A Tunnel Face Joint Information Recognition Method and Recognition System Based on Mask R-CNN

Abstract

The present invention proposes a method and system for identifying tunnel face joint information based on Mask R-CNN, which belongs to the technical field of tunnel engineering detection. In order to solve the problems of subjective bias and low efficiency in the current manual extraction of tunnel face joints, the method of the present invention includes: marking the joint features of the tunnel face image, constructing a tunnel face joint database, constructing a deep learning model based on the Mask R-CNN algorithm and optimizing its network structure, dividing the tunnel face joint database into images of different scales, using different scale images and the ExponentiaLR method to train and optimize the deep learning model, obtaining an optimized deep learning model, and using the optimized deep learning model to perform joint segmentation and joint information recognition on the tunnel face image to be tested. The method of the present invention can efficiently and accurately identify the joint features of the tunnel face and has good robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel engineering detection, and in particular to a tunnel face joint information recognition method and recognition system based on Mask R-CNN. Background Art

[0002] Tunnels play a vital role in numerous fields, including transportation, water conservancy, mining, military defense, and construction. They not only improve transportation efficiency and economic development, but also enhance residents' lives, resulting in significant social and economic benefits. In recent years, with socioeconomic progress, tunnel construction has entered a new phase of development. However, due to the hidden nature of tunnel construction, unpredictable geological and environmental risks are often a significant constraint on tunnel construction.

[0003] The tunnel face, as the most intuitive vehicle for revealing these complex geological conditions, carries a wide range of geological information and is crucial for ensuring construction safety. The joints and fissures in the face rock mass directly reflect its integrity and are crucial indicators for determining surrounding rock grade and construction safety. These indicators also impact construction costs and operational safety. Therefore, scientific testing of joints and fissures in the tunnel face is crucial for ensuring tunnel construction safety.

[0004] Currently, the extraction of tunnel face joints usually relies on traditional methods of manual observation and measurement. This method is time-consuming, labor-intensive, and prone to subjective bias. In addition, large-scale tunnel projects require processing a large amount of tunnel face joint data, and manual measurement methods are extremely inefficient.

[0005] With the rapid improvement of computer performance, digital image processing methods, three-dimensional point cloud processing methods, and artificial intelligence algorithms have begun to be gradually applied to the identification of joints and fissures in tunnel faces. Mask R-CNN is an instance segmentation algorithm. It is a multi-task network that can be used for "target detection", "target instance segmentation", and "target key point detection". It is an extension of Faster RCNN that adds a branch to predict the segmentation mask in parallel with the bounding box detection. It has been used in tasks such as instance segmentation, target detection, and human key point detection, achieving the best results.

[0006] Based on this, a method for identifying joints and cracks in tunnel faces using the Mask R-CNN algorithm was studied. This method can significantly improve the efficiency and accuracy of joint and crack identification in tunnel faces, and effectively improve the subjective bias and low efficiency of existing manual measurement. Summary of the Invention

[0007] In response to the deficiencies in the above-mentioned prior art, the purpose of the present invention is to provide a tunnel face joint information recognition method and recognition system based on Mask R-CNN. The method is based on the Mask R-CNN algorithm, uses the MobileNetV3 network structure as the backbone network, and further optimizes the MobileNetV3 network structure to improve the model's adaptability and convergence speed. The model is trained by combining multi-scale training with the ExponentiaLR method to enhance the robustness of the model, thereby efficiently and accurately identifying the joint features of the tunnel face, thereby solving the problems of insufficient accuracy, low robustness, and slow detection speed in existing recognition technologies.

[0008] In order to achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0009] In one aspect, the present invention discloses a method for identifying tunnel face joint information based on Mask R-CNN, comprising the following steps:

[0010] S1: Mark the joint features of the tunnel face image and build a tunnel face joint database;

[0011] S2: Build a deep learning model based on the Mask R-CNN algorithm;

[0012] S3: Divide the tunnel face joint database into images of different scales, use the images of different scales and the ExponentiaLR method to train and optimize the deep learning model, and obtain the optimized deep learning model;

[0013] S4: Input the tunnel face image to be tested into the optimized deep learning model to identify joint information, and further process the output image.

