An efficient multi-view stereo reconstruction method and system for outdoor scenes

Abstract

The application discloses an efficient multi-view stereo reconstruction method and system for an outdoor scene, and belongs to the field of multi-view three-dimensional reconstruction. The specific steps are as follows: preparing a multi-view stereo reconstruction data set for model training; constructing a multi-view stereo reconstruction network including an adaptive feature extraction converter module based on a large kernel attention, a cost regularization module and a depth estimation module; training and fine-tuning the network model to obtain a final model; and building a multi-view stereo reconstruction system and equipment. The application can solve the problems of poor reconstruction result accuracy and low completeness of existing multi-view stereo reconstruction methods when facing outdoor large-scale scenes, reduce the training and inference time and calculation cost of multi-view stereo reconstruction, and lay an important foundation for the application of outdoor large-scale scene image data in the field of multi-view three-dimensional reconstruction and the development of three-dimensional reconstruction technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of multi-view stereo reconstruction, computer graphics and computer vision, and in particular to an efficient multi-view stereo reconstruction method and system for outdoor scenes. Figure Three BACKGROUND TECHNICAL FIELD

[0002] In recent years, multi-view stereo reconstruction based on deep learning has attracted more and more interest due to its wide application in the fields of autonomous driving, virtual reality, augmented reality, etc. Although existing multi-view stereo reconstruction methods have made great progress in indoor controllable environments, these methods have difficulty in dealing with problems such as occlusion between objects, non-Lambertian surfaces, uneven lighting, and weak texture regions in outdoor complex scenes. Although recent research has introduced a visual transformer module to enable the network to focus on global context, enhancing the quality of the three-dimensional model, due to the quadratic complexity of self-attention, these multi-view stereo reconstruction methods with visual transformers consume a large amount of running time and high memory capacity, severely affecting the development and application of multi-view stereo three-dimensional reconstruction technology. Therefore, it is of great significance to design an efficient multi-view stereo reconstruction method for outdoor scenes.

[0003] Chinese patent publication No. CN115908723A, entitled "Limit-guided multi-view stereo reconstruction method based on interval perception", first captures the global context information within and across images through a ray-guided aggregation sub-module, then generates a cost volume using a visible fusion sub-module, and finally calculates the loss using a depth estimation module. Although this method addresses the problem of weak cross-information feature extraction capability of multi-view stereo reconstruction networks to some extent, it cannot solve problems such as occlusion between objects, non-Lambertian surfaces, uneven lighting, and weak texture regions, nor can it effectively handle problems such as low time efficiency, large model parameter quantity, and large model computation in the process of multi-view stereo reconstruction based on outdoor scenes. Therefore, designing a method that can efficiently complete the multi-view stereo reconstruction task in real outdoor scenes is the key problem that the present application focuses on solving. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides an efficient multi-view stereo reconstruction method and system for outdoor scenes, which can quickly calculate a high-quality dense point cloud model of the reconstructed scene from large-scale images in outdoor complex scenes, and has faster running time, less computational cost, and smaller storage cost than existing multi-view stereo methods, promoting the progress of multi-view stereo reconstruction technology for outdoor scenes. At the same time, the present application can improve the stability and performance of training, and has high generalizability.

[0005] The application specifically adopts the following technical solutions to achieve the above-mentioned purpose.

[0006] A high-efficiency multi-view stereo reconstruction method for outdoor scenes, comprising the following steps:

[0007] Step 1, preparing a data set: preparing a first multi-view stereo reconstruction data set and a second multi-view stereo reconstruction data set, and dividing the two data sets into a training data set and a test data set respectively; the training data set is used to train the network model, and the test data set is used to evaluate the model training effect;

[0008] Step 2, constructing a network model: the network is composed of a backbone network, an adaptive feature extraction converter module based on large kernel attention, a cost regularization module and a depth estimation module.

