Segmentation method and system for disaster scene point cloud data

Abstract

The invention relates to the field of artificial intelligence and point cloud segmentation, and provides a segmentation method and system for disaster scene point cloud data, and the method can effectively promote the network to improve the discrimination capability of new objects through carrying out the data resampling of new objects in a secondary subset in the network continuous learning of students. The teacher network prediction result is combined with the new-class original real label to generate a pseudo label, and training loss is obtained based on the difference between the generated pseudo label and the student network prediction result; point cloud features of a teacher network and a student network are subjected to synchronous layered pooling, and distillation loss is obtained based on the difference between the pooling features of the teacher network and the student network, so that the network does not need to construct total loss based on an old class, and the problem that a computer memory is greatly increased due to old class continuous storage is solved; and meanwhile, the trained teacher network on the main subset is utilized to guide the students to learn on the network, so that the post-disaster objects in the new class can be quickly identified.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and point cloud segmentation, and in particular to a method and system for segmenting point cloud data for disaster scenarios. Background Art

[0002] With the rapid development of disaster prevention and mitigation technologies, the use of computer algorithms to process point cloud data can provide efficient and accurate assistance for disaster assessment and decision-making. As a key technology for identifying object types and segmenting instance shapes, point cloud instance segmentation is an important part of it.

[0003] Currently, deep learning-based point cloud instance segmentation usually uses a neural network implicit fitting function, uses point cloud data to drive the network, self-learns complex patterns and high-dimensional features, and generates object classification and instance masks. It mainly targets closed scenes, that is, research and application are only carried out on indoor and street scene data with known categories, and lacks the ability to distinguish newly emerging object features and categories.

[0004] Disaster scenes are a special type of scene with complex environments, diverse elements, changing features, and strong urgency. When a disaster occurs, it is necessary to effectively distinguish the features and categories of newly emerged objects. For open scene points, methods based on deep learning point cloud instance segmentation + continuous learning are still in their infancy. The only public literature (BoudjoghraMEA,AlKhatibS,LahoudJ,etal.3DIndoorInstanceSegmentationinanOpen-World[J].AdvancesinNeuralInformationProcessingSystems,2024,36) involves methods for indoor open scenes. For outdoor open disaster scenes, this method has the following shortcomings:

[0005] (1) The constructed data does not contain any unique objects in disaster scenes and lacks depth information of outdoor scenes, making it impossible to identify all disaster elements in disaster scenes.

[0006] (2) A continuous learning method based on old class sample data is adopted. When a new object category is added, a data replay anti-forgetting mechanism is adopted, that is, the old class data and the new class data set are not divided, and the trained old class data and the new class data are formed into a data set. The network constructs the loss from the saved old class for continuous learning. In the case of a huge amount of data in the remote sensing point cloud scene, the old class sample data needs to be preserved, which will bring a heavy additional storage burden and greatly increase the computer memory. At the same time, the use of the synthetic data set of the old and new classes increases the complexity and intensity of training, making it impossible to quickly identify the elements of the disaster scene, and thus it is difficult to meet the urgency requirements.

[0007] To this end, a segmentation method and system for disaster scene point cloud data are proposed to solve the above technical problems. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for segmenting point cloud data of disaster scenes, aiming to solve the problems raised in the above background technology.

[0009] One embodiment of the present invention is implemented as follows: a method for segmenting point cloud data for disaster scenarios is provided, the method comprising the following steps:

[0010] Obtain manually annotated conventional urban point cloud data; obtain manually annotated open disaster scene point cloud data, wherein the open disaster scene point cloud data includes a learning set and a validation set, wherein the learning set includes a primary subset and a secondary subset, wherein the annotations of the primary subset only include common objects in general outdoor scenes, and the annotations of the secondary subset only include one or more disaster scene objects;

[0011] Constructing a network structure for segmenting the conventional urban point cloud data; pre-training the network on the conventional urban point cloud data to obtain network pre-training parameters; migrating the pre-trained network structure and pre-training parameters to the main subset;

[0012] The migrated networks are used as the teacher network and the student network respectively, and the student network is used to continuously learn in the sub-subsets. When each sub-subset starts continuous learning, the teacher network and the student network are initialized using the network parameters obtained from the continuous learning of the previous sub-subset, and then the teacher network is frozen;

[0013] Resample the point cloud data in the secondary subset, where background samples belonging to common objects are undersampled and target samples belonging to disaster scene objects are oversampled;

[0014] The resampled point cloud data is synchronously input into the teacher network and student network feature extractors to obtain the point cloud features of the teacher network and the student network respectively;

[0015] The obtained point cloud features of the teacher network and the student network are input into the decoder to generate prediction results of the teacher network and the student network respectively; the prediction results of the teacher network are combined with the original true label annotation of the target class to generate pseudo labels, and the training loss is calculated based on the difference between the pseudo labels and the prediction results of the student network; the point cloud features of the teacher network and the student network are simultaneously layered pooled to obtain the layered pooled features of the teacher network and the student network respectively, and the distillation loss is calculated based on the difference between the layered pooled features of the two;

[0016] The distillation loss and the training loss are added together to obtain a total loss, back-propagation is performed on the student network based on the total loss, and the student network parameters are optimized. After completing the training on all sub-subsets, the optimized parameters of the student network are obtained;

[0017] After initializing the student network with the optimized parameters, the point cloud data on the validation set is processed, and the obtained semantic-level predictions and instance-level masks are compared with the annotated true categories and masks to evaluate the point cloud data segmentation effect.

[0018] Another embodiment of the present invention is implemented as follows: providing a segmentation system for disaster scene point cloud data, the system comprising:

[0019] Point cloud data acquisition and annotation unit: The point cloud data acquisition and annotation unit is used to acquire manually annotated conventional urban point cloud data; and acquire and manually annotate open disaster scene point cloud data. The open disaster scene point cloud data includes a learning set and a validation set. The learning set includes a primary subset and a secondary subset. The labels of the primary subset only include common objects in general outdoor scenes, and the labels of the secondary subset only include one or more disaster scene objects.

