A fast binocular stereo matching method

Abstract

A fast binocular stereo matching method includes: S1. Inputting the left and right views into a ResNet-based feature extraction network to extract left and right view features, and downsampling to obtain context feature pyramids of different resolutions; S2. Using the left and right view features to calculate correlation features, downsampling the candidate matching dimensions of the correlation quantity to obtain a correlation quantity pyramid; S3. Based on the disparity estimate and a preset offset value, searching for feature vectors within the correlation region of each pixel at each level of the correlation pyramid; concatenating the vectors found at each level along the feature channel dimension; aggregating the concatenated features with the current disparity estimate through convolution, and then concatenating them along the feature channel dimension to form correlation features; S4. Inputting the correlation features and the context feature pyramid into a multi-scale efficient GRU module for disparity optimization and update, iteratively obtaining the final estimated disparity. This improves the restoration of fine structures and the accuracy of disparity estimation.

Description

Technical Field

[0001] The present invention relates to computer vision technology, and in particular to a fast binocular stereo matching method. Background Art

[0002] Stereo matching is a computer vision technique that aims to reconstruct the depth information of a three-dimensional scene by computing a dense disparity map by analyzing paired binocular images. As the most widely used passive sensing technology for depth estimation, stereo matching can obtain dense estimated disparity maps at a low cost and overcome the environmental limitations of active sensing technologies that rely on hardware. Therefore, it has high application value and research prospects in industrial scenarios.

[0003] Due to the stringent requirements for real-time algorithm performance and low deployment costs in industrial applications, current stereo matching algorithms still face significant challenges in practical deployment. Stereo matching techniques based on traditional methods offer low computational cost and high speed, but suffer from low disparity estimation accuracy, large estimation errors in ill-posed areas such as weak textures and occlusions, and poor generalization. Advanced deep learning-based stereo matching techniques, on the other hand, mostly rely on 3D convolution for cost aggregation, which is computationally intensive and requires the construction of a 4D cost volume. This results in large amounts of memory required at runtime and prevents direct inference of high-resolution images. Consequently, in industrial applications, greater attention is being paid to lightweight deep learning binocular stereo matching algorithms. To speed up the algorithm's inference and reduce the computational effort, it is often necessary to limit the resolution of the operation.

[0004] It should be noted that the information disclosed in the above background technology section is only used to understand the background of this application, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] The main purpose of the present invention is to overcome the defects in the above-mentioned background technology and provide a fast binocular stereo matching method.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A fast binocular stereo matching method comprises the following steps:

[0008] S1. Input the left and right views into the ResNet-based feature extraction network for feature extraction. The extracted left and right view features are downsampled through the residual block to obtain context feature pyramids of different resolutions.

[0009] S2. Using the correlation feature acquisition method, the correlation features are calculated using the left and right view features, and the candidate matching dimensions of the correlation quantity are downsampled to obtain different receptive fields and obtain a correlation quantity pyramid;

[0010] S3. Based on the current disparity estimate and the preset offset value, search for feature vectors within the correlation region of each pixel at each level of the correlation pyramid; concatenate the vectors found at each level along the feature channel dimension; aggregate the concatenated features with the current disparity estimate through convolution, and then concatenate along the feature channel dimension to form a correlation feature;

[0011] S4. Input the correlation features and context feature pyramid into the multi-scale efficient gated recurrent unit (GRU) module for disparity optimization and update. After multiple iterations, the final estimated disparity is obtained. The multi-scale efficient GRU module updates the hidden state and estimates the disparity residual through the deep separable HE-DeepSepConvGRU unit enhanced by high-frequency information at multiple input resolutions.

[0012] Further:

[0013] In step S4, a hierarchical resolution update strategy is adopted. Each complete multi-scale efficient GRU module update includes: first performing multiple hidden state updates in the lower resolution HE-DeepSepConvGRU unit, followed by a hidden state update and estimated disparity update in the highest resolution HE-DeepSepConvGRU unit; different resolution units interact with each other at multiple levels by taking each other's hidden states as input.

[0014] In step S1, feature extraction is performed through the ResNet network with shared weights to obtain Input the left and right view features of the resolution, and downsample the left and right view features through three residual blocks to obtain resolutions of Contextual feature pyramid at input resolution.

[0015] In step S1, the calculation of correlation features using left and right view features includes: The left and right view features of the input resolution are similarly calculated by dot product to obtain the correlation.

