination operation and further extends it by drawing on the series principle in equation (4). The specific process is as follows:

[0062]

[0063]

[0064] in, As the input of the current layer, These represent the three high-frequency graphs of the i-th layer described in formula (2).

[0065] To combine outputs of different frequencies, WTConv utilizes the linear properties of wavelet transform and its inverse transform, i.e., IWT(X+Y) = IWT(X) + IWT(Y). Therefore, it executes...

[0066]

[0067] This prompts the present invention to superimpose the outputs of convolutions at different levels, where Z (i) It is the summation output from the i-th layer and all subsequent layers. That is, it adds the outputs of convolutions of different sizes to obtain the final output. However, it does not apply to Y. LL ,Y H Normalization is performed because this normalization does not correspond to the normalization in the original domain. Instead, scaling only the channels is sufficient to balance the influence of each frequency component. The detailed algorithm is shown in Table 1.

[0068] Table 1, WTConv Algorithm

[0069]

[0070] 2. EUCBv2 Upsampling Optimization: An improved EUCBv2 algorithm is used to optimize the upsampling stage, further improving model accuracy.

[0071] Combination Figure 2The EUCBv2 module expands the feature map using an upsampling operation Up(·) with a magnification factor of 2. Next, it enhances the representational power of the upsampled feature map by deploying a 3×3 depthwise separable convolution DWC(·), followed by batch normalization (BN)(·) and h-Swish(·) activation functions. Finally, it reduces the number of channels using a 1×1 convolution C1×1(·) to match the feature map in the next stage. The specific architecture of EUCBv2(·) is as follows... Figure 2 As shown.

[0072] Its mathematical expression is shown in Formula 9:

[0073] EUCBv2(x)=C 1×1 (h_swish(BN(DWC(Up(x)))) (9)

[0074] h-Swish is a variant of Swish developed to simplify the calculation of the formula. The original Swish formula includes the sigmoid function, which is computationally expensive. h-Swish replaces the sigmoid function with a piecewise linear function, making the calculation much simpler.

[0075]

[0076] 3. Model Combination Method

[0077] For the task of detecting surface defects in steel wire ropes, this invention combines wavelet transform convolution (WTconv2d) with YOLOV11 to better improve its performance. For example... Figure 3 In YOLOv11, wavelet transform convolution has been replaced in two places: the feature extraction backbone module and the detection head module. This is to enhance the analysis of high-level semantic understanding, improve the model's ability to understand wire rope defects, and increase the model's classification performance. The replaced path is:

[0078] (a) Path 1 (backbone): yoloV11 --> c3k2 module --> c3k module --> bottleneck module --> WTconv2d

[0079] (b) Path 2 (head): yoloV11 --> head module --> DWConv module --> WTconv2d

[0080] 4. Dataset Construction and Training

[0081] A dataset containing various types of defects, such as broken cores, broken wires, twisting, and scars, was constructed and trained using the optimized YOLOv11 model described above.

[0082] Compared with existing technologies, this invention improves mAP to 0.889, which is 7.7% higher than the basic YOLOv11 model; it achieves simultaneous identification of multiple types of defects, enhancing the robustness and adaptability of the system.

[0083] In summary, the present invention provides a method for detecting surface defects in steel wire ropes based on a YOLOv11 network optimization, comprising: First, collecting and labeling a dataset of steel wire rope images containing various surface defect types. Next, optimizing the YOLOv11 network according to the proposed method, including integrating wavelet transform convolutional layers and an EUCBv2 upsampling module. Then, training the optimized model using the constructed dataset and adjusting hyperparameters to achieve optimal performance. Finally, deploying the system in a real-world environment, capturing steel wire rope surface images via a camera, and using the trained model for real-time defect detection and alarm.

[0084] This embodiment also provides a steel wire rope surface defect detection system, including:

[0085] The YOLOv11 network optimization unit constructs a convolutional module based on two-dimensional Haar wavelet transform; the feature extraction backbone module and the detection head module in the YOLOv11 network adopt convolutional modules based on two-dimensional wavelet transform; the EUCBv2 upsampling module in the YOLOv11 network is optimized.

