A method and system for depth estimation of pond monocular images

Abstract

The application discloses to belong to the computer vision technical field, especially relate to a kind of for pond monocular image depth estimation method and system, the method includes: constructing pond monocular image depth estimation model, pond image is input pond monocular image depth estimation model, obtain the depth mapping of pond scene, obtain the distance information of each pixel point to camera, generate pond image depth map based on the distance information, realize the depth estimation of pond image.The system includes pond monocular image depth estimation model includes:U-MonoVIT encoder, local plane guide layer, jump multi-scale extended self-attention mechanism.Can effectively solve the problem that pond image is blurred, contrast is low, overexposure or dim, and the depth estimation accuracy of the problem caused by low contrast, overexposure or dim is reduced, can accurately and efficiently realize the monocular depth estimation of pond scene, provides key support for comprehensive exploration of pond environment and biomass estimation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a depth estimation method and system for a pond monocular image. BACKGROUND

[0002] The main purpose of monocular depth estimation is to infer the depth information of each object or scene element in the scene from a single image, that is, the distance from each pixel to the camera, which has wide application in the fields of automatic driving, three-dimensional reconstruction and virtual reality. Depth estimation of images in ponds is a key basic technology for accurately perceiving and understanding the scene in underwater environments, which helps to promote the development and application of pond biomass estimation, aquaculture environment monitoring, pond biological behavior analysis and other tasks, and provides a cost-effective and superior perception solution for intelligent aquaculture and pond monitoring tasks.

[0003] Most of the existing research on monocular image depth estimation is for land scene research, which has been widely applied in the fields of automatic driving, virtual reality, robot navigation and has achieved remarkable results. The supervised monocular depth learning method is to regress a single pixel using a dataset with real depth data. For each pixel point in the input RGB image, the model needs to predict the depth value at the corresponding pixel position of the image. Its advantage is relatively high precision, which can preserve the key details of the depth map. The self-supervised method constructs labels based on the structure and characteristics of the data itself through geometric relationships during training, and then the model can be trained like supervised learning. Since the self-supervised monocular depth estimation method does not require real depth data, it cannot obtain real depth values, so there is still some accuracy gap compared with the supervised method.

[0004] Monocular image depth estimation in the pond scene shows different characteristics from monocular image depth estimation in the land scene in many ways. First, in the pond environment, due to the scattering of light and the influence of suspended particles in water, the image usually presents the characteristics of blurring and low contrast. These factors will affect the edge definition and detail information of objects in the pond image, making it difficult for the depth estimation algorithm to correctly identify the boundaries and shapes of objects. Second, the environment in the pond has complex light changes, which leads to overexposure or dimness in the pond image, thereby affecting the judgment of the distance and position of objects in the image by the depth estimation algorithm. Therefore, there is an urgent need for a depth estimation method for pond monocular images, which can accurately obtain underwater depth images and provide support for intelligent aquaculture research and provide a basis for pond environment monitoring. SUMMARY

[0005] The application aims to provide a depth estimation method for a pond monocular image, construct a pond monocular image depth estimation model, input a pond image into the pond monocular image depth estimation model, obtain a depth map of the pond scene, and generate a pond image depth map based on distance information, so as to realize depth estimation of the pond image.

[0006] The construction of the pond monocular image depth estimation model comprises the following steps:

[0007] In the encoding stage, given the current input image, the U-MonoVIT encoder serving as the backbone network extracts multi-scale image features; the U-MonoVIT encoder comprises a local perception unit and a parallel feature extraction module.

[0008] In the decoding stage, the local plane guide layer is adopted to obtain the local feature information around the pixel point by fitting the local image block through 4D plane coefficients.

[0009] Through the skip multi-scale expansion self-attention mechanism, different expansion rates are set for different heads, so as to effectively aggregate the semantic information of different scales in the receptive field of the encoder and the decoder.

[0010] The U-MonoVIT encoder further comprises:

[0011] In the encoding stage, the given current input image is sequentially input into a convolution layer and four Transformer blocks; the convolution layer sequentially comprises a first convolution, a second convolution and a third convolution; the convolution kernel size of the three convolutions is 3*3, the step length of the first convolution and the second convolution is 1, and the step length of the third convolution is 2; the four Transformer blocks correspond to stages 1, 2, 3 and 4, four stages correspond to four dimensions [C1, C2, C3, C4], which respectively correspond to 1 / 4, 1 / 8, 1 / 16 and 1 / 32 proportion features, and the depths of each stage are [3, 4, 6, 3]; the Transformer block comprises a patch embedding and a VIT layer; the patch embedding converts a picture into a sequence; the VIT layer comprises a parallel first branch and a second branch; the first branch comprises a local perception unit, a first normalization layer and a multi-head self-attention module; the second branch comprises a parallel feature extraction module; the outputs of the first branch and the second branch are added and input into a second normalization layer and a multi-layer perception mechanism.