[0020] A network pre-training unit, which is used to pre-train the network on manually labeled conventional urban point cloud data to obtain network pre-training parameters;

[0021] A network migration unit, configured to migrate the network and pre-trained parameters on a primary subset of manually labeled open disaster scene point cloud data;

[0022] A network continuous learning unit is configured to use the migrated network as a teacher network and a student network, respectively, and to use the student network to perform continuous learning on the secondary subset to obtain student network parameters. The network continuous learning unit specifically includes:

[0023] Initialization module: used for initializing the teacher network and the student network using the network parameters obtained from the continuous learning of the previous sub-subset when the student network starts continuous learning in each sub-subset, and then freezing the teacher network;

[0024] Sub-subset data resampling module: used to resample the point cloud data in the sub-subset, where background samples belonging to common objects are undersampled and target samples belonging to disaster scene objects are oversampled;

[0025] Point cloud feature extraction module: used to synchronously input the resampled point cloud data into the teacher network and student network feature extractors to obtain the point cloud features of the teacher network and the student network respectively;

[0026] Training loss calculation module: used to generate pseudo labels using the prediction results of the teacher network in combination with the original true labels of the target class; and calculate the training loss using the pseudo labels and the prediction results of the student network;

[0027] Distillation loss calculation module: used to synchronously perform layered pooling on the point cloud features of the teacher network and the student network to obtain pooled features; and calculate the distillation loss using the pooled features;

[0028] The network parameter optimization module is used to perform backpropagation training on the student network using the total loss obtained by adding the distillation loss and the training loss, optimize the student network parameters, and obtain the optimized parameters of the student network after completing the training on all sub-subsets.

[0029] Network application evaluation unit: The network application unit is used to use the optimization parameters to initialize the student network and process the point cloud data on the validation set, and compare the obtained semantic-level predictions and instance-level masks with the labeled true categories and masks to evaluate the point cloud data segmentation effect.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) Aiming at the actual application scenarios of disasters, a point cloud segmentation dataset is constructed and divided into multiple subsets. Each main subset contains various common objects in general scenes, and each secondary subset contains specific objects in disaster scenes. The disaster degree of specific objects in disaster scenes is further refined. While retaining the ability to process the original common objects, new post-disaster objects or objects that need to be paid attention to after the disaster are learned, so that all objects in the disaster scene can be recognized.

[0032] (2) In response to the need for data segmentation of specific object classes (new classes) after a disaster occurs, during the continuous learning process, data of new class objects in the subtask data are resampled to obtain samples of special attention for the new class data, avoiding the problem of class imbalance in subsequent learning tasks due to the small proportion of new class data in the entire data, thereby effectively promoting the network to improve its ability to distinguish new classes.

[0033] (3) In the continuous learning process, the prediction results of the teacher network are used to generate pseudo labels in combination with the real labels of the input target class, and the difference between the pseudo labels and the results predicted by the student network is used to obtain the training loss; by synchronously layering and pooling the features extracted by the teacher network and the student network, the difference between the pooled features is used to obtain the distillation loss, and then the loss of continuous learning of the network is obtained, so that the network does not need to be constructed based on the old class loss, and the problem of the increase and decrease of the old class greatly increasing the computer memory is solved; since the training loss and the distillation loss are both related to the teacher network, and the teacher network has been trained on the main subset (including regular objects) during pre-training, the teacher network can be used to guide the student network to learn the newly added post-disaster objects or objects of focus. The teacher network features are used to guide the generation of student network features, which helps the student network to retain the review of old knowledge while learning new classes, alleviate the problem of catastrophic forgetting, and realize the rapid identification of disaster elements, thereby meeting the requirements of urgency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of a segmentation method provided in one embodiment;

[0035] Figure 2 is a flow chart of a segmentation method in another embodiment;

[0036] Figure 3 is a flow chart of a segmentation method in another embodiment;

[0037] Figure 4 is a flow chart of a segmentation method in another embodiment;

[0038] Figure 5 is a flow chart of a segmentation method in another embodiment;

[0039] Figure 6 Schematic diagram of a distillation characteristic structure in a segmentation method in one embodiment;

[0040] Figure 7 This is a diagram showing the effect of a continuous learning method in a segmentation method in one embodiment;

[0041] Figure 8 A structural block diagram of a segmentation system provided in another embodiment;

[0042] Figure 9 4 is a structural block diagram of continuous learning in one embodiment. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] like Figure 1 FIG. 1 is a diagram showing a method for segmenting point cloud data for disaster scenarios provided by an embodiment of the present invention, the method comprising:

[0045] S101 Point cloud data acquisition and labeling: Acquire manually annotated conventional urban point cloud data; acquire manually annotated open disaster scene point cloud data, wherein the open disaster scene point cloud data includes a learning set and a validation set, wherein the learning set includes a primary subset and a secondary subset, wherein the annotations of the primary subset only include common objects in general outdoor scenes, and the annotations of the secondary subset only include one or more disaster scene objects;

[0046] S102 Network pre-training and migration: constructing a network structure for segmenting the conventional urban point cloud data; pre-training the network on the conventional urban point cloud data to obtain network pre-training parameters; migrating the pre-trained network structure and pre-training parameters to the main subset;

[0047] S103: Initializing the teacher network and the student network: using the migrated networks as the teacher network and the student network, respectively, and using the student network to continuously learn in the sub-subsets. When each sub-subset starts continuous learning, the teacher network and the student network are initialized using the network parameters obtained from the continuous learning of the previous sub-subset, and then the teacher network is frozen.