[0014] As a further preferred solution of the above technical solution: in step S1, the step of constructing a tunnel face joint database includes:

[0015] S11: Acquire tunnel face image;

[0016] S12: Annotate the joints and cracks in the tunnel face image in the form of linear frames, obtain a JOSN file after annotation, and convert the JOSN file into COCO format.

[0017] As a further preferred solution: In step S2, the process of building the deep learning model is as follows:

[0018] S21: Use the MobileNetV3 network structure as the backbone network, replacing the ResNet network in the Mask R-CNN algorithm;

[0019] S22: Optimize the MobileNetV3 network structure;

[0020] S23: Add the ECA module to the MobileNetV3 network structure to optimize the high-level low-resolution semantic information graph.

[0021] Based on the above technical solution, further, in step S21, the MobileNetV3 network structure adopts depthwise separable convolution and introduces an inverted residual structure into the network. In order to ensure the accuracy of the feature extraction process, the activation function H-Swish is added to the MobileNetV3 network structure, and the SE attention mechanism module P1 is introduced into the inverted residual structure to form a Bneck structure. The expression of the activation function H-Swish is:

[0022]

[0023] Where: Relu is the activation function, x is the input value of the activation function, and H-Swish[x] is the activation function optimized based on Relu.

[0024] Based on the above technical solution, further, in step S22, optimizing the MobileNetV3 network structure includes two steps: simplification and replacement, as follows:

[0025] S221: Delete 3 layers of Bneck in the MobileNetV3 network structure;

[0026] S222: Use the activation function PreLu to replace the original activation functions H-Swich and ReLu in the Bneck structure. The expression of the activation function PreLu is:

[0027]

[0028] Where: y represents the input of the nonlinear activation function in channel i, a i Used to reflect the slope of the function.

[0029] As a further preferred solution: in step S3, the step of obtaining the optimized deep learning model includes:

[0030] S31: Set images of different scales to train the deep learning model, and randomly select a scale for training every ten iterations;

[0031] S32: The ExponentiaLR method is used to further optimize the deep learning model. By reducing the learning rate by a fixed exponential factor after each epoch or each batch, the deep learning model is assisted in finding the optimal solution. The expression of the learning rate lr is:

[0032]

[0033] Where: initial_lr is the initial learning rate, gamma is the reduction factor, usually between 0 and 1, indicating the reduction factor of the learning rate for each epoch.

[0034] On the other hand, the present invention discloses a tunnel face joint information recognition system based on Mask R-CNN, which is used in any of the above-mentioned recognition methods, and includes:

[0035] The image processing module marks the joint features of the acquired tunnel face images and establishes a tunnel face joint database;

[0036] Model building module, which builds a deep learning model based on the Mask R-CNN algorithm and deploys it to the backend;

[0037] The training module uses images of different scales divided by the tunnel face joint database and the ExponentiaLR method to train and optimize the deep learning model to obtain the optimized deep learning model;

[0038] The recognition module calls the deep learning model to recognize the input tunnel face image to be tested and outputs the joint information of the tunnel face image to be tested.

[0039] On the other hand, the present invention also discloses an electronic device, which includes at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions that can be executed by the processor, and the instructions are executed by the processor so that the processor can execute any of the above-mentioned identification methods.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention is based on the Mask R-CNN algorithm and uses the lightweight MobileNetV3 network structure as the backbone network to replace the traditional ResNet network in the Mask R-CNN algorithm. At the same time, it introduces the inverted residual structure and the SE attention mechanism into the inverted residual structure to form a Bneck structure. While achieving a more efficient layer structure, it improves the accuracy of image feature extraction. By deleting the 3-layer Bneck structure, the MobileNetV3 network structure is simplified, thereby reducing the number of parameters and computational complexity of the model, making the model have a faster computing speed. In addition, a more efficient activation function PreLu is used to achieve adaptive learning in the process of correcting linear unit parameters, so that all neurons can update their weights during model training, thereby improving the model's adaptability and accelerating convergence.