[0012] The feature extraction link, the batch normalization layer BN and the convolution layer Conv; the feature extraction link comprises a convolution layer Conv, a batch normalization layer BN and a SILU activation function;

[0013] The working process of the parallel feature extraction module is that the feature map processed by the local perception unit is input into the parallel feature extraction module, first passes through a 3x3 convolution layer Conv, the convolution kernel size of the 3x3 convolution layer can effectively capture local details and global information, and enhance the feature expression ability; then, the feature map passes through a batch normalization layer BN, the batch normalization layer BN can accelerate network convergence, improve training stability, and reduce internal covariate shift; at the same time, a SILU (Sigmoid Linear Unit) activation function is used to further improve the feature extraction ability of the model for underwater blurred areas and enhance the perception effect of low contrast and unevenly illuminated areas; after the above process is executed twice, a 3x3 convolution layer Conv and a normalization layer BN are executed respectively;

[0014] The local perception unit comprises a 3X3 deep convolution;

[0015] The working process of the local perception unit is that first, the input image is divided into small blocks (patches) and converted into a sequence representation. Then, the patches pass through a 3x3 deep convolution layer to extract local features while preserving spatial information. In order to enhance the feature expression ability, the feature map processed by the deep convolution is added to the feature map without deep convolution, realizing feature complementation. The fused feature map is further processed by a convolution layer block to capture local details and global dependencies at the same time.

[0016] The U-MonoVIT encoder extracts multi-scale image features, which includes:

[0017] The definition phase i is:

[0018]

[0019] In the formula, DWConv(x) represents a deep convolution.

[0020] The local context is embedded in the visual marker using a parallel feature extraction module, and the fusion mode of the parallel feature extraction module is represented as:

[0021]

[0022] In the formula, is the feature map processed by the parallel feature extraction module of phase i, PFEM(x) is the parallel feature extraction module, is a multi-head attention mechanism;

[0023] The local relationship between the features is constructed based on the local perception unit and the parallel feature extraction module;

[0024] The multi-head self-attention module will Transformed by linear transformation as query Key and value In the formula, n = H x W is the number of visual markers:

[0025] MHA(Q, K, V) = Concat(head0,..., head N )W O (3)

[0026] head i = Attention(QW i Q , KW i K , VW i V ) (4)

[0027]

[0028] In the formula, MHA(Q, K, V) represents the output of multi-head attention, head N represents the output of the Nth attention head; W O represents a projection matrix used for concatenating the projected dimensions; Q represents the query, K T represents the key, d k represents the scaling factor, and V represents the value matrix; W i Q represents the Q learning weight matrix, W i K represents the K learning weight matrix, and W i V represents the V learning weight matrix; Softmax represents the normalized weight.

[0029] The fusion features are fed into a feedforward network, represented as:

[0030] Y = MHSA(Norm(X)) + X (6)

[0031] Z = MLP(Norm(Y)) + Y (7)

[0032] In the formula, MLP represents a multi-layer perceptron, Norm represents Layer Normalization on Y, which ensures numerical stability, X represents the input of the encoder layer, Y represents the output after multi-head self-attention calculation, and Z represents the final output feature.

[0033] The local plane guide layer obtains the local feature information around the pixel point by fitting the local image block through 4D plane coefficients, which specifically includes:

[0034] Given a feature map with spatial resolution H / k, the proposed layer estimates a 4D plane coefficient for each spatial cell; to guide the feature with local plane hypothesis, use ray-plane intersection to convert each estimated four-dimensional plane coefficient into k x k local depth cues c i :

[0035]

[0036] where n1n2n3n4 represents the estimated plane coefficient, u i , v i k x k block normalized coordinates of pixel i;

[0037] By stacking 1 x 1 convolution, repeatedly reduce the number of channels by 2 until the number of channels reaches 3; pass the feature map through the first method and the second method to get the local plane coefficient estimation; the first method is to convert into a unit normal vector, and the second method is to define a sigmoid function of the vertical distance between the plane and the origin; after the sigmoid function, multiply the output by the maximum distance k to get the actual depth value;