[0086] The training unit constructs a dataset containing surface defects of steel wire ropes and trains the optimized YOLOv11 network.

[0087] The detection unit uses a trained YOLOv11 network to detect surface defects in the wire rope.

[0088] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization, characterized in that, include: Construct a convolution module based on two-dimensional Haar wavelet transform; In the YOLOv11 network, the feature extraction backbone module and the detection head module adopt a convolutional module based on two-dimensional wavelet transform; Optimize the EUCBv2 upsampling module in the YOLOv11 network; A dataset containing surface defects of steel wire ropes was constructed, and an optimized YOLOv11 network was trained on it. Surface defects in wire ropes are detected using a trained YOLOv11 network.

2. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 1, characterized in that, The construction of the wavelet transform-based convolutional module specifically includes: One-dimensional Haar wavelet transform is implemented in two independent dimensions to form a depthwise convolution with four filter combinations and a stride of 2. Each channel input is decomposed and convolved through four filters to form a continuous wavelet decomposition hierarchy for the low-frequency components. Perform a k×k depthwise convolution on each layer component to form a convolution process involving k×k parameters; The convolutional outputs of each level are subjected to inverse wavelet transform through transposed convolution, and then superimposed to synthesize the output results.

3. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 1, characterized in that, The one-dimensional Haar wavelet transform process is as follows: For the input image X, the wavelet transform is implemented by applying convolution kernels of [1,1] / 2 and [1,-1] / 2 in a single spatial dimension, and then processed by downsampling by 2x.

4. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 3, characterized in that, The four filters are: Among them, f LL Represents a low-pass filter, f LH f HL f HH This forms a set of high-pass filters; Each channel input is processed by four filters to progressively decompose the low-frequency components through convolution, forming a continuous wavelet decomposition hierarchy: in, i represents the current decomposition level, i = 1, 2, ..., l, where l is the total number of levels. The function is:

5. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 4, characterized in that, The output of a k×k depthwise convolution is: in, Let W be the wavelet transform function. i It is a depthwise convolution kernel weight tensor of size k×k.

6. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 5, characterized in that, Performing the inverse wavelet transform yields: IWT(M)=Conv-transposed([f LL ,f LH ,f HL ,f HH ],[M LL ,M LH ,M HL ,M HH ]) Where Conv-transposed is the inverse wavelet transform function, Z (l+1) =0, Z i This represents the inverse wavelet transform result for each layer.

7. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 3, characterized in that, In the YOLOv11 network, the feature extraction backbone module and the detection head module adopt wavelet transform-based convolutional modules. Specifically, the wavelet transform-based convolutional module replaces the convolutional module after the bottleneck module in the feature extraction backbone module, and the wavelet transform-based convolutional module replaces the convolutional module after the DWConv module in the detection head module.

8. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 1, characterized in that, The optimization of the EUCBv2 upsampling module in the YOLOv11 network specifically includes: first, expanding the feature map using an upsampling operation Up(·) with an amplification factor of 2; then, enhancing the representational ability of the upsampled feature map through a 3×3 depthwise separable convolution DWC(·), followed by batch normalization BN(·) and h-Swish(·) activation function operations; finally, reducing the number of channels through a 1×1 convolution to match the feature map in the next stage.

9. The method for detecting surface defects in steel wire ropes based on YOLOv11 network optimization according to claim 8, characterized in that, The h-Swish activation function is: Here, ReLU(x) is the activation function.

10. A wire rope surface defect detection system for implementing the method of any one of claims 1-9, comprising: The YOLOv11 network optimization unit constructs a convolutional module based on two-dimensional Haar wavelet transform; the feature extraction backbone module and the detection head module in the YOLOv11 network adopt convolutional modules based on two-dimensional wavelet transform; the EUCBv2 upsampling module in the YOLOv11 network is optimized. The training unit constructs a dataset containing surface defects of steel wire ropes and trains the optimized YOLOv11 network. The detection unit uses a trained YOLOv11 network to detect surface defects in the wire rope.

Images