[0042] 2. The present invention adopts a combination of multi-scale training and learning rate adjustment method to train the model, which can improve the robustness of the model and make the performance of the trained model better in all aspects.

[0043] 3. The method of the present invention has the advantages of efficient and accurate identification of tunnel face joint features. Compared with existing manual observation and measurement methods, it greatly improves the work efficiency of tunnel face joint feature identification and avoids the influence of identification errors caused by subjective bias. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0045] Figure 1 This is a workflow diagram of a method for identifying tunnel face joint information based on Mask R-CNN of the present invention;

[0046] Figure 2 This is a simplified diagram of the MobileNetV3 network structure of the present invention;

[0047] Figure 3 This is the SPMV_3 structure diagram obtained by the present invention;

[0048] Figure 4 The overall structure of the model (i.e., Mask R-CNN_SPMV_3) after optimizing the network structure of the deep learning model of the present invention;

[0049] Figure 5 For the present invention Figure 4 Enlarged view of (a);

[0050] Figure 6 For the present invention Figure 4 Enlarged view of (b);

[0051] Figure 7 For the present invention Figure 4 Enlarged view of (c);

[0052] Figure 8 The precision and accuracy curves of the model training of the present invention are shown;

[0053] Figure 9 This is the mAP@0.5 and mAP@0.5-0.95 accuracy index curve graph of the present invention;

[0054] Figure 10 The mask skeleton extraction process of the present invention, wherein (a) is the mask image, (b) is the binarized image, and (c) is the skeletonized image;

[0055] Figure 11 The mAP@0.5 comparison results of the optimized deep learning model of the present invention and the basic Mask R-CNN model;

[0056] Figure 12 This is the loss function comparison result of the optimized deep learning model of the present invention and the basic Mask R-CNN model. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0058] Reference Figure 1-12 The present invention provides a method for identifying tunnel face joint information based on Mask R-CNN, comprising the following steps:

[0059] (1) Collect tunnel face images, mark the joint features of the tunnel face images, and build a tunnel face joint database;

[0060] A portable camera or mobile phone is used to track the tunnel construction process, and 1,000 photos of the tunnel face are taken. According to the requirements of multi-scale training in the subsequent process of the present invention, the image sizes are compressed to different sizes; then, the joints and cracks in the images are annotated using the Labelme software. The joints and cracks are annotated in the form of linear frames. The linear frames are covered on the joints and cracks, and the linear frames are required to contain complete joints and cracks. When a tunnel face image contains multiple cracks, they are annotated one by one with a single linear frame. After the images are annotated, a JOSN file is obtained. When the dataset is used for model training in the subsequent steps, the dataset format is required to be in COCO format. The josn_to_dataset function provided by the Labelme software can be used to convert the file format to the COCO format; finally, the entire dataset is divided into a training set and a test set with a ratio of 9:1. The training set is used for the training phase of the model. After obtaining the optimized deep learning model, the feature extraction performance of the optimized deep learning model is tested using the test set.

[0061] (2) Build a deep learning model based on the Mask R-CNN algorithm;

[0062] Based on the Mask R-CNN algorithm, the backbone layer uses the advanced MobileNetV3 network structure to replace the traditional RetNet network. A channel attention module (ECA) is added to the backbone feature extraction network to optimize high-level low-resolution semantic information maps and improve the expressiveness of features. Specifically:

[0063] ① Use the lightweight MobileNetV3 network structure as the backbone network. MobileNetV3 adopts depthwise separable convolution (DSConv) and introduces an inverted residual structure in the network to achieve a more efficient layer structure with its low-rank characteristics.

[0064] In order to ensure the accuracy of feature extraction, a new activation function H-Swish is used in the MobileNetV3 network structure. The function definition formula is:

[0065]

[0066] Where: Relu is the activation function, x is the input value of the activation function, and H-Swish[x] is the activation function optimized based on Relu.

[0067] On this basis, the MobileNetV3 network structure introduces the SE attention mechanism module P1 into the inverted residual structure to form a Bneck structure.