[0038] The first two channels of the given feature map are regarded as angles, and they are converted into unit normal vectors using the following formula:

[0039]

[0040] The unit normal vector includes two degrees of freedom: polar angle θ and azimuth angle

[0041] The up-sampling results of the feature map at the H / 8 and H / 4 stages are spliced together after passing through a 3*3 convolution and a bn layer, and the down-sampling results of the local plane guided layer restored to full resolution, to obtain local features Given global features and local features Perform 3 x 3 convolution on and to input the jump multi-scale expansion self-attention mechanism; by normalizing the global features and local features , the distribution of the input is more stable:

[0042]

[0043] In the formula, LayerNorm is layer normalization, is the normalized feature of the global feature , is the normalized feature of the local feature .

[0044] The self-attention mechanism that expands through multiple scales by skipping, and sets different expansion rates for different heads, specifically includes:

[0045] The upsampling results from the feature mapping downsampling to H / 8 and H / 4 stages and the downsampling results from the local plane guided layer restored to full resolution are concatenated together after passing through 3*3 convolution and bn layers and then fed into the jump expansion context aggregation module to aggregate with global features. By sparsely selecting query-centric keys and values, locality and sparsity are explicitly satisfied, and a model of long-distance dependencies is effectively established.

[0046] The skip-multiscale extended self-attention mechanism SMDSA includes:

[0047]

[0048] In the formula, This represents the feature map after SMDSA processing, where r represents the expansion rate, controlling sparsity. Q, K, and V represent the query matrix, key matrix, and value matrix generated from local features, respectively. Each row of the three matrices represents a query / key / value feature vector, as defined below:

[0049]

[0050] In the formula, W Q , and The parameters are linear mapping parameters. For position (i,j) in the original feature map, SMDSA sparse selection of keys and values ​​undergoes self-attention within a sliding window of size w×w centered at i. The output of the skip extended context aggregation module operation is defined as follows:

[0051]

[0052] Output using heads with different expansion rates The data is concatenated together and sent to a linear layer for feature aggregation.

[0053]

[0054] By setting different dilation rates for different heads, the skip extended context aggregation module can effectively aggregate semantic information from the encoder and decoder at different scales within the receptive domain; each head uses a different dilation rate, allowing the module to extract features at multiple scales.

[0055] The output of the linear layer is compared with the initial global features. and local features Add:

[0056]

[0057] Another object of the present application is to provide a pond monocular image depth estimation system for the pond monocular image depth estimation method according to the present application, comprising: a pond monocular image depth estimation model;

[0058] The pond monocular image depth estimation model is used to obtain the depth map of the pond scene based on the monocular pond image, obtain the distance information of each pixel point to the camera, generate the pond image depth map based on the distance information, and realize the depth estimation of the pond image.

[0059] The pond monocular image depth estimation model comprises: a U-MonoVIT encoder, a local plane guide layer, and a jump multi-scale extended self-attention mechanism.

[0060] The U-MonoVIT encoder is used to extract multi-scale image features, and comprises: a local perception unit and a parallel feature extraction module.

[0061] The local plane guide layer is used to obtain the local feature information around the pixel point by fitting the local image block through 4D plane coefficients.

[0062] The jump multi-scale extended self-attention mechanism is used to set different expansion rates for different heads, and effectively aggregate the semantic information of different scales in the receptive field of the encoder and the decoder.

[0063] Another object of the present application is to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the pond monocular image depth estimation method according to the present application.

[0064] Another object of the present application is to provide a computer readable storage medium having a computer program stored thereon, wherein when the processor executes the computer program, the processor executes the pond monocular image depth estimation method according to the present application.

[0065] The present application has the following advantages:

[0066] The pond monocular image depth estimation method and system disclosed by the present application can accurately obtain underwater depth images, provide support for intelligent aquaculture research, and provide a basis for pond environment monitoring.

[0067] With respect to the binocular system commonly used in the prior art, the depth estimation method and system for a pond monocular image disclosed by the application realizes monocular pond image depth estimation based on intrinsic scale invariance and jump expansion context fusion, and has the following technical advantages:

[0068] 1. The monocular method disclosed by the application avoids the need for more computing resources, including image preprocessing, matching cost calculation, disparity estimation, etc., due to the involvement of two image processing in the binocular system. The monocular method only processes one image, greatly reducing the consumption of computing resources. The monocular depth estimation method only needs one camera, and compared with the two cameras required by the binocular system, the hardware configuration is simpler and the calculation cost is lower.