[0048] S104 Sub-subset data resampling: resample the point cloud data in the sub-subset, undersampling background samples belonging to common objects and oversampling target samples belonging to disaster scene objects;

[0049] S105 Point cloud feature extraction of the teacher network and the student network: The resampled point cloud data is synchronously input into the teacher network and the student network feature extractor to obtain the point cloud features of the teacher network and the student network respectively;

[0050] S106 Calculation of training loss and distillation loss: Input the obtained point cloud features of the teacher network and the student network into the decoder to generate prediction results of the teacher network and the student network respectively; Use the prediction results of the teacher network in combination with the original true label annotation of the target class to generate pseudo labels, and calculate the training loss according to the difference between the pseudo labels and the prediction results of the student network; Synchronously perform layered pooling on the point cloud features of the teacher network and the student network to obtain layered pooling features of the teacher network and the student network respectively, and calculate the distillation loss according to the difference between the layered pooling features of the two;

[0051] S107: Optimizing student network parameters: Adding the distillation loss and the training loss to obtain a total loss, performing backpropagation training on the student network based on the total loss, optimizing the student network parameters, and obtaining optimized student network parameters after completing training on all sub-subsets;

[0052] S108: Evaluation of the effect of disaster scene point cloud data segmentation: Use the optimization parameters to initialize the student network and then process the point cloud data on the validation set to obtain semantic-level predictions and instance-level masks. Compare the obtained semantic-level predictions and instance-level masks with the annotated true categories and masks to evaluate the effect of point cloud data segmentation.

[0053] In an embodiment of the present invention, for step S101, a typical test environment and case drone laser point cloud data are used as raw data, covering the complete multi-view images and point cloud information of Beichuan County; by labeling the disaster object data, semantic classes are: background, vegetation, vehicles, waters, roads, houses, damaged houses and ruins, and instance classes are: collapsed houses, roads, ruins; the main subset includes 5 types of objects (background, vegetation, vehicles, waters, houses), and the secondary subset includes 3 types of objects (roads, ruins, damaged houses). Classes not included in each subset are set as background classes; at the same time, the disaster degree in the secondary subset objects is fine-grained, so as to train the network to make a refined judgment on the disaster degree;

[0054] For step S102, since there is relatively little data on disaster scenes, directly using the initialized network for training is likely to lead to overfitting or insufficient generalization. Therefore, the model is first pre-trained on a large urban dataset, including eight networks: PointGroup, HAIS, SoftGroup, DyCo3D, DKNet, BSeg, Mask3D, and OneFormer3D as the skeleton network. The BSeg, Mask3D, DKNet, and SoftGroup networks that perform well in the pre-training process are migrated to the main subset of disaster scene point cloud data.

[0055] For step S103, when continuous learning on disaster scene point cloud data begins, the migrated network is used as the teacher network and student network respectively, and the student network is used to train on the sub-subsets. At the beginning of training on each sub-subset, the teacher network and the student network are initialized using the parameters trained on the previous sub-subset, and then the teacher network is frozen. That is, at the beginning of training on each sub-subset, the teacher network and the student network are exactly the same.

[0056] For step S104, since the current sub-subset only contains background classes and target classes, and the background class is composed of true background classes and pseudo background classes (vegetation, vehicles, waters, houses) that have been learned in the pre-training task, the target class accounts for a relatively small proportion in the entire data. This leads to a class imbalance problem in subsequent learning tasks, which can easily cause overfitting of the background class (old class) and neglect of the target class (new class). Therefore, before each round of training begins, it is necessary to oversample the target class samples belonging to disaster scene objects in the current sub-task data set to achieve special attention to the newly added target class (new class) data, thereby effectively prompting the network to improve its ability to distinguish new classes.

[0057] For step S105, the teacher network and student network feature extractors are used to obtain the point cloud features of the teacher network and the student network respectively. For the same subset, in the first round of training, the point cloud features of the teacher network and the student network are the same. Starting from the second round of training, due to changes in the student network, there are differences in the point cloud features of the teacher network and the student network.

[0058] For step S106, based on the dynamic threshold, the background class labels judged to be low-quality in the prediction results of the teacher network are eliminated to improve the credibility of the prediction, and the original real label annotation of the target class data is added to generate pseudo labels, and the training loss is obtained according to the difference between the pseudo labels and the results predicted by the student network; the point cloud features of each layer obtained by the teacher network and the student network are simultaneously layered and pooled, and the difference between the pooled features of each layer of the two is evaluated to obtain the distillation loss. In the first round of training of each sub-subset, since the pooled features of the teacher network and the student network are the same, the distillation loss is 0. Starting from the second round of training, there is a distillation loss due to changes in the student network; the training loss is added to the distillation loss The total loss is obtained to optimize the training of the student network. Since both the training loss and the distillation loss are related to the teacher network, and the teacher network has been trained on the main subset (including regular objects) during pre-training, the teacher network can be used to guide the student network to learn newly added post-disaster objects or objects of focus. Guiding the generation of student network features through the teacher network features helps the student network to retain the review of old knowledge while learning new categories, alleviate the problem of catastrophic forgetting, and thus realize the network without constructing losses based on old categories, solving the problem of greatly increasing computer memory due to the continued existence of old categories; at the same time, using the teacher network trained on the main subset to guide the student network learning, it is possible to quickly identify post-disaster objects in new categories.

[0059] As another preferred embodiment of the present invention, a method for segmenting point cloud data for disaster scenes is provided. In step S101, the network structure includes a structure constructed using a sparse U-Net network as a skeleton network;

[0060] In this embodiment, the network is pre-trained on the conventional urban point cloud instance segmentation data. The skeleton networks with better performance are BSeg, Mask3D, DKNet and SoftGroup. At the same time, pre-training parameters of various networks are obtained. These skeleton networks can effectively cope with outdoor scene point cloud data, and these networks are selected for subsequent learning experiments.

[0061] like Figure 2 As shown, as another preferred embodiment of the present invention, a method for segmenting point cloud data for disaster scenes is provided. In step S102, the step of migrating the network and pre-trained parameters on the main subset includes:

[0062] S201: Initialize the network using pre-trained parameters;

[0063] S202: Connect a multi-layer perceptron network after the network feature extractor, and use LeakyReLU as the activation function to obtain the feature offset extracted by the network on each main subset through training and learning;

[0064] S203: Add the network extracted features and the feature offset to obtain a new feature;

[0065] S204: Based on the new features, make minor adjustments to the pre-trained network parameters and network structure after training by migrating to each main subset at 0.2-0.3 times the pre-trained learning rate.

[0066] In this embodiment, to reduce the impact of the domain migration process on the original network, a multi-layer perceptron (MLP) network is connected after the network feature extractor to learn the feature offsets extracted by the network on different data sets. The network is migrated to the disaster scene point cloud data, and training is carried out on the main subset based on the new features and pre-trained network parameters. The network structure is trained with a small learning rate and fine-tuned to achieve a rapid response to disaster scenarios.