[0009] Step 3, training the network model: inputting the data set prepared in step 1 into the network model constructed in step 2 for training; by minimizing the loss between the output of the network and the label, a better depth prediction result is achieved;

[0010] Step 4, screening the optimal network model; using the test data set as input, evaluating and screening the optimal network model parameters;

[0011] Step 5, fine-tuning the network model; fine-tuning the network model with the second multi-view stereo reconstruction data set, so that the model is generalized from the indoor scene with controllable light to the outdoor large-scale scene.

[0012] Step 6, saving the network model; after the network training is completed, the optimal network parameters are saved, and then the multi-view scene image to be three-dimensionally reconstructed is input into the network to obtain the predicted depth map, and through depth filtering and fusion operation, the reconstructed three-dimensional point cloud is obtained.

[0013] Further, the first multi-view stereo reconstruction data set is a DTU data set, and the multi-view images, depth maps and corresponding camera internal and external parameters of each scene are obtained.

[0014] Further, the backbone network is a feature pyramid feature extraction network, which is used to extract multi-scale features of the input reference image and the source image;

[0015] Further, the adaptive feature extraction converter module based on large kernel attention includes an adaptive large kernel attention block, a hybrid feedforward neural network block, a linear cross attention block and an up-sampling module.

[0016] The adaptive large kernel attention block combines multiple scale convolution kernels with different expansion coefficients into a large size convolution kernel, so that the model can effectively reduce the calculation cost and improve the calculation efficiency of the model without sacrificing performance. On this basis, by introducing deformable convolution, the large size convolution kernel can adaptively learn the multi-view target of different scales to obtain a larger receptive field, so that the model can better understand the target context information in the outdoor scene and improve the extraction ability of complex features.

[0017] The mixed feedforward neural network block better aggregates global and multi-scale local information, and extends the design of layer normalization to a general form, promoting the optimization of the skip connection and improving the network feature extraction capability.

[0018] The linear cross-attention block enables the network to learn the feature correlation between different view images, helping the model better understand the spatial structure of the object and improving the accuracy and stability of the stereo reconstruction.

[0019] Further, the cost regularization module replaces the three-dimensional convolution kernel with a three-dimensional adaptive large kernel attention, which can expand the receptive field without increasing the model parameters and computational complexity, so that the model can better capture the global information in the scene and improve the accuracy of stereo reconstruction.

[0020] Further, the depth estimation module uses maximum probability depth estimation to select the depth value with the highest probability value from the probability volume for each pixel position as the depth value of the pixel position in the depth estimation map.

[0021] Further, the preset threshold in training the network model includes a loss function preset value, an iteration number preset value, and an optimization learning rate preset value.

[0022] Further, the loss function is a focal loss to strengthen the single hot spot supervision in the blurred area.

[0023] Further, the process of training the network model further includes evaluating the accuracy and integrity of the algorithm reconstruction result by an evaluation index.

[0024] Further, the second multi-view stereo reconstruction dataset is a BlendedMVS dataset.

[0025] An outdoor scene-oriented multi-view stereo reconstruction system, the system comprising:

[0026] An image acquisition module for acquiring training data and test data; the training data and test data include multi-view images and corresponding depth images.

[0027] Model building module, used to build the network model proposed by this method and initialize the model weights;

[0028] A dataset and model loading module, used to load input images and network models onto hardware devices for model training; the hardware devices refer to image processors;

[0029] The model training module is used to perform supervised model training on the network model constructed by this method for the multi-view images of each scene in the training data to obtain the optimal model weights;

[0030] The depth estimation module is used to estimate the depth of the multi-view images of each scene in the test data through a network model equipped with the optimal model weights to obtain a multi-view depth map of each scene;

[0031] The depth filtering and fusion module is used to obtain a 3D point cloud file for each scene after performing depth filtering and fusion operations on the multi-view depth map predicted by the network;

[0032] The point cloud visualization module is used to parse and process 3D point cloud files through the MeshLab software platform to achieve 3D point cloud visualization;

[0033] A storage medium is used to store a multi-view stereo reconstruction system.