[0069] 2. The encoder framework of U-MonoVIT is disclosed in the application, and U-MonoVIT has multiple spatial pyramid reduction modules. In each VIT layer, U-MonoVIT has a module parallel to the multi-head self-attention module, which fuses and feeds the features into the feedforward network. This design indeed embodies the intrinsic scale invariance of the model, and solves the problem of low accuracy of underwater depth estimation in the pond caused by underwater image blurring and low contrast.

[0070] 3. The application adopts a local plane guide layer, which obtains the local feature information around the pixel point by fitting the local image block with 4D plane coefficients. Each pixel will obtain a 4D plane coefficient related to the local image block around it, which is used to represent the local feature of the pixel. By performing such an operation on the entire feature map, the model can capture the local features at different positions in the image, thereby improving the understanding of depth information and the processing ability of local details, and effectively solving the problem of loss of detail information caused by color deviation of underwater images.

[0071] 4. The application aggregates the semantic information of different scales within the receptive field of the encoder and the decoder through the proposed jump multi-scale expansion self-attention mechanism (SMDSA), establishes a more fine interaction between the wide-area feature and the local perception, effectively solves the problem of overexposure and dimness caused by the pond image when facing light changes, and better adapts to the depth map generation demand under different light environments. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 The flowchart of the depth estimation method for a pond monocular image of the application;

[0073] Figure 2 The overall flowchart of the pond monocular image depth estimation in the embodiment of the application;

[0074] Figure 3This is a schematic diagram of the overall structure of the U-MonoVIT encoder in an embodiment of the present invention;

[0075] Figure 4 This is a schematic diagram of the workflow of the local planar guide layer in an embodiment of the present invention;

[0076] Figure 5 This is a schematic diagram of the skipping multi-scale extended self-attention mechanism in an embodiment of the present invention;

[0077] Figure 6 This is a qualitative comparison of depth estimation results under low contrast conditions in an embodiment of the present invention, wherein: (a) is the input image, (b) is Groundtruth, (c) is GLPDepth, (d) is Adabins, (e) is PixelFormer, (f) is DepthFormer, (g) is LapDepth, and (h) is the depth estimation result obtained by the method disclosed in the present invention;

[0078] Figure 7 The following is a qualitative comparison of depth estimation results under different illumination changes in the embodiments of the present invention, wherein: (a) is the input image, (b) is Groundtruth, (c) is GLPDepth, (d) is Adabins, (e) is PixelFormer, (f) is DepthFormer, (g) is LapDepth, and (h) is the depth estimation result obtained by the method disclosed in the present invention;

[0079] Figure 8 This is a schematic diagram of the comparison of detail images for depth estimation in an embodiment of the present invention, wherein: (a) is the input image, (b) is the ground truth (GT), (c) is the GLPDepth, (d) is the Adabins, (e) is the LapDepth, and (f) is the detail image of depth estimation obtained by the method disclosed in the present invention. Detailed Implementation

[0080] This invention provides a method and system for depth estimation of monocular images of ponds. The invention will be further described in detail below with reference to the accompanying drawings.

[0081] like Figure 1 The embodiment of the present invention disclosed in the present invention provides a depth estimation method for monocular images of ponds. A depth estimation model for monocular images of ponds is constructed, the pond image is input into the depth estimation model of monocular images of ponds to obtain a depth mapping of the pond scene, the distance information from each pixel to the camera is obtained, and a depth map of the pond image is generated based on the distance information to realize the depth estimation of the pond image.

[0082] The construction of the pond monocular image depth estimation model includes the following steps:

[0083] In the encoding stage, given the current input image, the U-MonoVIT encoder serving as the backbone network extracts multi-scale image features; the U-MonoVIT encoder comprises a local perception unit and a parallel feature extraction module;

[0084] In the decoding stage, the local plane guide layer is adopted to obtain the local feature information around the pixel point through the fitting of the 4D plane coefficients to the local image block.

[0085] Through the jump multi-scale extended self-attention mechanism, different expansion rates are set for different heads, and the semantic information of different scales in the receptive field participated by the encoder and the decoder is effectively aggregated.

[0086] The application discloses a depth estimation method for a pond monocular image, which can solve the problem of reduced depth estimation accuracy caused by the influence of light scattering and suspended particles in water, and can effectively solve the problem of overexposure and dimness caused by light changes in the pond image, and can better adapt to the depth map generation demand in different environments, and provide a cost-effective and superior perception solution for intelligent pond breeding.