[0067] The ablation experiments in Table 1-2 are based on the BSeg network. Baseline refers to direct training without pre-training; transfer without MLP refers to pre-training followed by transfer without MLP; and transfer with MLP refers to pre-training followed by transfer with MLP feature shifting. The quantitative results in Table 1-2 show that using the pre-trained and feature-shifted MLP network significantly improves instance segmentation metrics such as AP, AP50, and AP25, as well as semantic segmentation metrics such as Accuracy and mIoU. The semantic segmentation achieves optimal results across all object metrics, demonstrating the positive impact of pre-training and feature-shifted MLP on both instance and semantic segmentation.

[0068] Table 3-4 shows a comparative experiment on model migration of four networks for disaster scenarios (MLP feature offset features are added after pre-training). From the quantitative results in Table 3-4, it can be seen that after using the pre-training and feature offset MLP network, better results are achieved in all kinds of object indicators in semantic segmentation.

[0069] Table 1. Ablation experiment results of model transfer for disaster scenarios

[0070]

[0071] Table 2 Intersection-comparison results of model migration ablation experiments for disaster scenarios

[0072]

[0073] Table 3 Comparative experiments on model transfer of four networks for disaster scenarios (transferring with MLP)

[0074]

[0075] Table 4. Intersection-over-Union (IoU) results of model transfer comparison experiments for disaster scenarios (transferring w / MLP)

[0076]

[0077] like Figure 3 In another preferred embodiment of the present invention, a method for segmenting point cloud data for disaster scenarios is provided. In step 104, the step of resampling the point cloud instance segmentation data in the secondary subset includes:

[0078] S301: Counting the number of points belonging to the target class and the number of points belonging to the non-target class in the point cloud data;

[0079] S302: Calculate the ratio of the number of target class points to the number of non-target class points as the sampling probability of the non-target class sample, and take the value 1 as the sampling probability of the target class sample;

[0080] S303: Sampling to generate training samples;

[0081] In this embodiment, the target class is resampled according to the sampling probability of non-target class samples and the sampling probability of target class samples, so as to overcome the problem of class imbalance in subsequent learning tasks caused by the small proportion of target class in the entire data, which may easily lead to overfitting of old classes and neglect of new classes. t The selection probability of each sample ρ ti for:

[0082]

[0083] Where: T is the target class, N t is the number of target class samples, N0 is the number of other class samples, D t For a sub-subset, d t For a certain sample.

[0084] like Figure 4 As shown, as another preferred embodiment of the present invention, a method for segmenting point cloud data for disaster scenes is provided. In step 105, the steps of using the prediction results of the teacher network in combination with the original true labels of the target class to generate pseudo labels, and calculating the training loss based on the difference between the pseudo labels and the prediction results of the student network include:

[0085] S401: Use the teacher network to predict semantic probabilities and calculate the difference between the average predicted probability and the maximum predicted probability for each type of object;

[0086] Calculate the ratio between the predicted average probability and the predicted maximum probability difference. If it is greater than a fixed value, take the minimum predicted probability as the semantic threshold. Otherwise, take the maximum value between the minimum predicted probability and the fixed threshold as the semantic threshold.

[0087] The predicted points smaller than the semantic threshold are classified as background, and the predicted points larger than the semantic threshold are processed according to the actual prediction results. The real category annotated by the target class label is added to generate a pseudo semantic label.

[0088] S402: Using pseudo semantic labels, remove mask points belonging to the background class in the teacher network prediction mask;

[0089] Use the teacher network to predict the mask score and calculate the difference between the average mask prediction score and the maximum prediction score;

[0090] Calculate the ratio between the predicted average score and the predicted maximum score difference. If it is greater than a fixed value, take the minimum mask score as the mask threshold. Otherwise, take the maximum value between the minimum mask score and the fixed threshold as the mask threshold.

[0091] Remove the mask smaller than the mask threshold, treat the mask larger than the mask threshold as the actual predicted mask, and add the real mask annotated with the target class label to generate a pseudo instance mask;

[0092] S403: Use the teacher network to predict the instance mask score as the pseudo mask score, and calculate the intersection-over-union ratio between the pseudo instance mask corresponding to the pseudo mask score and the instance mask corresponding to the student network as the loss weight;

[0093] S404: If the network has a mask center position prediction branch, use the teacher network to predict the center position and calculate the average and standard deviation of the predicted center;

[0094] Use the pseudo instance mask to calculate the difference between the mask center and the average value of the predicted center. If it is greater than the standard deviation, remove the calculated mask center. If it is less than the standard deviation, retain the center as the pseudo center position.

[0095] S405: Using the pseudo semantic label, pseudo instance mask, pseudo mask score and pseudo center position to compare with the student network prediction results respectively, and calculate the training loss.

[0096] In this embodiment, for step S401, the currently trained sub-subset includes the target class and the background class, wherein the background class includes the pseudo background class (conventional objects, which have been learned in the pre-training task) and the true background class; the teacher network can be used to perform semantic prediction results for conventional objects; the set dynamic threshold can select an appropriate threshold according to the probability distribution. When all probabilities of a certain class (or a certain class instance) are small, a fixed threshold is used; when there are points with higher probabilities and the distribution within the class is more concentrated, the prediction of this class is considered to be more credible, and the minimum value is used as the threshold; when there are points with higher probabilities and the distribution within the class is more dispersed, the low-score prediction is considered to be unreliable, and the maximum value of the minimum value and the fixed threshold is used as the threshold to eliminate the low-score threshold; based on the dynamic semantic threshold, semantic labels less than the semantic threshold are considered to be low-quality and unreliable, and are classified as background classes for elimination, semantic labels greater than the mask threshold are retained, and the real categories annotated with the target class labels are added and added to generate pseudo-semantic labels.