[0034] Beneficial effects

[0035] Compared with the existing technology, the present invention provides an efficient multi-view stereo reconstruction method and system for outdoor scenes, which has the following beneficial effects:

[0036] 1. This paper designs a new multi-view stereo reconstruction network architecture, constructs a multi-view feature extraction and probabilistic volume prediction network based on adaptive large-core attention, solves the problems of long training and inference time and high computational cost of existing multi-view stereo reconstruction methods when facing large-scale outdoor scenes, and provides a new way for outdoor large-scale scene image data to be reconstructed in multiple views. Figure Three It lays an important foundation for the application of 3D reconstruction in the field and the development of 3D reconstruction technology.

[0037] 2. The adaptive large-core attention module proposed in this paper takes into account the global and local information capture capabilities of convolution and self-attention mechanisms while avoiding their limitations of high quadratic computational complexity, redundant model parameters, and fixed receptive field size. Without compromising the accuracy and completeness of stereo reconstruction, it accelerates model training and inference, improving computational efficiency, particularly when processing large-scale outdoor scene data.

[0038] 3. The hybrid feedforward neural network module disclosed by the present application utilizes multi-scale convolution kernels and layer normalization expansion to capture the relationship between different scale features of multi-view targets in the same scene, further enhances the information interaction between different channels, aggregates the corresponding relationship between different depth feature mappings and targets, and thus obtains more rich feature information.

[0039] 4. The cost volume converter based on three-dimensional adaptive large kernel attention disclosed by the present application can improve the feature expression ability, information interaction ability and receptive field coverage range of the model, and thus shows stronger stability and generalization when facing complex scenes and input changes.

[0040] 5. The multi-view stereo reconstruction network disclosed by the present application shows good effect in the DTU dataset and the BlendedMVS dataset, while ensuring the reconstruction quality, the memory occupation, the calculation amount and the inference time are reduced by 33%, 32% and 41% respectively, effectively reducing the training and use cost of the model and improving the reconstruction efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a flow chart of an efficient multi-view stereo reconstruction method for an outdoor scene;

[0042] Figure 2 It is a general architecture diagram of an efficient multi-view stereo reconstruction network model for an outdoor scene;

[0043] Figure 3 It is a structure diagram of the adaptive feature extraction converter module based on large kernel attention described in the present application;

[0044] Figure 4 It is a specific composition schematic diagram of all adaptive large kernel attention modules described in the present application;

[0045] Figure 5 It is a specific composition schematic diagram of all hybrid feedforward neural network modules described in the present application;

[0046] Figure 6 It is a specific composition schematic diagram of all cost volume converters based on three-dimensional adaptive large kernel attention described in the present application;

[0047] Figure 7 It is a related index comparison schematic diagram of the method proposed in the present application;

[0048] Figure 8 It is a main module schematic diagram of a multi-view stereo reconstruction system for an outdoor scene;

[0049] Figure 9 It is a dense point cloud model diagram provided by the embodiment of the present application;

[0050] Figure 10 Schematic diagram of the internal structure of an electronic device for multi-view stereo reconstruction of outdoor scenes. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0052] Example

[0053] like Figure 1 As shown, an efficient multi-view stereo reconstruction method for outdoor scenes in this embodiment includes the following steps:

[0054] Step 1, prepare the data set; prepare the first multi-view stereo reconstruction data set and the second multi-view stereo reconstruction data set, obtain the multi-view images, depth maps and corresponding internal and external parameters of the camera for each scene; divide the two data sets into training data sets and test data sets respectively; the training data set is used to train the network model, and the test data set is used to evaluate the model training effect; the first multi-view stereo reconstruction data set uses the DTU data set; the DTU data set is a large-scale multi-view stereo reconstruction data set with precise camera trajectory collected in a controlled laboratory environment, with an image resolution of 640×512; the data set contains 119 scenes in 49 views under 7 different lighting conditions, each set of data consists of RGB images and phase and provides a true point cloud scanned using structured light; we selected 97 scenes as the training set of the network model and 22 scenes as the test set of the network model;

[0055] Step 2: Build a network model. We designed an efficient multi-view stereo reconstruction network model for outdoor scenes. The model architecture is as follows: Figure 2 As shown;