[0016] The regression loss L1 distance commonly used in binocular stereo matching algorithms is used as the network loss function. The L1 distance between the iterative output disparity result of the disparity optimization network and the true value is calculated and weighted and summed. The weight increases exponentially according to the iteration order.

[0017] The HE-DeepSepConvGRU unit uses separable convolution to replace the original convolution operation, and replaces the original convolution layer in ConvGRU with a pointwise convolution layer (Pointwise Convolution) with a convolution kernel of 1×1 and a channel-by-channel convolution layer (Depthwise Convolution) with a convolution kernel of 3×3.

[0018] The HE-DeepSepConvGRU unit includes the interaction between high-frequency input information HF-x and the hidden state h.

[0019] In step S4, the hidden state at the current resolution is used as h t-1 ; The features at the current resolution in the input context feature pyramid are subjected to separable convolution to obtain r t , z t ,q t ; Use downsampled high-resolution hidden state and correlation features as high-frequency input features The upsampled low-resolution hidden state is used as the remaining input feature x t ;

[0020] First, the high-frequency input features With the current hidden state h t-1 By performing the interaction, the interaction formula is as follows:

[0021]

[0022]

[0023] Where δ is the hyperparameter of the balanced distribution. The enhanced high-frequency input features are obtained With the rest of the input features x t Connect along the channel dimension to get the complete input features With the enhanced hidden state and r t , z t ,q t Enter GRU for update.

[0024] In step S4, for the deep separable ConvGRU unit with high-frequency information enhancement at the highest resolution, the updated hidden state h t The disparity residual is estimated through two layers of convolution, and the estimated residual value is added to the current estimated disparity to obtain the updated disparity estimate.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the fast binocular stereo matching method.

[0026] The present invention has the following beneficial effects:

[0027] This paper proposes an efficient binocular stereo vision method with high-frequency information enhancement, which improves the lightweight stereo matching method's ability to restore fine structures and increases the accuracy of disparity estimation. The advantages of this method over traditional methods are:

[0028] We propose a deep separable ConvGRU unit with enhanced high-frequency information (HE-DeepSepConvGRU). Compared to the original ConvGRU, this unit enhances the interaction between input high-frequency features and hidden state features, enhancing the role of high-frequency information in disparity optimization and leveraging local detail information to better restore fine structures. Furthermore, separable convolution replaces the original convolution operation, significantly reducing the model's computational complexity and making it easier to deploy in real-world industrial scenarios.

[0029] Using a hierarchical resolution update strategy, each complete multi-scale efficient GRU module update consists of multiple hidden state updates at lower resolutions, followed by a hidden state update and estimated disparity update at higher resolutions. Units at different resolutions interact with each other at multiple levels, using their hidden states as input.

[0030] Compared with the existing lightweight binocular stereo matching network, the network of the present invention makes full use of high-frequency information, can better deal with detailed texture areas, further reduce the network volume, accelerate the network inference speed, and meet the strict requirements of industrial applications for algorithm real-time performance and low deployment cost.

[0031] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The overall framework of a stereo matching network based on an iterative optimization architecture with good generalization performance according to an embodiment of the present invention is shown.

[0033] Figure 2 The specific update strategy of the multi-scale efficient GRU module of an embodiment of the present invention is shown.

[0034] Figure 3 The detailed structure of the depth-wise separable ConvGRU unit with high-frequency information enhancement according to an embodiment of the present invention is shown.

[0035] Figure 4 Flowchart of a fast binocular stereo matching method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0037] Abbreviations and key terms:

[0038] High-frequency information refers to the details in an image, mainly including edges, textures, noise, and local areas of rapid change. These features often reflect the complexity of the internal structure of the image and therefore play an important role in image processing and computer vision tasks.

[0039] Parallax: The horizontal displacement of corresponding surface points on an object in a stereoscopic binocular image on the imaging plane. This displacement is caused by the different positions of the object in the left and right camera perspectives. Parallax can be used to infer depth information using the principle of similar triangles, given known camera extrinsics.

[0040] A feature pyramid is a multi-scale representation used in computer vision and image processing to represent image features at different scales and resolutions. It is constructed by scaling the original image to a series of progressively smaller sizes and extracting features at each size. This captures details and structural information at different levels of the image, helping algorithms better handle scale variations and varying resolutions.

[0041] Most current lightweight binocular stereo matching algorithms perform calculations at a lower-resolution feature scale, making insufficient use of high-frequency information. High-frequency information contains local detail texture content, which greatly affects the estimation of the scene's fine structure. This results in limitations in the accuracy of disparity estimation by lightweight stereo matching algorithms. The efficient binocular stereo vision algorithm with enhanced high-frequency information proposed in this paper improves the lightweight stereo matching algorithm's ability to restore fine structures and improves the accuracy of disparity estimation.