[0097] In step S402, based on the pseudo semantic label generation process, the mask points classified as true background in the teacher network prediction mask are first removed; then the pseudo background masks smaller than the mask threshold are also removed, and the masks larger than the mask threshold are retained as the actual predicted masks, and the real masks annotated with the target class label are added to generate the pseudo instance mask;

[0098] For step S403, the larger the intersection-over-union ratio between the pseudo-instance mask corresponding to the pseudo-mask score and the instance mask corresponding to the student network, the greater the fit between the pseudo-instance mask and the instance mask corresponding to the student network, and the greater the training loss of the design.

[0099] For step S404, if the network has a mask center position prediction branch, the difference between the mask pseudo center calculated by the pseudo instance mask and the average value of the teacher network predicted center is compared with the standard deviation to determine whether to retain the center as the pseudo center position.

[0100] For step S405, the pseudo semantic label, pseudo instance mask, pseudo mask score and pseudo center position are compared with the results predicted by the student network to obtain the training loss.

[0101]

[0102] Where: They are calculated from the semantic judgment results, instance mask, instance center and instance score respectively.

[0103] To achieve better training results, the difference comparison method is the same as the training loss acquisition method in the network (teacher network) pre-training. Then, while the network is learning new classes, it reviews the knowledge of old classes. Under the guidance of the teacher network, the network can construct the distillation loss without being based on the old classes, thus solving the problem of the continued existence of old classes greatly increasing computer memory.

[0104] like Figure 5-6 As shown, as another preferred embodiment of the present invention, a method for segmenting point cloud data for disaster scenes is provided. In step 105, the point cloud features of the teacher network and the student network are simultaneously layered and pooled to obtain layered pooled features of the teacher network and the student network respectively. The step of calculating the distillation loss based on the difference between the layered pooled features of the two includes:

[0105] S501 uses a hole average pooling layer with different jumps to extract multi-scale pooling features for low-level features;

[0106] S502 uses a second-order pooling layer to extract context-related pooling features for high-level features;

[0107] S503 calculates the L2 loss as the distillation loss based on the pooled features of each layer of the teacher network and the student network;

[0108] In this embodiment, for step 501, the low-level convolutional features contain rich spatial details (such as edges, textures, and local structural information), but are sensitive to position and have a low level of semantic abstraction. Therefore, atrous spatial pyramid pooling (ASPP) is used to process the features. ASPP can effectively capture multi-scale local contextual information without significantly reducing the spatial resolution of the feature map by applying atrous convolution operations with different expansion rates in parallel.

[0109] For step 502, the high-level convolutional features have strong semantics (representing object parts, wholes, or scene concepts) and spatial robustness (insensitive to position and scale changes) due to the layer-by-layer abstraction of the deep network. However, the spatial detail information has been greatly attenuated, so global second-order pooling is used to process the features. It describes how different feature channels change together by calculating the covariance matrix.

[0110] In step 503, the difference between the pooled features of each layer of the teacher network and the student network is evaluated using the L2 loss. As the number of training rounds increases, the weight of the KL divergence L2 loss gradually decays to prevent the original knowledge from limiting network learning. At the same time, the deeper the feature, the more important it is and the higher the weight is. The loss weight of each layer is as follows:

[0111]

[0112] Where: λ is a constant; n l , N l is the current layer and the total number of layers of this type of pooling method; γ is a constant less than 1, n e , N e is the current number of training rounds and the total number of rounds;

[0113] By designing dynamic loss weights, different weights are applied to the distillation loss at different layers and rounds to obtain the final distillation loss. The teacher network features can then be used to guide student feature generation. While the network is learning new classes, it can review the knowledge of old classes without having to rely on old classes to build distillation loss, thus solving the problem of the continued existence of old classes significantly increasing computer memory.

[0114] Add the original training loss and the distillation loss to get the total loss l:

[0115] l=l train +Ση KDi l KDi

[0116] Where: l KDi is the distillation loss per layer, calculated from the L2 loss.

[0117] Table 5 shows the experimental results of using pseudo-label generation, data resampling, and feature distillation in the continuous learning process. It can be concluded from the quantitative data given in Table 5 that by adding data resampling on the basis of pseudo-label generation, the network can effectively improve its ability to distinguish new classes. After generating pseudo-labels for training, through data resampling and distilling the features between the teacher network and the student network, the student network can be guided by the teacher network features, thereby alleviating the problem of catastrophic forgetting.

[0118] Table 5. Intersection-over-Union results of ablation experiments on the continuous learning segmentation framework for disaster scene point clouds

[0119]

[0120] Figure 7The qualitative results of the continuous learning effect are given, where the leftmost are the original image, semantic label and mask label respectively, (a) is the result of adding the pseudo label generation module, (b) is the result of continuing to add the data resampling module, and (c) is the result of continuing to add the feature distillation module. Figure 7 It can be seen that in the continuous learning process, on the basis of generating pseudo-labels, data resampling and feature distillation are successively added, so that the predicted semantic-level predictions and instance-level masks are increasingly close to the real categories and masks recently annotated, which further illustrates the effectiveness of the continuous learning method in the present invention.

[0121] The present invention transfers the pre-trained network to the student network and the teacher network, and trains the student network using the data in the secondary subset. In the continuous learning process, the teacher network prediction results are used, combined with the target class real label annotation to generate pseudo labels, and the difference between the pseudo labels and the student network prediction results is used to obtain the training loss; by synchronously layering and pooling the features extracted by the teacher network and the student network, the difference between the pooled features is used to obtain the distillation loss, and then the loss of the network continuous learning is obtained. Since the training loss and the distillation loss are both associated with the teacher network, and the teacher network has been trained on the main subset (including conventional objects) during pre-training, the teacher network can be used to guide the student network to learn newly added post-disaster objects or objects of focus. The teacher network features are used to guide the generation of student network features, which helps the student network to retain a review of old knowledge while learning new classes, alleviate the problem of catastrophic forgetting, and thus achieve the network without constructing losses based on old classes, solving the problem that the continued increase and decrease of old classes greatly increases computer memory.