[0087] In the embodiment, the depth estimation method for the pond monocular image based on the intrinsic scale invariance and the jump extended context fusion pond monocular image depth estimation model has the following characteristics:

[0088] (1) The model designs an encoder framework named U-MonoVIT, the U-MonoVIT has multiple spatial pyramid reduction modules, and the design of the method embodies the intrinsic scale invariance of the model, which helps to reduce the interference of the blurred and low-contrast pond image and improve the clarity and accuracy of the underwater image depth map in the pond;

[0089] (2) The local plane guide layer is adopted to obtain the local feature information around the pixel point through the fitting of the local image block. The model can capture the local features at different positions in the image, thereby improving the understanding of the depth information and the processing ability of the local details;

[0090] (3) Through the jump multi-scale extended self-attention mechanism (SMDSA), the semantic information of different scales in the receptive field participated by the encoder and the decoder is aggregated, a more fine interaction between the wide-area features and the local perception is established, the problem of overexposure and dimness caused by light changes in the underwater image is effectively solved, and the depth map generation demand in different environments can be better adapted.

[0091] The implementation process of the method for depth estimation of a pond monocular image disclosed in the application is specifically introduced through a complete embodiment.

[0092] The U-MonoVIT encoder further comprises:

[0093] In the encoding stage, a given current input image is sequentially input into a convolution layer and four Transformer blocks; the convolution layer sequentially comprises a first convolution, a second convolution and a third convolution; the convolution kernels of the three convolutions have a size of 3*3, the step length of the first convolution and the second convolution is 1, and the step length of the third convolution is 2; the four Transformer blocks correspond to stages 1, 2, 3 and 4, the four stages correspond to four dimensions [C1, C2, C3, C4], which respectively correspond to proportional features of 1 / 4, 1 / 8, 1 / 16 and 1 / 32, and the depths of each stage are [3, 4, 6, 3]; the Transformer block comprises patch embedding and a VIT layer; the patch embedding converts a picture into a sequence; the VIT layer comprises a first branch and a second branch in parallel; the first branch comprises a local perception unit, a first normalization layer and a multi-head self-attention module; the second branch comprises a parallel feature extraction module; the outputs of the first branch and the second branch are added and then input into a second normalization layer and a multi-layer perception mechanism. In this embodiment, the pond monocular image depth estimation model adopts the encoder U-MonoVIT as a backbone network and adopts a spatial pyramid progressive reduction strategy to generate a feature map. The use of the backbone encoder network U-MonoVIT respectively through the local perception unit and the multi-head self-attention module parallel module comprehensively extracts multi-scale image features, which specifically includes:

[0094] First, in the encoding stage, given the current input image, the present application adopts a Conv-stem block composed of three convolutions, the size of the convolution kernel is 3*3, the step of the first two convolutions is 1, the size of the generated feature map is unchanged, and the step of the third convolution is 2, the size of the generated feature map becomes 1 / 2 of the original. Then the feature map passes through four Transformer blocks to generate four dimensions [C1, C2, C3, C4] corresponding to 1 / 4, 1 / 8, 1 / 16, 1 / 32 scale features, and the depths of each stage are [3, 4, 6, 3]. Reducing the number of Transformer layers at each scale can significantly reduce the computational load and memory consumption, making the model more efficient during inference. For the feature map of each stage, it is first converted into a sequence through a patch embedding, and then processed through a local perception unit in a complementary manner through convolution layers and transformer blocks. In each transformer layer, U-MonoVIT has a module parallel to the multi-head self-attention module, which fuses its features and feeds them into the feedforward network. This way of considering local and global information helps the model better understand the features of different scales in the image, thus improving its adaptability to scale changes and reflecting the scale invariance within the model.

[0095] In the present embodiment, the VIT layer is based on the traditional Vision Transformer (ViT), and the proposed local perception unit and parallel feature extraction module are introduced, thereby improving the feature extraction capability in the underwater blurred area. Then, the decoder gradually restores the spatial resolution of the image and generates a depth map using the semantic information extracted by the encoder. The local plane guiding layer and the proposed skip multi-scale extended self-attention mechanism (SMDSA) effectively solve the problem of overexposure and dimness caused by underwater images in the pond when facing light changes, thereby better adapting to the depth map generation needs in different environments.

