A building change detection method based on a double-stage task mechanism network

Abstract

This invention discloses a building change detection method based on a two-stage task mechanism network, belonging to the field of image processing technology. It utilizes a two-stage task mechanism network to construct a building change detection model, which outputs building change detection results based on previous and subsequent temporal images. The two-stage task mechanism network divides the building change detection task into two stages: feature optimization selection and cross-scale progressive fusion. This achieves multi-scale, multi-temporal, and multi-feature integration of images, reducing the impact of human intervention on the accuracy of building change detection. This invention improves the accuracy of building change detection, achieving better detection results and providing strong technical support for urban planning and geological hazard assessment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for detecting building changes based on a two-order task mechanism network. Background Technology

[0002] Building change detection (CD) captures and analyzes information about changes in buildings and has wide applications in urban planning, illegal construction, and disaster management. Traditional remote sensing-based building change detection methods primarily rely on manual visual interpretation. However, these methods are hampered by image imaging conditions and human error, resulting in drawbacks such as high time consumption, low accuracy, low efficiency, and unsuitability for large-area detection. In recent years, intelligent remote sensing image change detection technology has developed rapidly, providing new insights for research in this field.

[0003] Deep learning, with its advantages of high automation, fast processing speed, and wide detection range, has been widely applied in computer vision fields such as scene classification, semantic segmentation, and object detection. Currently, commonly used deep learning-based building change detection methods are divided into early fusion methods and late fusion methods. Early fusion methods use a single network to compare the feature extraction capabilities and differences of images from different time periods, greatly increasing the risk of gradient vanishing, which can lead to interference between features from different time periods. Late fusion methods use multiple independent network structures to extract features from remote sensing images from different time periods. Due to the diverse types and parameters of remote sensing data, and the specific characteristics of buildings themselves, the extracted results in building change detection often appear fragmented and incomplete, and perform poorly when processing building edges. Feature fusion methods combine the extracted features into a single feature with greater discriminative power than the input features, improving the accuracy of late fusion methods. However, this approach reduces the feature extraction capability of remote sensing images from different time periods, resulting in a large number of false positives and false negatives in building change detection based on feature fusion methods, especially regarding the fragmentation and uneven edges of small buildings.

[0004] Under current technological conditions, deep learning-based building change detection technology mainly faces the following challenges: 1) Due to adverse factors such as differences in remote sensing image imaging conditions, building diversity, different objects with the same spectrum, and different spectra of the same object, manual visual interpretation cannot perform large-scale building change detection, and its generalization ability is poor; 2) Remote sensing data are diverse in type and parameters, and the features of buildings themselves are inconsistent in images at different times. Feature fusion deep learning methods will reduce the remote sensing feature extraction capability at different times, resulting in a large number of false positives and false negatives; 3) In reality, the proportion of areas with building changes is relatively small, and most remote sensing images have the problem of imbalance between positive and negative building samples, and the dataset has certain defects. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a building change detection method based on a two-stage task mechanism network, thereby improving the accuracy of building change detection.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0007] A building change detection method based on a two-order task mechanism network includes the following steps:

[0008] S1, acquire sample data, the sample data including: previous phase image, subsequent phase image and building change labels;

[0009] S2, Construct a building change detection network, which is used to output the building change detection results based on the previous time phase image and the subsequent time phase image;

[0010] S3. The building change detection network is trained using sample data to obtain the trained building change detection model.

[0011] S4, Validate the building change detection model using sample data;

[0012] S5 uses a building change detection model to detect changes in the preceding and following time-phase images, and outputs the building change detection results between the preceding and following time-phase images.

[0013] Preferably, in step S1, sample data is obtained from three public datasets for building change detection: WHU-CD, LEVIR-CD, and GZ-CD. The sample data includes previous phase images, subsequent phase images, and building change labels. Each sample data is cropped to obtain sample slice data, and all sample slice data are divided into training set, validation set, and test set.

[0014] Preferably, in step S2, the building change detection network adopts a two-stage task mechanism network, which is a two-branch network structure. The inputs of the two branches are the previous time phase image and the next time phase image, respectively.