[0068] ②Simplify the MobileNetV3 network structure, delete the 3-layer Bneck structure, reduce the number of parameters and calculations of the deep learning model, so that the deep learning model has faster computing speed and lower power consumption. The simplified network architecture refers to Table 1 and Figure 2 shown.

[0069] The more efficient activation function PreLu (Paramatric Rectified Linear Unit) replaces the original activation functions H-Swich and ReLu in the Bneck structure. The biggest advantage of the PReLu activation function is that it can achieve adaptive learning during the correction of linear unit parameters, allowing neurons to have non-zero gradient values ​​when negative inputs are input. During model training, all neurons can update their weights, which can improve the model's adaptability and accelerate the model's convergence. The expression of the activation function PReLu is:

[0070]

[0071] Where: y represents the input of the nonlinear activation function in channel i, a i Used to reflect the slope of the function, by a i Adjustment can be made so that different channels use different activation functions.

[0072] After adjusting the activation function PreLu, the obtained SPMV_3 structure is as follows Figure 3 shown.

[0073] Table 1 Simplified network architecture of MobileNetV3

[0074]

[0075] In Table 1, 5*5 is the convolution kernel size, 7*7 is the pooling window size, Exp_size is the number of channels raised by the first 1*1 convolution in Bneck, #Out is the number of output channels, NL is the nonlinear activation function, Pre is the PreLu activation function, and SE is squeeze Excitation module ( ), Conv2d is a two-dimensional convolution, Pool is a pooling layer, BN is data normalization, NBN does not use BN, and S is the stride. Conv2d is a classification layer, "—" indicates "no", and "√" indicates "yes".

[0076] because a i It is capable of adaptive learning, so the PreLu function definition can be converted to:

[0077] ③ Add a channel attention module (ECA) to the backbone feature extraction network to optimize the high-level low-resolution semantic information map and improve the expressiveness of features. By introducing the channel attention module, the deep learning model can learn the importance of each channel feature and adjust the output weights of different channels accordingly, making the network pay more attention to features that are beneficial to the current task.

[0078] After optimizing the network structure of the deep learning model, the Mask R-CNN_SPMV_3 network is obtained, and its overall structure is as follows: Figure 4 shown.

[0079] (3) The tunnel face joint database is divided into images of different scales. The deep learning model is trained and optimized using images of different scales and the ExponentiaLR method to obtain the optimized deep learning model.

[0080] When training a deep learning model, the training set is set to images of different scales. During training, a scale is randomly selected for training every ten iterations to enhance the generalization ability of the deep learning model.

[0081] The setting parameters for images of different scales in this invention are min_dimension to 800 and max_dimension to 1024. That is, for any image, the shortest side is at least 800 and the longest side is at most 1024. The input image size is within this range. If the input image is not within this range, it will be adjusted according to the fixed ratio of the original image. For example, if the input image size is 100*100, it will be adjusted to 800*800; if the input is 2000*2000, it will be adjusted to 1024*1024; if the input is 900*1000, the short side will be adjusted to 800 first, and then the long side will be adjusted to (800 / 900)*1000 proportionally.

[0082] ExponentiaLR (exponentially decreasing learning rate regulator) is used to further optimize the deep learning model. By reducing the learning rate by a fixed exponential factor after each epoch or each batch, the deep learning model is helped to approach the optimal solution in a smoother and more gradual manner. The expression of the learning rate lr is:

[0083]

[0084] Where initial_lr is the initial learning rate, that is, the learning rate set at the beginning of training; gamma is the reduction factor, usually between 0 and 1 (for example, 0.9), which represents the reduction factor of the learning rate for each epoch and is used to control the speed at which the learning rate decreases.

[0085] In this step, epoch refers to the entire training set that has been trained and returned once by the deep learning model; batch refers to the number of batches into which the deep learning model divides the training set during training. After training is complete, the optimized deep learning model is obtained.

[0086] (4) The image of the tunnel face to be tested is input into the optimized deep learning model for joint information recognition, and the output image is further processed.