[0056] First, we use the feature pyramid as the backbone network to extract the input reference image I0∈R H×W×3 and the source image Multi-scale features Where i represents the i-th input view and N represents the number of input images. To expand the capabilities of long-term context aggregation and global feature interaction across scales, we design an adaptive feature extraction converter module based on large-core attention, which uses a combination of adaptive large-core attention modules and cross-attention modules within and across multi-view images at each stage. Secondly, the source image features are mapped to the reference image through homography transformation and fused to obtain a three-dimensional cost volume V, where V∈R H×W×C×D Then, we use the cost volume transformer based on 3D adaptive large kernel attention to perform regularization operation and generate a probability volume P, where P∈R H×W×DFinally, a Softmax operation is applied to the probability volume to estimate the depth of the reference image.

[0057] The specific content of the above module is as follows:

[0058] As shown in Figure 3 , the adaptive feature extraction converter module based on large kernel attention is composed of a self-attention module, an up-sampling module and a cross-attention module; the self-attention module adopts the converter architecture, uses the adaptive large kernel attention module and the hybrid feedforward neural network module proposed in this work to efficiently capture the global and local context features within the image, so that the network pays more attention to the challenging areas in multi-view stereo matching, such as weak texture and non-Lambertian surfaces; the up-sampling module adopts bilinear interpolation up-sampling to fuse multi-scale features; the cross-attention module uses linear cross-attention blocks to capture cross-image context information, and also uses hybrid feedforward neural network modules to aggregate global and local context information;

[0059] The specific composition diagram of the adaptive large kernel attention module is shown in Figure 4 ; a large kernel convolution can be decomposed into a depth separable convolution, a depth separable dilated convolution and a pointwise convolution. Specifically, we use a 5x5 depth separable convolution, a 5x5 depth separable dilated convolution with an expansion coefficient of 3 and a 1x1 pointwise convolution to approximately represent a 13x13 large convolution kernel. By the above method, we can capture long-range relationships with slight computational cost and parameters, replace the multi-head self-attention module in the converter, and complete feature extraction;

[0060] Although the decomposition method can enable the network to extract local and global context information of the feature map while reducing the computational cost and the number of parameters, the decomposition method artificially fixes the calculation area of the convolution kernel, which prevents the network from adaptively extracting multi-view feature map reconstruction target features. To solve the above problem, we introduce deformable convolution into the large kernel attention module without significantly increasing the computational cost and the number of parameters, further expand the receptive field, and increase the feature extraction range, so that the network can dynamically and adaptively extract image features during the training process, and improve the reconstruction accuracy of the model. The adaptive large kernel attention module can be written as

[0061] F′ i =GELU(Conv 1×1 (F i ))

[0062] A=Conv 1×1 (DDW-D-Conv(DDW-Conv(F′ i )))

[0063]

[0064] First, we use 1×1 convolution and GELU activation function to transform the input feature F of the i-th view into i Perform channel downsampling operation to obtain F' i ∈R H×W×C’ Then, through the adaptive large core attention module, we get the attention feature map A∈R H ×W×C′ The attention feature map indicates the importance of each feature. Finally, the attention feature map A is combined with F' i After pixel-level multiplication, 1×1 convolution is used to restore the feature channel dimension Output∈R H×W×C .

[0065] The specific composition diagram of the hybrid feedforward neural network module is as follows Figure 5 As shown in the figure, traditional converters usually use fully connected layers as feedforward neural networks, relying entirely on attention to capture inter-pixel dependencies. This results in the network being poor at learning local information of two-dimensional images, affecting the final depth estimation results. To address the above issues, this paper designs a new feedforward neural network to replace the fully connected layers in the self-attention module and the cross-attention module, improving the network's ability to extract significant features across images. Specifically, we first use a 1×1 convolutional layer to expand the input features. Figure X ∈R H×W×C The channel dimension is X'∈R H×W×3C Next, the feature map is evenly divided into three parts along the channel direction. Each part is then subjected to a 1×1 convolution, a 3×3 depthwise convolution, and a 5×5 depthwise convolution before channel concatenation to fully extract multi-scale spatial morphological feature information. The skip connection is then further optimized through three layers of normalization to adaptively retain foreground target features and suppress background clutter interference. Finally, the GELU activation function and 1×1 convolution are used to restore the feature map channel dimension.