[0015] Preferably, the two-stage task mechanism network includes two stages of tasks;

[0016] The first stage task is to learn image features. The process of the first stage is as follows: In the input part, the two-branch network performs the same convolution operation on the previous and subsequent time-phase images respectively. After the convolutional layer calculation, it enters the downsampling layer and the upsampling layer in sequence, and adjacent layers are connected by dense connection blocks. Skip connections are used to connect the simple features obtained by downsampling and the complex features obtained by upsampling. After feature optimization by the feature selection module, three feature layers are extracted in each branch network, including: deep features C1 and shallow features A1 of the first branch network, deep features C2 and shallow features A2 of the second branch network, and output features B1 of the first branch network and output features B2 of the second branch network after feature learning in the first stage.

[0017] The second stage task is to fuse image features using a cross-scale progressive fusion method. The process of the second stage is as follows: First, for the j-th branch network (j=1,2), the deep feature Cj is deconvolved and concatenated with the shallow feature Aj to obtain the j-th fusion feature ACj. At the same time, for the j-th branch network (j=1,2), the shallow feature Aj is deconvolved and concatenated with the merged feature AC to obtain the j-th concatenated feature ACAj. Then, the first fusion feature AC1 and the second fusion feature AC2 of the two branches are concatenated to obtain the merged feature AC. The merged feature AC is then placed into a pooling layer, and the output features B1 of the first branch network and B2 of the second branch network are also placed into the pooling layer. After unified pooling, they are concatenated to obtain the fusion concatenated feature ACB. Finally, the fusion concatenated feature ACB, the first concatenated feature ACA1, and the second concatenated feature ACA2 are concatenated to obtain the fusion feature layer.

[0018] Subsequently, the two-stage task mechanism network obtains building change detection results based on this fused feature layer.

[0019] Preferably, in step S3, the IoU loss function and the cross-entropy loss function are combined with certain weights to obtain the Bayesian combined loss function, which is used as the optimal loss function for the model. The formula is defined as follows:

[0020]

[0021] Where T represents the optimal loss function (i.e., the Bayesian combination loss function), L represents the cross-entropy loss function, I represents the IoU loss function, W represents the weights, and y represents the label. This represents the probability of the model outputting the label y, where N represents the number of samples in the dataset. i This represents the value of the i-th pixel of the actual label. The value of the i-th pixel in the model's predicted value is represented.

[0022] Preferably, a Bayesian optimization method is chosen to determine the values ​​of hyperparameters, i.e., weights W, in the hyperparameter space.

[0023] Preferably, in step S4, the detection results are evaluated for accuracy using four metrics: precision, recall, F1 score, and mIoU. The specific formulas are as follows:

[0024]

[0025] In the formula, TP refers to the number of samples that the model predicts as positive and that are also positive in reality; FP refers to the number of samples that the model predicts as positive and that are also negative in reality; TN refers to the number of samples that the model predicts as negative and that are also negative in reality; and FN refers to the number of samples that the model predicts as negative and that are also positive in reality.

[0026] A storage medium storing a computer program, which, when executed, implements the above-described building change detection method based on a two-stage task mechanism network.

[0027] An apparatus includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described building change detection method based on a two-stage task mechanism network.

[0028] The advantages of this invention are:

[0029] (1) This invention discloses a building change detection method based on a two-stage task mechanism network. Addressing the issues of data mis-extraction and omission in current building change detection methods, particularly the fragmentation and uneven edges of small buildings, this invention utilizes a two-stage task mechanism network for building change detection. The building change detection task is divided into two stages: feature optimization selection and cross-scale progressive fusion. This achieves multi-scale, multi-temporal, and multi-feature integration of images, reducing the impact of human intervention on the accuracy of building change detection. In application, this invention achieves a practically valuable fully automated building change detection technology, improving detection accuracy and yielding better detection results, providing strong technical support for urban planning and geological hazard assessment. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the network structure of the two-stage task mechanism of the present invention.

[0031] Figure 2 This is a schematic diagram of feature optimization selection in the first stage of a two-stage task mechanism network.

[0032] Figure 3This is a schematic diagram of cross-scale progressive fusion in the second stage of a two-stage task mechanism network.