[0096] The feature extraction link, the batch normalization layer BN, and the convolution layer Conv; the feature extraction link includes a convolution layer Conv, a batch normalization layer BN, and a SILU activation function;

[0097] The working process of the parallel feature extraction module is as follows: the feature map processed by the local perception unit is input into the parallel feature extraction module, first passes through a 3*3 convolution layer Conv, the convolution kernel size of the 3*3 convolution layer can effectively capture local details and global information, and enhance the feature expression ability; then, the feature map passes through a batch normalization layer BN, the batch normalization layer BN can accelerate network convergence, improve training stability, and reduce internal covariate shift; at the same time, a SILU (Sigmoid Linear Unit) activation function is used to further improve the feature extraction ability of the model for underwater blurred areas and enhance the perception effect on low contrast and uneven illumination areas; after the above process is executed twice, a 3*3 convolution layer Conv and a normalization layer BN are executed respectively; the design aims to fully extract local details and nonlinear features in the image, and improve the representation ability and robustness of the features.

[0098] In the decoding stage, first, at the stage of reducing the feature mapping resolution to H / 8, a more dense spatial pyramid pooling layer is used as the context information extractor of the method. Then, at the stage of reducing the feature mapping resolution to H / 8, H / 4 and H / 2, the method uses a local plane guide layer to locate the geometric guidance of the required depth estimation, and restores the features to full resolution. The up-sampling results at the stages of reducing the feature mapping to H / 8 and H / 4 are spliced together after passing through a 3*3 convolution and a bn layer, and are sent to the jump expansion context aggregation module, aggregated with global features, and the method explicitly meets the locality and sparsity by querying the center-based key and value, which can effectively establish a model of long-range dependencies. The method effectively solves the problem of overexposure and dimness of underwater images in the pool when facing light changes, and can better adapt to the depth map generation requirements in different lighting environments.

[0099] The following will be specifically explained and described for each sub-step:

[0100] In the embodiment, the arrangement process of the depth estimation method for the pool monocular image is as shown in Figure 2 The application will be further described with reference to the accompanying drawings.

[0101] In the model training part, first, input the pond image with depth annotation, use the backbone encoder network U-MonoVIT to extract multi-scale image features through the parallel module of local perception unit and multi-head self-attention module. Then, in the decoding stage, the local plane guide layer is adopted, and the local image block is fitted through the 4D plane coefficient to obtain the local feature information around the pixel point. Finally, through the proposed jump multi-scale expansion self-attention mechanism (SMDSA), different expansion rates are set for different heads, which effectively aggregates the semantic information of different scales in the receptive field of the encoder and decoder.

[0102] The overall structure of the U-MonoVIT encoder in the embodiment is as shown in Figure 3

[0103] In the model training process, the backbone network U-MonoVIT adopts a spatial pyramid reduction strategy to generate feature maps. If you want to generate 5 feature maps of different scales, you need 5 stages. In the first stage of encoding, given the current input image, the method adopts a Conv-stem block composed of three convolutions, and the size of the convolution kernel is 3*3. The step of the first two convolutions is 1, and the size of the generated feature map is unchanged. The step of the third convolution is 2, and the size of the generated feature map becomes 1 / 2 of the original. For the remaining 4 scale feature maps, they are first converted into a sequence through a Patch embedding process, and then processed through a local perception unit in a complementary manner through convolution layers and transformer blocks.

[0104] In the embodiment, the U-MonoVIT encoder extracts multi-scale image features, including:

[0105] Stage i is defined as:

[0106]

[0107] In the formula, DWConv(x) represents a deep convolution.

[0108] Local context is embedded in visual markers using a parallel feature extraction module (PFEM), and the fusion mode of the parallel feature extraction module is represented as:

[0109]

[0110] In the formula, is the feature map processed by the parallel feature extraction module of stage i, and PFEM(x) is the parallel feature extraction module, ​For multi-head attention mechanism;

[0111] When the local perception unit and the parallel feature extraction module construct the local relationship between the intra-feature neighbors, the transformer block models the information interaction and long-distance dependence on the entire input space within the feature, benefiting from the self-attention mechanism.