[0066] During cost volume regularization, the global receptive field significantly influences the final probability volume prediction. As the receptive field expands, the voxels of the probability volume with the highest probability become smoother, more complete, and more confident in the depth dimension. In practice, we upgrade the 2D adaptive large-kernel attention module to a 3D representation and propose a new cost volume transformer based on 3D adaptive large-kernel attention.

[0067] like Figure 6As shown, the cost volume converter based on three-dimensional adaptive large kernel attention adopts a U-shaped network architecture, which is composed of an encoder, a bottleneck layer, a decoder, and a skip connection; given a cost volume, three three-dimensional feature extraction stages are used in the encoder in succession, each of which contains a three-dimensional self-attention block and a down-sampling layer. The three-dimensional feature map with depth information is input into the three-dimensional self-attention block composed of a three-dimensional adaptive large kernel attention block and a hybrid feedforward neural network block, and the global cost information is regularized and aggregated. Compared with the three-dimensional convolution layer and the non-regularization module, the three-dimensional self-attention block proposed by us can produce a higher-quality probability volume. The down-sampling layer reduces the spatial resolution of the feature map to half of the original while increasing the channel size by one time. The central bottleneck layer uses another three-dimensional self-attention block to perceive the semantic information of the image. The structure of the decoder is symmetrical to that of the encoder. In order to reduce the channel dimension while improving the feature resolution, the up-sampling layer uses transposed convolution to replace the original bilinear interpolation up-sampling method, and adjusts the spatial and channel dimensions of the cost volume; in order to reduce the spatial information loss of the network, we use the skip connection to concatenate the shallow features and the deep features, fuse the multi-scale features in the encoder and the decoder, and retain the multi-scale features from the encoder and the decoder; finally, the channel information of each spatial pixel point is aggregated by using a 1x1x1 convolution kernel to obtain the final probability volume P.

[0068] Step 3, training the network model; input the data set prepared in step 1 into the network model constructed in step 2 for training;

[0069] By minimizing the loss between the output of the network and the label, a better depth prediction result is achieved; we use focal loss as the loss function in the network training process, and regard depth estimation as a classification task to strengthen the single hot spot supervision in the fuzzy area; the focal loss can be expressed as:

[0070]

[0071] wherein represents the prediction probability of depth hypothesis d at pixel p, represents the depth value closest to the ground truth among all hypotheses, {p v represents the subset of pixels with valid ground truth values, and γ represents the focus parameter;

[0072] Set the input image batch size to 2, select the Adam optimizer, set the training iteration number to 16, and gradually decrease the learning rate from 0.0005 to 0. Save the iteration result weight of 8 to 16 rounds;

[0073] Step 4, screening the optimal network model; using the test data set as input, evaluating and screening the optimal network model parameters; the evaluation index selects two official indexes, which are precision evaluation and integrity evaluation; the precision evaluation is the average point cloud to point cloud absolute distance of the reconstruction result to the Ground Truth, and the integrity evaluation is the reconstruction result compared with the Ground Truth. The overall is the average of the accuracy and integrity, indicating the overall performance of the model.

[0074] Step 5, fine-tuning the network model; the network model is fine-tuned using the second multi-view stereo reconstruction data set, so that the model is generalized from the light controllable indoor scene to the outdoor large-scale scene. The second multi-view stereo reconstruction data set uses the BlendedMVS data set; the data set contains 113 outdoor scenes, each scene has 30 to 40 multi-view images, and the scenes include buildings, streets, sculptures and small objects; compared with the DTU data set, the scenes in the BlendedMVS data set contain various camera trajectories; the unstructured camera trajectory can better model different image capture styles, so that the network has better generalization performance for real world reconstruction. We use 106 scenes for training and 7 scenes for testing.

[0075] Step 6, saving the network model; after the network training is completed, the optimal network parameters are saved, and then the multi-view scene image to be three-dimensionally reconstructed is input into the network to obtain the predicted depth map. Through depth filtering and fusion operation, the three-dimensional point cloud after reconstruction is obtained.