[0033] Figure 4 This is a schematic diagram illustrating the detection results of different methods based on the WHU-CD dataset in this embodiment.

[0034] Figure 5 This diagram illustrates the detection results of different methods used to detect changes in small, dense building clusters based on the LEVIR-CD dataset in this embodiment. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Depend on Figure 1 As shown, a building change detection method based on a two-order task mechanism network according to the present invention includes the following steps:

[0037] S1: Acquisition and preprocessing of the building change detection dataset, constructing training set, validation set, and test set.

[0038] The specific process of step S1 is as follows:

[0039] S11: Obtain sample data from three publicly available building change detection datasets: WHU-CD, LEVIR-CD, and GZ-CD. The sample data includes previous and subsequent time-phase images and change labels. Each dataset must contain buildings of different regions and varying sizes and shapes.

[0040] An earlier phase image refers to an image with vegetation cover or obvious non-construction traces, while a later phase image refers to an image with obvious construction fill traces (obvious construction traces, such as foundations, buildings, structures, squares, etc.). The change label indicates the difference between the earlier and later phase images.

[0041] Some buildings with irregular edges are selected and added to the training samples to facilitate the study of the effect of the present invention on building edge recognition, and the edge recognition effect is verified in the accuracy evaluation in step S5.

[0042] S12, crop each sample data to a size of 512×512 to obtain sample slice data; divide all sample slice data into training set, validation set and test set according to a ratio of 7:2:1.

[0043] S2. Construct a building change detection network, specifically a two-stage task mechanism network (TSTM-Net). The building change detection network is used to output the building change detection results based on the previous time-phase image and the subsequent time-phase image.

[0044] Depend on Figure 1 As shown, the dual-stage task mechanism network consists of two branches, with the inputs to the preceding and following temporal images, respectively. The first branch receives the preceding temporal image as input, and the second branch receives the following temporal image as input. This reduces the computational complexity of a single branch network, thereby improving the ability to preserve detailed information and enhancing operational efficiency. The dual branches perceive building features at different temporal stages and, with the collaborative action of the feature optimization and selection module, extract features at different levels and perform cross-scale progressive fusion to fully exploit the effective change features of buildings in multi-temporal images, learn effective feature information, and ultimately extract the target building change information.

[0045] The two-stage task mechanism network consists of two stages of tasks.

[0046] The first stage focuses on guiding building feature learning to acquire building features. This stage simultaneously considers deep and shallow features, providing relatively smooth decision boundaries based on these features. The process of the first stage is as follows: In the input part, a dual-branch network performs the same convolution operation on the previous and subsequent time-phase images respectively. After computation through the convolutional layer (Conv), the images are fed into downsampling and upsampling layers. Downsampling involves reducing the image size from large to small, while upsampling involves enlarging the image size from small to large. Simultaneously, dense connection blocks (Dense Layers) are introduced between adjacent layers. By connecting features of the same scale in adjacent layers and reusing information from previous layers, the computational load can be significantly reduced, mitigating gradient vanishing to some extent. This embodiment includes four downsampling layers and four upsampling layers. Skip connections connect the shallow features obtained from downsampling with the deep features obtained from upsampling, reducing the loss of low-level features during computation.

[0047] The feature optimization selection module takes two feature layers with downsampling and upsampling skip connections as input. It calculates attention weights by calculating feature similarity and complementarity to ensure that important features are focused on. The feature optimization selection module focuses on strengthening features and extracts the strengthened features into the second stage.

[0048] Figure 2 for Figure 1The specific operation of the feature optimization and selection module is divided into two branches. After the upper branch performs a skip connection operation, it performs a 3*3 convolution operation and adds it to the lower branch's low-level features to densely cover the missing details of the low-level features. Then, it performs 5*5 and 7*7 convolution operations. The lower branch uses an attention mechanism to emphasize the importance of the low-level features. Then, it performs feature multiplication to better handle the interaction information between shallow features and low-level features and capture the correlation between features so as to reflect the differences in the subsequent feature comparison process.