[0112] Specifically, the multi-head self-attention module transforms the original input to query through linear transformation and value . Where n = H x W is the number of visual markers:

[0113] MHA(Q, K, V) = Concat(head0, …, head N )W O (3)

[0114] head i = Attention(QW i Q , KW i K , VW i V ) (4)

[0115]

[0116] In the formula, MHA(Q, K, V) represents the output of multi-head attention, head N represents the output of the Nth attention head; W O represents the projection matrix used to splice the projected dimension; Q represents the query Query, K T represents the key Key, d k represents the scaling factor, V represents the value matrix Value, W i Q represents the Q learning weight matrix, W i K represents the K learning weight matrix, and W i V represents the V learning weight matrix; Softmax represents the normalized weight;

[0117] The fused features are fed into the feedforward network, which is represented by the following formula:

[0118] Y = MHSA(Norm(X)) + X (6)

[0119] Z = MLP(Norm(Y)) + Y (7)

[0120] In the formula, MLP represents a multi-layer perceptron (MLP), Norm represents Layer Normalization on YYY to ensure numerical stability, X represents the input of the encoder layer, Y represents the output after multi-head self-attention calculation, and Z represents the final output feature.

[0121] This design embodies the intrinsic scale invariance of the model, solving the problem of low depth estimation accuracy caused by pond image blur and low contrast.

[0122] The workflow of the local plane guide layer in the embodiment is as shown in Figure 4 .

[0123] In the model training process, in the decoding stage, first, at the stage of reducing the feature mapping resolution to H / 8, a more dense spatial pyramid pooling layer is used as the context information extractor of the method. Then, at the stage of reducing the feature mapping resolution to H / 8, H / 4 and H / 2, the method uses a local plane guide layer to locate the required depth estimation geometric guide and restore the feature to full resolution.

[0124] In the embodiment, the local plane guide layer is used to obtain the local feature information around the pixel point by fitting the local image block with 4D plane coefficients, which specifically includes:

[0125] Given a feature map with spatial resolution H / k, the proposed layer estimates a 4D plane coefficient for each spatial unit. In order to guide the feature with local plane hypothesis, the ray-plane intersection is used to convert each estimated four-dimensional plane coefficient into kxk local depth cues c i :

[0126]

[0127] where n1n2n3n4 represents the estimated plane coefficient, u i , v i kxk block-normalized coordinates of pixel i.

[0128] Through a stack of 1x1 convolutions, the number of channels is repeatedly reduced by a factor of 2 until it reaches 3. Then, the feature map is passed through two different methods to get the local plane coefficient estimation: one method is to convert to a unit normal vector, and the other is to define a sigmoid-type function of the distance from the plane to the origin. After the sigmoid function, the output is multiplied by the maximum distance k to get the actual depth value. Because the unit normal vector has only two degrees of freedom (i.e., the polar angle θ and the azimuth angle φ). The first two channels of the given feature map are considered as angles, and they are converted to a unit normal vector using the following formula:

[0129]

[0130] The unit normal vector includes two degrees of freedom: polar angle theta and azimuth angle In this way, the model can capture local features at different positions in the image, thereby improving the understanding of depth information and the processing ability of local details.

[0131] The up-sampling results of the feature map reduction to H / 8 and H / 4 stage are spliced together with the down-sampling results of the local plane guide layer restored to full resolution after passing through the 3*3 convolution and bn layer, to obtain the local feature Given the global feature And the local feature The method performs 3*3 convolution on And The jump expansion context aggregation module. In order to avoid the problem of gradient disappearance and gradient explosion, the method normalizes the global feature And the local feature To make the distribution of the input more stable:

[0132]

[0133] In the formula, LayerNorm is layer normalization, is the normalized feature of the global feature , is the normalized feature of the local feature ;

[0134] In the decoding stage, first, in the stage of feature map resolution reduction to H / 8, a more dense spatial pyramid pooling layer is used as the context information extractor of the method; then in the stage of feature map resolution reduction to H / 8, H / 4 and H / 2, the local plane guide layer is used to locate the required depth estimation geometry guide, and the feature is restored to full resolution.

[0135] The arrangement process of the jump multi-scale expansion self-attention mechanism in the embodiment is as shown in Figure 5 .

[0136] According to the locality and sparsity of shallow global attention in Vision transformer (ViTs), the method is proposed to center on the query patch generated by local feature, and then the keys and values generated by local feature are sparsely selected by sliding window mechanism, and then the representative patches are self-attended. First, the channels of feature mapping are divided into multiple heads, and different heads use different dilation rates. By default, a 3*3 convolution kernel is used, and the dilation rates are set to r=1, 2 and 3 respectively. In this way, the receptive field sizes of different heads are 3*3, 5*5 and 7*7 respectively. By setting different dilation rates for different heads, the skip-extended context aggregation module effectively aggregates the semantic information of different scales within the receptive field, effectively reduces the redundancy of the self-attention mechanism, and does not need complex operations and additional computational cost.