[0122] After continuous learning was completed for all sub-subsets, the optimized student network was used to process the point cloud data on the validation set. The obtained semantic-level predictions and instance-level masks were compared with the annotated true categories and masks to evaluate the point cloud data segmentation performance. The IoU data obtained experimentally in Table 6 show that among the four networks (BSeg, Mask3D, DKNet, and SoftGroup), the DKNet network achieved the highest IoU value for predictions, indicating the highest degree of agreement with the annotated true categories and masks, and the best performance in segmenting and distinguishing objects in disaster scenes. Furthermore, using the DKNet network, the IoU values ​​for vegetation, water bodies, vehicles, houses, and ruins approached or exceeded 0.7, indicating that the continuous learning described in this invention, used to optimize and improve the DKNet network after training, can effectively segment and distinguish objects in disaster scenes, thereby providing strong support for emergency rescue decision-making in disaster areas.

[0123] Table 6 Intersection-comparison results of the comparative experiment on the continuous learning segmentation framework for disaster scene point clouds

[0124]

[0125] like Figure 8-9 FIG. 1 is a segmentation system for point cloud data of disaster scenes provided by another embodiment of the present invention, and the system includes:

[0126] S601: a point cloud data acquisition and labeling unit, which is used to acquire manually labeled conventional urban point cloud data; and acquire and manually label open disaster scene point cloud data, wherein the open disaster scene point cloud data includes a learning set and a validation set, wherein the learning set includes a primary subset and a secondary subset, wherein the labels of the primary subset only include common objects in general outdoor scenes, and the labels of the secondary subset only include one or more disaster scene objects;

[0127] S602: a network pre-training unit, wherein the network pre-training unit is used to pre-train the network on manually labeled conventional urban point cloud data to obtain network pre-training parameters;

[0128] S603: a network migration unit, configured to migrate the network and pre-trained parameters on a primary subset of manually labeled open disaster scene point cloud data;

[0129] S604: a network continuous learning unit, wherein the network continuous learning unit is configured to use the migrated network as a teacher network and a student network, respectively, and use the student network to perform continuous learning on the secondary subset to obtain parameters of the student network. The network continuous learning unit specifically includes:

[0130] S6041 initialization module: used for initializing the teacher network and the student network using the network parameters obtained from the previous sub-subset continuous learning when the student network starts continuous learning in each sub-subset, and then freezing the teacher network;

[0131] S6042 Sub-subset data resampling module: used to resample the point cloud data in the sub-subset, where background samples belonging to common objects are undersampled and target samples belonging to disaster scene objects are oversampled;

[0132] S6043 point cloud feature extraction module: used to synchronously input the resampled point cloud data into the teacher network and student network feature extractors to obtain the point cloud features of the teacher network and the student network respectively;

[0133] S6044 training loss calculation module: used to generate pseudo labels using the prediction results of the teacher network in combination with the original true labels of the target class; and calculate the training loss using the pseudo labels and the prediction results of the student network;

[0134] S6045 distillation loss calculation module: used to synchronously perform layered pooling on the point cloud features of the teacher network and the student network to obtain pooled features; and calculate the distillation loss using the difference between the pooled features of the two;

[0135] S6046 student network parameter optimization module; used to train the student network through backpropagation using the total loss obtained by adding the distillation loss and the training loss, optimize the student network parameters, and obtain the optimized student network parameters after completing the training on all sub-subsets;

[0136] S605 network application unit, the network application unit uses the optimization parameters to initialize the student network and then processes the point cloud data on the validation set, and compares the obtained semantic-level predictions and instance-level masks with the labeled true categories and masks to evaluate the point cloud data segmentation effect.

[0137] In this embodiment, the description of the above system may correspond to the segmentation method. For the description of the above segmentation method, please refer to the above text and will not be repeated here.

[0138] As another preferred embodiment of the present invention, a segmentation system for disaster scene point cloud data is provided, wherein the network migration unit specifically includes:

[0139] Migration initialization module: initialize the network using pre-trained parameters;

[0140] Domain transfer module: By connecting a multi-layer perceptron network after the network feature extractor and using LeakyReLU as the activation function, the feature offset extracted by the network on each main subset is obtained through learning; the network extracted features and the feature offset are added to obtain new features;

[0141] Transfer fine-tuning module: Use the new features to fine-tune the parameters and network structure of the pre-trained network migrated on each main subset at 0.2-0.3 times the pre-training learning rate.

[0142] In this embodiment, the description of the above system may correspond to the segmentation method. For the description of the above segmentation method, please refer to the above text and will not be repeated here.

[0143] As another preferred embodiment of the present invention, a segmentation system for disaster scene point cloud data is provided, wherein the training loss calculation module includes:

[0144] Pseudo-semantic label generation module: The teacher network is used to predict semantic probabilities. The ratio between the average predicted probability and the maximum predicted probability difference for each object type is calculated. If the ratio is greater than a fixed value, the minimum predicted probability is taken as the semantic threshold. Otherwise, the maximum value between the minimum predicted probability and the fixed threshold is taken as the semantic threshold. Prediction points below the semantic threshold are classified as background, and prediction points above the semantic threshold are processed according to the actual predicted category. The actual category annotated by the target class label is added to generate a pseudo-semantic label.

[0145] Pseudo-instance mask generation module: Using pseudo-semantic labels, the mask points belonging to the background class are removed from the teacher network predicted mask. The teacher network predicts the mask score and calculates the ratio between the predicted average score and the predicted maximum score difference. If it is greater than a fixed value, the minimum mask score is taken as the mask threshold. Otherwise, the maximum value between the minimum mask score and the fixed threshold is taken as the mask threshold. Masks smaller than the mask threshold are removed, and masks larger than the mask threshold are treated as the actual predicted mask. The real mask annotated with the target class label is added to generate a pseudo-instance mask.

[0146] Pseudo-mask score generation module: Calculate the maximum intersection-over-union (IoU) of the student network predicted mask and the pseudo-instance mask, use the mask score of the pseudo-instance mask as the pseudo-mask score, and use the IoU as the loss weight;

[0147] Pseudo center position generation module: If the network has a mask center position prediction branch, use the teacher network to predict the center position and calculate the average and standard deviation of the predicted center; use the pseudo instance mask to calculate the difference between the mask center and the predicted center average. If it is greater than the standard deviation, remove the calculated mask center; if it is less than the standard deviation, retain the center as the pseudo center position;

[0148] Training loss acquisition module: The pseudo semantic label, pseudo instance mask, pseudo mask score and pseudo center position are compared with the student network prediction results to calculate the training loss

[0149] In this embodiment, the description of the above system may correspond to the segmentation method. For the description of the above segmentation method, please refer to the above text and will not be repeated here.