[0049] In this embodiment, two features and the next bottom-level feature are extracted from the feature optimization selection module, including: deep feature C1 and shallow feature A1 of the first branch network, deep feature C2 and shallow feature A2 of the second branch network, and output feature B1 of the first branch network and output feature B2 of the second branch network after the first stage of feature learning.

[0050] The second stage employs a cross-scale progressive fusion method, fusing different levels of features extracted by the input feature optimization and selection module at multiple scales to output a fused feature layer containing rich semantic information about building changes. For example... Figure 3 As shown, the process of the second stage is as follows:

[0051] First, for each network, such as the first network, the deep feature C1 is deconvolved to make its size the same as the shallow feature A1. Then, A1 and C1 are concatenated along the channel dimension to obtain the first fused feature AC1. Similarly, the second network yields the second fused feature AC2.

[0052] Then, the first fusion feature AC1 and the second fusion feature AC2 of the dual-branch network are merged to obtain the merged feature AC; the merged feature AC is put into the pooling layer, and the output features B1 of the first branch network and B2 of the second branch network are also put into the pooling layer. After pooling, they are spliced ​​to obtain the fused spliced ​​feature ACB.

[0053] Simultaneously, for each network, such as the first network, the shallow feature A1 is deconvolutioned and concatenated with the merged feature AC to obtain the first concatenated feature ACA1. Similarly, the second network yields the second concatenated feature ACA2.

[0054] Finally, after splicing the fused splicing feature ACB, the first splicing feature ACA1, and the second splicing feature ACA2, a fused feature layer containing rich semantic information about building changes is obtained.

[0055] Subsequently, the two-stage task mechanism network obtains building change detection results based on this fused feature layer.

[0056] S3. Using the training set and validation set, the building change detection network, i.e., the two-stage task mechanism network, is trained to obtain the trained building change detection model.

[0057] This invention combines the IoU loss function and the cross-entropy loss function with certain weights to obtain a Bayesian combined loss function. This function adjusts the loss function in real time for buildings against complex backgrounds, highlighting the building objects and weakening the background areas. It further separates the building objects from the background, making it easier for the network to extract the building's distinctive features. The formula is defined as follows:

[0058]

[0059] Where T represents the optimal loss function, i.e., the Bayesian combined loss function, L represents the cross-entropy loss function, I represents the IoU loss function, W represents the weights, and y represents the binary label (0 or 1). This represents the probability of the model outputting the label y, where N represents the number of samples in the dataset. i This represents the value of the i-th pixel of the actual label. The value of the i-th pixel in the model's predicted value is represented.

[0060] This invention, based on Bayesian statistical theory, combines prior knowledge with data to improve the efficiency and accuracy of model search. It selects the Bayesian optimization (BO) method to determine the values ​​of hyperparameters, i.e., weights W, in the hyperparameter space, as defined by the following formula:

[0061]

[0062] Among them, Y * Let P(Y|W) be a certain threshold, P(Y|W) be the posterior probability, and EI be the expectation of the hyperparameter W. Given W, That is, Y relative to the threshold Y * How much did the average performance improve? The quality of the hyperparameters is evaluated by calculating the EI (Extreme Indicator). W next The formula is Find the maximum value (i.e., the maximum likelihood estimate) to obtain the estimated value of W.

[0063] S4. The building change detection model is validated using a test set.

[0064] S41, Input the sample data in the test set into the building change detection model of the present invention for detection, and obtain the building change detection results;

[0065] S42 uses four metrics—precision, recall, F1 score, and mIoU—to evaluate the accuracy of the detection results. The specific formulas are as follows:

[0066]

[0067]

[0068] In the formula, TP refers to the number of samples that the model predicts as positive and that are also positive. FP refers to the number of samples that the model predicts as positive and that are also negative. TN refers to the number of samples that the model predicts as negative and that are also negative. FN refers to the number of samples that the model predicts as negative and that are also positive.

[0069] S5 uses a building change detection model to detect the previous and subsequent time-phase images to be detected, and outputs the building change portion between the previous and subsequent time-phase images.