[0042] See Figure 1 , an embodiment of the present invention provides a fast binocular stereo matching method, comprising the following steps:

[0043] S1. Input the left and right views into the ResNet-based feature extraction network for feature extraction. The extracted left and right view features are downsampled through the residual block to obtain context feature pyramids of different resolutions.

[0044] S2. Using the correlation feature acquisition method, the correlation features are calculated using the left and right view features, and the candidate matching dimensions of the correlation quantity are downsampled to obtain different receptive fields and obtain a correlation quantity pyramid;

[0045] S3. Based on the current disparity estimate and the preset offset value, search for feature vectors within the correlation region of each pixel at each level of the correlation pyramid; concatenate the vectors found at each level along the feature channel dimension; aggregate the concatenated features with the current disparity estimate through convolution, and then concatenate along the feature channel dimension to form a correlation feature;

[0046] S4. Input the correlation features and context feature pyramid into the multi-scale efficient gated recurrent unit (GRU) module for disparity optimization and update. After multiple iterations, the final estimated disparity is obtained. The multi-scale efficient GRU module updates the hidden state and estimates the disparity residual through the deep separable HE-DeepSepConvGRU unit enhanced by high-frequency information at multiple input resolutions.

[0047] In a preferred embodiment, a hierarchical resolution update strategy is adopted, and each complete multi-scale efficient GRU module update includes: first performing multiple hidden state updates in the lower resolution HE-DeepSepConvGRU unit, followed by hidden state update and estimated disparity update in the highest resolution HE-DeepSepConvGRU unit; different resolution units interact with each other at multiple levels by taking each other's hidden states as input.

[0048] Specific embodiments of the present invention are further described below.

[0049] Stereo Matching Network Architecture

[0050] A stereo matching network based on iterative optimization architecture with good generalization performance is used. Its overall framework is as follows Figure 1 shown.

[0051] The overall network includes a feature extraction network based on ResNet and an iterative disparity optimization network based on multi-scale efficient GRUs.

[0052] The left and right views are used as input and feature extraction is performed through the ResNet network with shared weights to obtain Input the left and right view features of the resolution, and downsample the left and right view features through three residual blocks to obtain resolutions of Contextual feature pyramid at input resolution.

[0053] Then, the correlation feature acquisition method of the RAFT-Stereo network is used to calculate the correlation feature using the left and right view features. Specifically: The left and right view features of the input resolution are similarly calculated through dot products to obtain a correlation. Then, multiple pooling layers are used to downsample the candidate matching dimensions of the correlation to obtain different receptive fields, thus obtaining a correlation pyramid.

[0054] Using the search operator defined in RAFT-Stereo, the algorithm searches for feature vectors within the correlation region of each pixel at each level of the correlation pyramid, which contains a large amount of data, based on the current disparity estimate (initialized to an all-zero matrix in the first iteration) and a preset offset value. The vectors found at each level are then concatenated along the feature channel dimension. The concatenated features are aggregated with the current disparity estimate through two layers of convolution and then concatenated along the feature channel dimension to form the correlation feature.

[0055] Afterwards, the correlation features and context feature pyramid are input into the proposed multi-scale efficient GRU module for disparity optimization and update. After multiple iterations, the final estimated disparity is obtained.

[0056] The regression loss L1 distance commonly used in binocular stereo matching algorithm is used as the loss function of the network, and the disparity results {d1,…,d N}Calculation and truth value d th The L1 distance between them is weighted and summed. The weight increases exponentially in the order of iteration. The formula is as follows:

[0057]

[0058] Where γ is the weight hyperparameter, which is set to 0.9.

[0059] Multi-scale efficient GRU module

[0060] Feature operations at high resolution are beneficial for preserving more detailed features to restore fine structures. However, in binocular stereo matching, when there are large areas of weak texture, local information recognition is low, and a larger receptive field is needed to estimate such areas. Features at low resolution have a more global receptive field, so a multi-scale update strategy is adopted. The multi-scale efficient GRU module is composed of The deep separable ConvGRU unit (HE-DeepSepConvGRU) with high-frequency information enhancement at three input resolutions updates the hidden state and estimates the disparity residual. Units of different resolutions are cross-connected by using each other's hidden states as inputs, where the HE-DeepSepConvGRU unit at lower resolutions is only used to update the hidden state, and the HE-DeepSepConvGRU at the highest resolution not only updates the hidden state but also updates the disparity estimation based on the hidden state and correlation features. In order to speed up the computational speed of the network, a hierarchical resolution update strategy is adopted: each complete multi-scale efficient GRU module update includes multiple updates on the low-resolution HE-DeepSepConvGRU unit, and then an update on the highest resolution HE-DeepSepConvGRU unit. The specific update strategy of the multi-scale efficient GRU module is as follows. Figure 2 shown.