[0150] As another preferred embodiment of the present invention, a segmentation system for disaster scene point cloud data is provided, wherein the distillation loss calculation module includes:

[0151] The low-level pooling module uses a dilated average pooling layer with different jumps to extract multi-scale pooling features;

[0152] High-level pooling module: uses a second-order pooling layer to extract context-dependent pooling features;

[0153] Distillation loss acquisition module: Based on the pooled features of each layer of the teacher network and the student network, the L2 loss is calculated as the distillation loss.

[0154] In this embodiment, the description of the above system may correspond to the segmentation method. For the description of the above segmentation method, please refer to the above text and will not be repeated here.

[0155] Those skilled in the art will appreciate that all or part of the processes in the above-described embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-described methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0156] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for segmenting point cloud data for disaster scenes, characterized in that: The method comprises the following steps: Obtain manually annotated conventional urban point cloud data; obtain manually annotated open disaster scene point cloud data, wherein the open disaster scene point cloud data includes a learning set and a validation set, wherein the learning set includes a primary subset and a secondary subset, wherein the annotations of the primary subset only include common objects in general outdoor scenes, and the annotations of the secondary subset only include one or more disaster scene objects; Constructing a network structure for segmenting the conventional urban point cloud data; pre-training the network on the conventional urban point cloud data to obtain network pre-training parameters; migrating the pre-trained network structure and pre-training parameters to the main subset; The migrated networks are used as the teacher network and the student network respectively, and the student network is used to continuously learn in the sub-subsets. When each sub-subset starts continuous learning, the teacher network and the student network are initialized using the network parameters obtained from the continuous learning of the previous sub-subset, and then the teacher network is frozen; Resample the point cloud data in the secondary subset, where background samples belonging to common objects are undersampled and target samples belonging to disaster scene objects are oversampled; The resampled point cloud data is synchronously input into the teacher network and student network feature extractors to obtain the point cloud features of the teacher network and the student network respectively; The obtained point cloud features of the teacher network and the student network are input into the decoder to generate prediction results of the teacher network and the student network respectively; the prediction results of the teacher network are combined with the original true label annotation of the target class to generate pseudo labels, and the training loss is calculated based on the difference between the pseudo labels and the prediction results of the student network; the point cloud features of the teacher network and the student network are simultaneously layered pooled to obtain the layered pooled features of the teacher network and the student network respectively, and the distillation loss is calculated based on the difference between the layered pooled features of the two; The distillation loss and the training loss are added together to obtain a total loss, back-propagation is performed on the student network based on the total loss, and the student network parameters are optimized. After completing the training on all sub-subsets, the optimized parameters of the student network are obtained; The student network is initialized with the optimized parameters and then the point cloud data on the validation set is processed to obtain semantic-level predictions and instance-level masks. The semantic-level predictions and instance-level masks are compared with the true categories and masks marked on the validation set to evaluate the point cloud data segmentation effect.

2. The method for segmenting point cloud data for disaster scenes according to claim 1, characterized in that: The network structure includes a structure built with a sparse U-Net network as a skeleton network.

3. The method for segmenting point cloud data for disaster scenes according to claim 1, characterized in that: The step of migrating the network and pre-trained parameters on the main subset includes: Initialize the network using pre-trained parameters; A multi-layer perceptron network is connected after the network feature extractor, and LeakyReLU is used as the activation function. The offset of the network extracted features on each main subset is obtained through training and learning; Adding the network extracted features and the feature offset to obtain new features; Based on the new features, the pre-trained network parameters and network structure migrated on each main subset are slightly adjusted at 0.2-0.3 times the pre-training learning rate.

4. The method for segmenting point cloud data for disaster scenes according to claim 1, characterized in that: The step of resampling the point cloud data in the secondary subset includes: Count the number of points belonging to the target class and background class in the point cloud data; Calculate the ratio of the number of target class points to the number of background class points as the sampling probability of the background class sample, and take the value 1 as the sampling probability of the target class sample; Sampling generates training samples.

5. The method for segmenting point cloud data for disaster scenes according to claim 1, characterized in that: The steps of generating pseudo labels by using the prediction results of the teacher network in combination with the original true label annotations of the target class, and calculating the training loss based on the difference between the pseudo labels and the prediction results of the student network include: Use the teacher network to predict semantic probabilities and calculate the difference between the average predicted probability and the maximum predicted probability for each type of object; Calculate the ratio between the predicted average probability and the predicted maximum probability difference. If it is greater than a fixed value, take the minimum predicted probability as the semantic threshold. Otherwise, take the maximum value between the minimum predicted probability and the fixed threshold as the semantic threshold. The predicted points smaller than the semantic threshold are classified as background, and the predicted points larger than the semantic threshold are processed according to the actual prediction results. The real category annotated by the target class label is added to generate a pseudo semantic label. Using pseudo semantic labels, we remove the mask points belonging to the background class from the teacher network prediction mask. Use the teacher network to predict the mask score and calculate the difference between the average mask prediction score and the maximum prediction score; Calculate the ratio between the predicted average score and the predicted maximum score difference. If it is greater than a fixed value, take the minimum mask score as the mask threshold. Otherwise, take the maximum value between the minimum mask score and the fixed threshold as the mask threshold. Remove the mask smaller than the mask threshold, treat the mask larger than the mask threshold as the actual predicted mask, and add the real mask annotated with the target class label to generate a pseudo instance mask; Use the teacher network to predict the instance mask score as the pseudo mask score, and calculate the intersection-over-union ratio between the pseudo instance mask corresponding to the pseudo mask score and the instance mask corresponding to the student network as the loss weight; If the network has a mask center position prediction branch, use the teacher network to predict the center position and calculate the average and standard deviation of the predicted center; Use the pseudo instance mask to calculate the difference between the mask center and the average value of the predicted center. If it is greater than the standard deviation, remove the calculated mask center. If it is less than the standard deviation, retain the center as the pseudo center position. The pseudo semantic label, pseudo instance mask, pseudo mask score and pseudo center position are respectively compared with the prediction results of the student network to calculate the training loss.