[0087] The tunnel face image to be tested here is the test set mentioned above. The tunnel face image to be tested is input into the optimized deep learning model for joint segmentation and joint information extraction. After completion, a detection image containing joint features is output. The detection image is then grayscaled and binarized with threshold segmentation to obtain a binarized tunnel face image. Skeleton extraction is performed on the binarized image to obtain a skeleton-extracted image. Pixel-level geometric analysis is performed based on the binarized image and the skeleton-extracted image to obtain various parameter values ​​of the joint information, such as the number, length, and width of the tunnel face joints and fissures.

[0088] Specifically, the grayscale processing method is as follows: for RGB images, when the three channels of the image are the same or one of the three components is taken, the image is displayed as a grayscale image. The method of the present invention is based on python3.9 and uses the opencv method to read the image in a single-channel manner to obtain a grayscale image; image binarization is the process of setting the grayscale value of the pixel points on the image to 0 or 255, that is, the entire image presents an obvious black and white effect. By binarizing the image, the shape and contour of the object can be better analyzed.

[0089] Skeleton extraction is performed on the basis of the binarized image, pixel-level geometric analysis is performed based on the binarized image and the skeleton-extracted image, and quantitative statistics of multiple joint information are performed.

[0090] Image binarization involves processing an image so that its pixels are represented by only 0 and 1 (only black and white, with black representing the background and white representing the foreground). The key is to find a threshold value, T, such that pixels below the threshold value become 0, and pixels above the threshold value become 255. Threshold-based image binarization methods include global thresholding and local thresholding. Global thresholding applies a fixed threshold value to all pixels in the entire image, while local thresholding dynamically adjusts the threshold value based on the grayscale value of the local area surrounding the pixel. In this paper, a global threshold-based image binarization algorithm is used, with 127 set as the global threshold. The specific code is: _, img_binary = cv2.threshold(img_gray,127,255,cv2.THRESH_BINARY). The cv2.threshold function binarizes the grayscale image, where 127 is the set global threshold and 255 is the value when the pixel is above the threshold. cv2.THRESH_BINARY indicates that the binarization method sets pixels above the threshold to white and pixels below the threshold to black.

[0091] After the joint features are binarized, skeletonization is performed. Skeletonization refers to extracting the skeleton part of the shape from a binary image. The pixels at the edge of the object in the image are continuously deleted through an iterative method until only a skeleton with a width of one pixel remains. The present invention uses the Zhang-Suen algorithm for skeleton extraction. The specific process is as follows:

[0092] (1) Traverse each pixel of the image (binarized image), and check whether the following conditions are met for the foreground color (pixel value non-zero point): a. The current pixel is white; b. There is at least one black pixel in the upper, lower, left, right, upper left, lower right, and lower right 8-neighborhood of the current pixel P; c. The number of white pixels in the 8-neighborhood of the current pixel is between 2 and 6. If all of the above conditions are met, then the current pixel is marked as black; (2) Traverse each pixel of the image again, and check whether the following conditions are met for the foreground color: a. The current pixel is white; b. There is at least one black pixel in the 8-neighborhood of the current pixel; c. The number of white pixels in the 8-neighborhood of the current pixel is between 2 and 6; d. The black pixels in the 8-neighborhood of the current pixel are not connected (that is, not a connected component). If all of the above conditions are met, then the current pixel is marked as black (background). The above steps (1) and (2) are performed alternately until no more pixels can be deleted. If no pixels are deleted after one round of iteration, the algorithm terminates, otherwise it continues to the next round of iteration. After multiple rounds of iteration, the Zhang-Suen algorithm will refine the edge of the white target object in the image and eventually obtain a skeleton with a width of one pixel. The entire skeleton extraction process is as follows: Figure 10 (a)- Figure 10 As shown in (c), after the skeleton is extracted, the actual length of the joint can be calculated through the proportional relationship between the image and the actual tunnel face.

[0093] In this invention, in order to determine the effectiveness of the deep learning model and optimization constructed by the present invention, the mAP@0.5 and loss function of the optimized deep learning model (i.e., Mask R-CNN_SPMV_3 model) obtained by the present invention are compared with those of the basic Mask R-CNN model. The number of iterations is set to 300, and the following results are obtained: Figure 11-12 The model performance evaluation results.