[0061] A complete update of the multi-scale efficient GRU module is as follows: The HE-DeepSepConvGRU unit at the resolution simultaneously inputs the hidden state of the corresponding resolution (when the first update is made, the hidden state is initialized to the feature of the corresponding resolution scale in the context feature pyramid), the hidden state of the context feature pyramid Features at different resolutions and downsampled Resolution hidden state, updated by HE-DeepSepConvGRU unit Resolution hidden state. Repeat afterwards The HE-DeepSepConvGRU unit update operation at the resolution upsamples the updated hidden features and High-resolution hidden state, context feature pyramid Features at different resolutions and downsampling Resolution hidden state input together HE-DeepSepConvGRU unit at resolution Resolution hidden state update. Next, repeat it separately and The HE-DeepSepConvGRU unit update operation at resolution is updated resolution hidden state upsampling, with High-resolution hidden state, context feature pyramid The features at the resolution and the correlation features are input together HE-DeepSepConvGRU unit at resolution Resolution hidden state updates and parallax updates.

[0062] Deep Separable ConvGRU Unit with High-Frequency Information Enhancement (HE-DeepSepConvGRU)

[0063] To enhance the network's utilization of high-frequency information, the Mogrifier LSTM draws inspiration from the LSTM. Mogrifier LSTM enhances context modeling by adding early interaction between input information x and hidden state h. By adding interaction between high-frequency input information HF-x and hidden state h to the original ConvGRU unit, the Mogrifier LSTM enhances the utilization of high-frequency information and learns more detailed features.

[0064] At the same time, in order to enable the network to better cope with the real-time and lightweight requirements in practical applications, with reference to separable convolution, the original convolution layer in the original ConvGRU is replaced with a pointwise convolution layer (PointwiseConvolution) with a convolution kernel of 1×1 and a channel-by-channel convolution layer (DepthwiseConvolution) with a convolution kernel of 3×3, so that the network has fewer parameters and faster speed while maintaining performance.

[0065] The detailed structure of the deep separable ConvGRU unit with high-frequency information enhancement is as follows Figure 3 As shown. The hidden state at the current resolution is used as h t-1 ; The features at the current resolution in the input context feature pyramid are subjected to separable convolution to obtain r t , z t ,q t ; Use downsampled high-resolution hidden state and correlation features as high-frequency input features The upsampled low-resolution hidden state is used as the remaining input feature x t .

[0066] First, the high-frequency input features With the current hidden state h t-1 By performing the interaction, the interaction formula is as follows:

[0067]

[0068]

[0069] Where δ is the hyperparameter of the balanced distribution and is set to 2. The enhanced high-frequency input features are obtained With the rest of the input features x t Connect along the channel dimension to get the complete input features With the enhanced hidden state and rt , z t ,q t Enter GRU for update. The update formula is:

[0070]

[0071]

[0072]

[0073]

[0074] For the deep separable ConvGRU unit with high-frequency information enhancement at the highest resolution, the updated hidden state h t The disparity residual is estimated through two layers of convolution, and the estimated residual value is added to the current estimated disparity to obtain the updated disparity estimate.

[0075] The advantages of the embodiments of the present invention over traditional methods are as follows:

[0076] We propose a deep separable ConvGRU unit with enhanced high-frequency information (HE-DeepSepConvGRU). Compared to the original ConvGRU, this unit enhances the interaction between input high-frequency features and hidden state features, enhancing the role of high-frequency information in disparity optimization and leveraging local detail information to better restore fine structures. Furthermore, separable convolution replaces the original convolution operation, significantly reducing the model's computational complexity and making it easier to deploy in real-world industrial scenarios.

[0077] Using a hierarchical resolution update strategy, each complete multi-scale efficient GRU module update consists of multiple hidden state updates at lower resolutions, followed by a hidden state update and estimated disparity update at higher resolutions. Units at different resolutions interact with each other at multiple levels, using their hidden states as input.