6. The method for segmenting point cloud data for disaster scenes according to claim 1, characterized in that: The step of synchronously performing layered pooling on the point cloud features of the teacher network and the student network to obtain layered pooled features of the teacher network and the student network, respectively, and calculating the distillation loss based on the difference between the layered pooled features of the two comprises: For low-level features, a dilated average pooling layer with different jumps is used to extract multi-scale pooling features; For high-level features, a second-order pooling layer is used to extract context-dependent pooling features; Based on the pooled features of each layer of the teacher network and the student network, the L2 loss is calculated as the distillation loss.

7. A segmentation system for point cloud data of disaster scenes, characterized by: The system comprises: Point cloud data acquisition and annotation unit: The point cloud data acquisition and annotation unit is used to acquire manually annotated conventional urban point cloud data; and acquire and manually annotate open disaster scene point cloud data. The open disaster scene point cloud data includes a learning set and a validation set. The learning set includes a primary subset and a secondary subset. The labels of the primary subset only include common objects in general outdoor scenes, and the labels of the secondary subset only include one or more disaster scene objects. A network pre-training unit, which is used to pre-train the network on manually labeled conventional urban point cloud data to obtain network pre-training parameters; A network migration unit, configured to migrate the network and pre-trained parameters on a primary subset of manually labeled open disaster scene point cloud data; A network continuous learning unit is configured to use the migrated network as a teacher network and a student network, respectively, and to use the student network to perform continuous learning on the secondary subset to obtain student network parameters. The network continuous learning unit specifically includes: Initialization module: used for initializing the teacher network and the student network using the network parameters obtained from the previous subset continuous learning when the student network starts continuous learning in each subset, and then freezing the teacher network; Sub-subset data resampling module: used to resample the point cloud data in the sub-subset, where background samples belonging to common objects are undersampled and target samples belonging to disaster scene objects are oversampled; Point cloud feature extraction module: used to synchronously input the resampled point cloud data into the teacher network and student network feature extractors to obtain the point cloud features of the teacher network and the student network respectively; Training loss calculation module: used to generate pseudo labels using the prediction results of the teacher network in combination with the original true labels of the target class; and calculate the training loss using the pseudo labels and the prediction results of the student network; Distillation loss calculation module: used to synchronously perform layered pooling on the point cloud features of the teacher network and the student network to obtain pooled features; and calculate the distillation loss using the difference between the pooled features of the two; The network parameter optimization module is used to train the student network through backpropagation using the total loss obtained by adding the distillation loss and the training loss, optimize the student network parameters, and obtain the optimized parameters of the student network after completing the training on all sub-subsets. Network application evaluation unit: The network application unit is used to use the optimization parameters to initialize the student network and process the point cloud data on the validation set, and compare the obtained semantic-level predictions and instance-level masks with the labeled true categories and masks to evaluate the point cloud data segmentation effect.

8. The segmentation system for disaster scene point cloud data according to claim 7, characterized in that: The network migration unit specifically includes: Migration initialization module: initialize the network using pre-trained parameters; Domain transfer module: By connecting a multi-layer perceptron network after the network feature extractor and using LeakyReLU as the activation function, the feature offset extracted by the network on each main subset is obtained through learning; the network extracted features and the feature offset are added to obtain new features; Transfer fine-tuning module: Using new features, fine-tune the parameters and network structure of the pre-trained networks transferred on each main subset at 0.2-0.3 times the pre-training learning rate.

9. The segmentation system for disaster scene point cloud data according to claim 7, characterized in that: The training loss calculation module includes: Pseudo-semantic label generation module: The teacher network is used to predict semantic probabilities. The ratio between the average predicted probability and the maximum predicted probability difference for each object type is calculated. If the ratio is greater than a fixed value, the minimum predicted probability is taken as the semantic threshold. Otherwise, the maximum value between the minimum predicted probability and the fixed threshold is taken as the semantic threshold. Prediction points below the semantic threshold are classified as background, and prediction points above the semantic threshold are processed according to the actual predicted category. The actual category annotated by the target class label is added to generate a pseudo-semantic label. Pseudo-instance mask generation module: Using pseudo-semantic labels, the mask points belonging to the background class are removed from the teacher network predicted mask. The teacher network predicts the mask score and calculates the ratio between the predicted average score and the predicted maximum score difference. If it is greater than a fixed value, the minimum mask score is taken as the mask threshold. Otherwise, the maximum value between the minimum mask score and the fixed threshold is taken as the mask threshold. Masks smaller than the mask threshold are removed, and masks larger than the mask threshold are treated as the actual predicted mask. The real mask annotated with the target class label is added to generate a pseudo-instance mask. Pseudo-mask score generation module: Calculate the maximum intersection-over-union (IoU) of the student network predicted mask and the pseudo-instance mask, use the mask score of the pseudo-instance mask as the pseudo-mask score, and use the IoU as the loss weight; Pseudo center position generation module: If the network has a mask center position prediction branch, use the teacher network to predict the center position and calculate the average and standard deviation of the predicted center; use the pseudo instance mask to calculate the difference between the mask center and the predicted center average. If it is greater than the standard deviation, remove the calculated mask center; if it is less than the standard deviation, retain the center as the pseudo center position; Training loss acquisition module: using the pseudo semantic label, pseudo instance mask, pseudo mask score and pseudo center position to compare with the student network prediction results respectively, and calculating the training loss.

10. The segmentation system for disaster scene point cloud data according to claim 7, characterized in that: The distillation loss calculation module includes: The low-level pooling module uses a dilated average pooling layer with different jumps to extract multi-scale pooling features; High-level pooling module: uses a second-order pooling layer to extract context-dependent pooling features; Distillation loss acquisition module: Based on the pooled features of each layer of the teacher network and the student network, the L2 loss is calculated as the distillation loss.

Images