[0094] Depend on Figure 11 It can be seen that the optimized deep learning model of the present invention improves accuracy faster during training and achieves better final results. The mAP@0.2 of the Mask R-CNN_SPMV_3 model is 0.93, which is 0.05 higher than that of the basic Mask R-CNN model. Figure 12 From the loss function curve in , it can be seen that the Mask R-CNN_SPMV_3 model has a smaller loss error value of 1.037, which is 0.366 less than that of the basic Mask R-CNN model, indicating that the Mask R-CNN_SPMV_3 model of the present invention has better robustness.

[0095] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying tunnel face joint information based on Mask R-CNN, characterized in that: The following steps are involved: S1: Mark the joint features of the tunnel face image and build a tunnel face joint database; S2: Build a deep learning model based on the Mask R-CNN algorithm; S21: Use the MobileNetV3 network structure as the backbone network, replacing the ResNet network in the Mask R-CNN algorithm; The MobileNetV3 network structure adopts depthwise separable convolution and introduces an inverted residual structure into the network. In order to ensure the accuracy of feature extraction, the activation function H-Swish is added to the MobileNetV3 network structure, and the SE attention mechanism module P1 is introduced into the inverted residual structure to form a Bneck structure. The expression of the activation function H-Swish is: Where: Relu is the activation function, x is the input value of the activation function, and H-Swish[x] is the activation function optimized based on Relu; S22: Optimize the MobileNetV3 network structure; Optimizing the MobileNetV3 network structure includes two steps: simplification and replacement, as follows: S221: Delete 3 layers of Bneck in the MobileNetV3 network structure; S222: Use the activation function PreLu to replace the original activation functions H-Swich and ReLu in the Bneck structure. The expression of the activation function PreLu is: Where: y represents the input of the nonlinear activation function in channel i, a i Used to reflect the slope of the function; S23: Add channel attention module to the MobileNetV3 network structure to optimize high-level low-resolution semantic information graph; S3: Divide the tunnel face joint database into images of different scales, use the images of different scales and the ExponentiaLR method to train and optimize the deep learning model, and obtain the optimized deep learning model; S4: Input the tunnel face image to be tested into the optimized deep learning model to identify joint information, and further process the output image.

2. The method for identifying tunnel face joint information based on Mask R-CNN according to claim 1, characterized in that: In step S1, the steps of constructing a tunnel face joint database include: S11: Acquire tunnel face image; S12: Annotate the joints and cracks in the tunnel face image in the form of linear frames, obtain a JOSN file after annotation, and convert the JOSN file into COCO format.

3. According to the method for identifying tunnel face joint information based on Mask R-CNN in claim 1, in step S3, the step of obtaining the optimized deep learning model comprises: S31: Set images of different scales to train the deep learning model, and randomly select a scale for training every ten iterations; S32: The ExponentiaLR method is used to further optimize the deep learning model. By reducing the learning rate by a fixed exponential factor after each epoch or each batch, the deep learning model is assisted in finding the optimal solution. The expression of the learning rate lr is: Where: initial_lr is the initial learning rate, gamma is the reduction factor, between 0 and 1, indicating the reduction factor of the learning rate for each epoch.

4. A tunnel face joint information recognition system based on Mask R-CNN, characterized by: The identification system is used in the identification method according to any one of claims 1 to 3, and the identification system includes: The image processing module marks the joint features of the acquired tunnel face images and establishes a tunnel face joint database; Model building module, which builds a deep learning model based on the Mask R-CNN algorithm and deploys it to the backend; The training module uses images of different scales divided by the tunnel face joint database and the ExponentiaLR method to train and optimize the deep learning model to obtain the optimized deep learning model; The recognition module calls the deep learning model to recognize the input tunnel face image to be tested and outputs the joint information of the tunnel face image to be tested.

5. An electronic device, characterized in that: The electronic device includes at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, and the instructions are executed by the processor so that the processor can perform the identification method according to any one of claims 1 to 3.

Images