[0070] Example 1

[0071] This invention is applied to the extraction of building change detection information. Three different study areas were selected, and building images at various scales and resolutions were obtained based on the WHU-CD, LEVIR-CD, and GZ-CD datasets. The images were cropped to 512×512 pixels, and the training, validation, and test sets were allocated in a 7:2:1 ratio. The WHU-CD dataset was taken from the area affected by the February 2011 earthquake and subsequently reconstructed. Images were cropped to 512×512 pixels, resulting in 5887 samples for the training set, 1392 samples for the validation set, and 660 samples for the test set. The LEVIR-CD dataset, from Texas, was cropped into 512×512 image slices, and data augmentation was applied to increase the number of images, resulting in 4005 samples for the training set, 876 samples for the validation set, and 408 samples for the test set. The GZ-CD dataset, from the suburbs of Guangzhou, China, was randomly cropped to 512×512 pixels, resulting in 4722 samples for the training set, 969 samples for the validation set, and 326 samples for the test set. The results of the model in this invention were compared with those of three other models: FC-C, FC-D, and MDFFNet. This embodiment uses the TensorFlow 1.4 framework and Python 3.6 programming language for experiments, employing a GTX 1080Ti GPU. Each experiment involves 50 training epochs, processing 4 sets of images per epoch, with an initial learning rate of 10. -3 Starting from round 15, the decay occurs every 10 rounds, with a decay value of 10. -1 The method of this invention converged within 50 rounds of training on three datasets, and other network models were also set up using the same method.

[0072] like Figure 4As shown in Table 1, the detection results of different methods based on the WHU-CD dataset are compared. It can be seen from Table 1 that, except for a slightly lower recall rate than other methods, the precision, F1 and mIoU are higher than other methods. Compared with the edge missing and blurring phenomena of other methods, the model of this invention performs better in terms of the smoothness of building edges.

[0073] Table 1 Comparison of detection results using different methods based on the WHU-CD dataset

[0074] FC-C 0.8424 0.9577 0.8963 0.8948 FC-D 0.6910 0.9716 0.8076 0.8200 MDFFNet 0.2385 0.9179 0.3852 0.5782 This invention 0.9242 0.9415 0.9327 0.9293

[0075] A comparison of different methods for detecting changes in small, dense building clusters was conducted using the LEVIR-CD dataset, such as... Figure 5 As shown in the table, in scenarios involving small, densely packed buildings, both the method of this invention and other methods can identify areas of building change. However, the method of this invention exhibits fewer blurring and fragmentation issues compared to other methods. Even the FC-C method, which performs well, shows some missed detections compared to the method of this invention. Table 2 also shows that the method of this invention has a significant advantage in accuracy compared to other methods, with higher F1 and mIoU scores.

[0076] Table 2 Comparison of detection results using different methods based on the LEVIR-CD dataset

[0077] FC-C 0.5598 0.7345 0.6353 0.7159 FC-D 0.3190 0.7704 0.4512 0.6256 MDFFNet 0.0726 0.7426 0.1324 0.5111 This invention 0.6490 0.6840 0.6660 0.7325

[0078] As shown in Table 3, based on the GZ-CD dataset, the method of this invention shows excellent results in all metrics. Compared with the better-performing FC-D method, the method of this invention improves precision, recall, F1 score, and mIoU by approximately 6%, 20%, 13%, and 14%, respectively. This demonstrates that the model of this invention also has superior performance in detecting changes in target buildings on satellite remote sensing imagery. The two-stage task mechanism of this invention highlights the characteristics of the building itself and better distinguishes between objects with the same spectrum and those with different characteristics. Scene diversity also has little impact on the method of this invention.