[0078] Compared with the existing lightweight binocular stereo matching network, the network of the present invention makes full use of high-frequency information, can better deal with detailed texture areas, further reduce the network volume, accelerate the network inference speed, and meet the strict requirements of industrial applications for algorithm real-time performance and low deployment cost.

[0079] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0080] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0081] An embodiment of the present invention further provides a processor, which executes a computer program and at least performs the method described above.

[0082] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a magnetic disk memory or a magnetic tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0083] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0084] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0085] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0086] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0087] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0088] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0089] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0090] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0091] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. Those skilled in the art will recognize that, without departing from the scope of the present invention, several equivalent substitutions or obvious variations can be made, and the performance or use of the same should be considered to fall within the scope of protection of the present invention.

Claims

1. A fast binocular stereo matching method, characterized in that: The steps include: S1. Input the left and right views into the ResNet-based feature extraction network for feature extraction. The extracted left and right view features are downsampled through the residual block to obtain context feature pyramids of different resolutions. S2. Using the correlation feature acquisition method, the correlation features are calculated using the left and right view features, and the candidate matching dimensions of the correlation quantity are downsampled to obtain different receptive fields and obtain a correlation quantity pyramid; S3. Based on the current disparity estimate and the preset offset value, search for feature vectors within the correlation region of each pixel at each level of the correlation quantity pyramid; concatenate the vectors found at each level along the feature channel dimension; aggregate the concatenated features with the current disparity estimate through convolution, and then concatenate along the feature channel dimension to form a correlation feature; S4, the correlation features and context feature pyramid are input into the multi-scale efficient gated recurrent unit GRU module for disparity optimization and update, and the final estimated disparity is obtained after multiple iterations; wherein, the multi-scale efficient GRU module updates the hidden state and estimates the disparity residual through the deep separable HE-DeepSepConvGRU unit enhanced by high-frequency information under multiple input resolutions; the HE-DeepSepConvGRU unit uses separable convolution to replace the original convolution operation, and replaces the original convolution layer in ConvGRU with the convolution kernel The point-by-point convolution layer and convolution kernel are The channel-by-channel convolutional layer is replaced by a hierarchical resolution update strategy. Each complete multi-scale efficient GRU module update includes: multiple hidden state updates in the lower resolution HE-DeepSepConvGRU unit, followed by hidden state updates and estimated disparity updates in the highest resolution HE-DeepSepConvGRU unit; different resolution units interact with each other at multiple levels by taking each other's hidden states as input.

2. The fast binocular stereo matching method according to claim 1, wherein: In step S1, feature extraction is performed through the ResNet network with shared weights to obtain Input the left and right view features of the resolution, and downsample the left and right view features through three residual blocks to obtain resolutions of Contextual feature pyramid at input resolution.

3. The fast binocular stereo matching method according to claim 2, wherein: In step S1, the calculation of correlation features using left and right view features includes: The left and right view features of the input resolution are similarly calculated by dot product to obtain the correlation.

4. The fast binocular stereo matching method according to any one of claims 1 to 3, wherein: The regression loss L1 distance commonly used in binocular stereo matching algorithms is used as the network loss function. The L1 distance between the iterative output disparity result of the disparity optimization network and the true value is calculated and weighted and summed. The weight increases exponentially according to the iteration order.

5. The fast binocular stereo matching method according to any one of claims 1 to 3, wherein: The HE-DeepSepConvGRU unit includes the interaction between high-frequency input information HF-x and the hidden state h.

6. The fast binocular stereo matching method according to any one of claims 1 to 3, wherein: In step S4, the hidden state at the current resolution is used as ; The features at the current resolution in the input context feature pyramid are obtained by separable convolution ; Use downsampled high-resolution hidden state and correlation features as high-frequency input features , the upsampled low-resolution hidden state is used as the remaining input features ; First, the high-frequency input features With the current hidden state By performing the interaction, the interaction formula is as follows: ； ； in is the hyperparameter of the balanced distribution, and the enhanced high-frequency input features are obtained , and the rest of the input features Connect along the channel dimension to get the complete input features , and the enhanced hidden state and Enter GRU for update.

7. The fast binocular stereo matching method according to claim 6, wherein: In step S4, for the deep separable ConvGRU unit with high-frequency information enhancement at the highest resolution, the updated hidden state The disparity residual is estimated through two layers of convolution, and the estimated residual value is added to the current estimated disparity to obtain the updated disparity estimate.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the fast binocular stereo matching method according to any one of claims 1 to 7 is implemented.

Images