[0076] From Figure 7 It can be known that while ensuring that the precision and integrity indexes are not reduced, the method provided by the present application has less memory occupation, less calculation amount and less inference time than the existing method; wherein the memory occupation unit is MB, the calculation amount unit is G, and the inference time unit is second; these indexes further illustrate that the method provided by the present application has better three-dimensional reconstruction quality and lower calculation cost;

[0077] As Figure 8 shown, the present application also provides an outdoor scene-oriented multi-view stereo reconstruction system, mainly including an image acquisition module, a model building module, a data set and model loading module, a model training module, a depth estimation module, a depth filtering and fusion module and a point cloud visualization module.

[0078] The image acquisition module is used for acquiring training data and test data; the training data and test data include multi-view images and corresponding depth images.

[0079] The model building module is used for building the network model provided by the present application and initializing the model weight.

[0080] A dataset and model loading module, used to load input images and network models onto hardware devices for model training; the hardware devices refer to image processors;

[0081] The model training module is used to perform supervised model training on the network model constructed by this method for the multi-view images of each scene in the training data to obtain the optimal model weights;

[0082] The depth estimation module is used to estimate the depth of the multi-view images of each scene in the test data through a network model equipped with the optimal model weights to obtain a multi-view depth map of each scene;

[0083] The depth filtering and fusion module is used to obtain a 3D point cloud file for each scene after performing depth filtering and fusion operations on the multi-view depth map predicted by the network;

[0084] The point cloud visualization module is used to analyze and process the 3D point cloud file through the MeshLab software platform to realize 3D point cloud visualization; the visualization results are as follows Figure 9 As shown;

[0085] like Figure 10 As shown, the present invention also provides an electronic device for multi-view stereo reconstruction of outdoor scenes, which mainly includes a memory, a processor, a communication interface and a bus; wherein the memory, the processor and the communication interface are connected to each other through the bus;

[0086] A Samsung 1TB solid-state drive is used as memory to store data sets, programs, and intermediate products during program execution. When the program stored in the memory is executed by a processor, the processor and the communication interface are used to perform each step of the multi-view stereo reconstruction training method according to an embodiment of the present invention.

[0087] Using an Intel Core i7-14700KF as a processor to execute relevant programs to implement the functions required to be performed by the units in the multi-view stereo reconstruction training system of the present invention, or to perform the multi-view stereo reconstruction training method of the present invention;

[0088] The communication interface uses a transceiver system such as, but not limited to, a transceiver to achieve communication between the system and other devices or communication networks; for example, the image to be processed or the initial feature map of the image to be processed can be obtained through the communication interface;

[0089] The bus may include a pathway for transferring information between various components of the system (e.g., memory, processor, communication interface);

[0090] The embodiments of the application have been described above. However, these embodiments are merely meant to be illustrative, and not meant to limit the scope of the application. Although each of the embodiments has been described above separately, this does not mean that the measures in the individual embodiments cannot be used advantageously in combination. The scope of the application is defined by the claims appended hereto and their equivalents. Various alternatives and modifications can be made to the embodiments of the application without departing from the scope of the application, and it is intended that all such alternatives and modifications be included within the scope of the disclosure.