[0079] Table 3 Comparison of detection results using different methods based on the GZ-CD dataset

[0080] FC-C 0.4890 0.8040 0.6082 0.6222 FC-D 0.8092 0.6953 0.7480 0.7063 MDFFNet 0.6544 0.5778 0.6137 0.5887 Ours 0.8724 0.9003 0.8850 0.8578

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting building changes based on a two-order task mechanism network, characterized in that, Includes the following steps: S1, acquire sample data, the sample data including: previous phase image, subsequent phase image and building change labels; S2, Construct a building change detection network, which is used to output the building change detection results based on the previous time phase image and the subsequent time phase image; S3. The building change detection network is trained using sample data to obtain the trained building change detection model. S4, Validate the building change detection model using sample data; S5, use the building change detection model to detect the previous and subsequent time-phase images to be detected, and output the building change detection results between the previous and subsequent time-phase images; In step S2, the building change detection network adopts a two-stage task mechanism network, which is a two-branch network structure. The inputs of the two branches are the previous time phase image and the next time phase image, respectively. The two-stage task mechanism network consists of two stages of tasks; The first stage task is to learn image features. The process of the first stage is as follows: In the input part, the two-branch network performs the same convolution operation on the previous and subsequent time-phase images respectively. After the convolutional layer calculation, it enters the downsampling layer and the upsampling layer in sequence, and adjacent layers are connected by dense connection blocks. Skip connections are used to connect the simple features obtained by downsampling and the complex features obtained by upsampling. After feature optimization by the feature selection module, three feature layers are extracted in each branch network, including: deep features C1 and shallow features A1 of the first branch network, deep features C2 and shallow features A2 of the second branch network, and output features B1 of the first branch network and output features B2 of the second branch network after feature learning in the first stage. The second stage task is to fuse image features using a cross-scale progressive fusion method. The process of the second stage is as follows: First, for the j-th branch network (j=1,2), the deep feature Cj is deconvolved and concatenated with the shallow feature Aj to obtain the j-th fusion feature ACj. At the same time, for the j-th branch network (j=1,2), the shallow feature Aj is deconvolved and concatenated with the merged feature AC to obtain the j-th concatenated feature ACAj. Then, the first fusion feature AC1 and the second fusion feature AC2 of the two branches are concatenated to obtain the merged feature AC. The merged feature AC is then placed into a pooling layer, and the output features B1 of the first branch network and B2 of the second branch network are also placed into the pooling layer. After unified pooling, they are concatenated to obtain the fusion concatenated feature ACB. Finally, the fusion concatenated feature ACB, the first concatenated feature ACA1, and the second concatenated feature ACA2 are concatenated to obtain the fusion feature layer. Subsequently, the two-order task mechanism network obtains building change detection results based on this fused feature layer; In step S3, the IoU loss function and the cross-entropy loss function are combined with certain weights to obtain the Bayesian combined loss function, which is used as the optimal loss function for the model. The formula is defined as follows: Where T represents the optimal loss function (i.e., the Bayesian combination loss function), L represents the cross-entropy loss function, I represents the IoU loss function, W represents the weights, and y represents the label. This represents the probability of the model outputting the label y, where N represents the number of samples in the dataset. This represents the value of the i-th pixel of the actual label. The value of the i-th pixel in the model's predicted value is represented.

2. The building change detection method based on a two-stage task mechanism network according to claim 1, characterized in that, In step S1, sample data is obtained from three public datasets for building change detection: WHU-CD, LEVIR-CD, and GZ-CD. The sample data includes previous time-phase images, subsequent time-phase images, and building change labels. Each sample data is cropped to obtain sample slice data, and all sample slice data are divided into training set, validation set, and test set.

3. The building change detection method based on a two-order task mechanism network according to claim 1, characterized in that, The Bayesian optimization method is chosen to determine the values ​​of hyperparameters, i.e., weights W, in the hyperparameter space.

4. The building change detection method based on a two-stage task mechanism network according to claim 1, characterized in that, In step S4, using accuracy Recall rate The accuracy of the detection results is evaluated using four metrics: F1, mIoU, and F1. The specific formulas are as follows: In the formula, TP refers to the number of samples that the model predicts as positive and that are also positive in reality; FP refers to the number of samples that the model predicts as positive and that are also negative in reality; TN refers to the number of samples that the model predicts as negative and that are also negative in reality; and FN refers to the number of samples that the model predicts as negative and that are also positive in reality.

5. A storage medium, characterized in that, It stores a computer program, which, when executed, implements a building change detection method based on a two-stage task mechanism network as described in any one of claims 1-4.

6. A device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a building change detection method based on a two-stage task mechanism network as described in any one of 1-4.

Images