Claims

1. An efficient multi-view stereo reconstruction method for outdoor scenes, characterized by: The steps include: Step 1: Prepare a data set: prepare a first multi-view stereo reconstruction data set and a second multi-view stereo reconstruction data set, and divide the two data sets into a training data set and a test data set respectively; The training data set is used to train the network model, and the test data set is used to evaluate the model training effect; Step 2: Build the network model: The network consists of a backbone network, a large-core attention-based adaptive feature extraction converter module, a cost regularization module, and a depth estimation module. The backbone network in step 2 is a feature pyramid feature extraction network, which is used to extract multi-scale features of the input reference image and the source image; The large-core attention-based adaptive feature extraction converter module includes an adaptive large-core attention block, a hybrid feedforward neural network block, a linear cross attention block and an upsampling module; The cost regularization module replaces the 3D convolution kernel with a 3D adaptive large kernel attention, thereby expanding the receptive field without increasing model parameters and computational complexity, enabling the model to better capture global information in the scene, thereby improving the accuracy of stereo reconstruction; The depth estimation module uses maximum probability depth estimation to select the depth value with the highest probability value from the probability volume for each pixel position and use it as the depth value of the pixel position in the depth estimation map; The adaptive large-core attention block in step 2 uses multi-scale convolution kernels with different expansion coefficients to form a large-size convolution kernel, which effectively reduces the computational cost and improves the computational efficiency of the model without sacrificing performance. On this basis, by introducing deformable convolution, the large-size convolution kernel can adaptively learn multi-view targets of different scales, obtain a larger receptive field, and enable the model to better understand the target context information in outdoor scenes, thereby improving the ability to extract complex features; The hybrid feedforward neural network block better aggregates global and multi-scale local information and extends the design of layer normalization to a general form, promoting the optimization of skip connections and improving the network's feature extraction capabilities; The linear cross attention block enables the network to learn feature correlations between images from different perspectives, helping the model better understand the spatial structure of objects and improving the accuracy and stability of stereo reconstruction. Step 3: Train the network model: Input the dataset prepared in step 1 into the network model built in step 2 for training; by minimizing the loss between the network output and the label, better depth prediction results can be achieved; Step 4: Screen the optimal network model; use the test dataset as input to evaluate and screen the optimal network model parameters; Step 5: Fine-tune the network model using the second multi-view stereo reconstruction dataset to generalize the model from indoor scenes with controllable lighting to large-scale outdoor scenes. Step 6: Save the network model. After the network training is completed, save the optimal network parameters. Then, input the multi-view scene image to be 3D reconstructed into the network to obtain the predicted depth map. Through depth filtering and fusion operations, the reconstructed 3D point cloud is obtained.

2. The efficient multi-view stereo reconstruction method for outdoor scenes according to claim 1, characterized in that: The first multi-view stereo reconstruction dataset in step 1 is a DTU dataset, which obtains multi-view images, depth maps and corresponding camera internal and external parameters of each scene.

3. The efficient multi-view stereo reconstruction method for outdoor scenes according to claim 1, characterized in that: In step 3, the preset thresholds in the training network model include a loss function preset value, an iteration number preset value, and an optimized learning rate preset value.

4. The efficient multi-view stereo reconstruction method for outdoor scenes according to claim 1, characterized in that: The loss function in step 3 is focal loss to strengthen single hotspot supervision in the fuzzy area.

5. The efficient multi-view stereo reconstruction method for outdoor scenes according to claim 1, characterized in that: The step 4 also includes evaluating the accuracy and completeness of the algorithm reconstruction results through evaluation indicators during the training of the network model.

6. The efficient multi-view stereo reconstruction method for outdoor scenes according to claim 1, characterized in that: The second multi-view stereo reconstruction dataset in step 5 is a BlendedMVS dataset.

7. A multi-view stereo reconstruction system for outdoor scenes, characterized by: The system comprises: An image acquisition module is used to acquire training data and test data; the training data and test data include multi-view images and corresponding depth images; Model building module, used to build the proposed network model and initialize the model weights; A dataset and model loading module, used to load input images and network models onto hardware devices for model training; the hardware devices refer to image processors; The model training module is used to perform supervised model training on the multi-view images of each scene in the training data through the constructed network model to obtain the optimal model weight; The depth estimation module is used to estimate the depth of the multi-view images of each scene in the test data through a network model equipped with the optimal model weights to obtain a multi-view depth map of each scene; The depth filtering and fusion module is used to obtain a 3D point cloud file for each scene after performing depth filtering and fusion operations on the multi-view depth map predicted by the network; The point cloud visualization module is used to parse and process 3D point cloud files through the MeshLab software platform to achieve 3D point cloud visualization; A storage medium for storing a multi-view stereo reconstruction system; The method is applied to the efficient multi-view stereo reconstruction method for outdoor scenes as described in any one of claims 1 to 6.

Images