[0137] In the embodiment, the specific process of setting different dilation rates for different heads by the skip multi-scale extended self-attention mechanism is as follows:

[0138] The up-sampling results of the feature mapping reduced to H / 8 and H / 4 stage and the down-sampling results recovered to full resolution by using local plane guide layer are spliced together after passing through 3*3 convolution and bn layer, and then sent to the skip-extended context aggregation module to be aggregated with global features, and the keys and values centered on the query are sparsely selected to explicitly meet the locality and sparsity, and effectively establish the model of long-range dependency;

[0139] The skip multi-scale extended self-attention mechanism comprises:

[0140]

[0141] Wherein, r represents the dilation rate, which controls the sparsity, Q, K, V represent the query matrix generated by local feature, the key and value matrix generated by local feature respectively, and each row of the three matrices represents a query / key / value feature vector, which is defined as follows:

[0142]

[0143] Wherein, W Q 、 and are parameters of linear mapping, and for the position (i,j) in the original feature mapping, the SMDSA sparse selection key and value are self-attended in the sliding window with the size of w*w centered on i, and the output definition of the skip-extended context aggregation module operation is as follows:

[0144]

[0145] The outputs of the heads using different dilation rates are combined to obtain the output of the skip multi-scale extended self-attention mechanism. Concatenate together and then send to linear layer for feature aggregation:

[0146]

[0147] By setting different expansion rates for different heads, the skip-extended context aggregation module can effectively aggregate semantic information of different scales of the encoder and decoder within the receptive field. Each head adopts a different expansion rate, enabling the module to extract features at multiple scales. This approach not only enhances the model's ability to capture multi-scale information but also reduces the redundancy of the self-attention mechanism while maintaining computational efficiency. In this way, the model can more comprehensively understand and process the complex structure of input data. Finally, the output of the linear layer is added to the initial global feature and local feature :

[0148]

[0149] After the training process, the model for pond monocular depth estimation is obtained. In the depth estimation process, based on the trained model, the depth map of the pond scene can be obtained, that is, the distance information from each pixel point to the camera, which will be represented by a depth map.

[0150] In this embodiment, the above-mentioned depth estimation method of the pond monocular image is a full encoder-decoder structure, first, the U-MonoVIT encoder structure is designed, U-MonoVIT is to adopt multiple spatial pyramid reduction modules as the backbone network, and the multi-scale image features are extracted. In each VIT layer, U-MonoVIT has a convolution block in parallel with the multi-head self-attention module, which fuses and feeds the features into the feedforward network. This way of considering local and global information helps the model better understand the features of different scales in the image, thus improving its adaptability to scale changes. This design reflects the intrinsic scale invariance of the model, which helps to reduce the interference of blurred and low-contrast pond images and improve the clarity and accuracy of the underwater image depth map. Then the local plane guide layer is adopted, and the 4D plane coefficients are used to fit the local image blocks to obtain the local feature information around the pixel point. Each pixel will get a 4D plane coefficient related to the local image block around it, which is used to represent the local features of the pixel. By performing this operation on the entire feature map, the model can capture the local features at different positions in the image, thus improving the understanding of depth information and the ability to handle local details, solving the problem of loss of detail information due to light absorption. The proposed jump multi-scale extended self-attention mechanism (SMDSA) sets different expansion rates for different heads, effectively aggregates the semantic information of different scales within the receptive field of the encoder and decoder, effectively reduces the redundancy of the self-attention mechanism, without complex operations and additional computational cost, and adaptively adjusts the weights of global context and low-level information, effectively solving the problem of overexposure and dimness caused by light changes in underwater images in the pond. The method can better adapt to the depth map generation requirements under different lighting environments.

[0151] In this embodiment, the effectiveness of the disclosed pond monocular image depth estimation method and system is verified through ablation experiments on underwater datasets. By systematically removing or replacing different components of the model, the influence of each part on the overall performance is demonstrated. By replacing the encoder U-MonoVIT, the SMDSA module and the LPG layer, it is observed that the three components disclosed in the present application have an important influence on the accuracy of the depth estimation result. The experimental results are shown in Table 1, in which the larger the index value of the upward arrow is, the better, and the smaller the index value of the downward arrow is, the better.

[0152] Table 1: Results of ablation experiment

[0153]

[0154] With the increase of the core factors, the overall performance is improved, and U-MonoVIT is improved by 32.8% and 45.8% in AbsRel and SqRel compared with baseline. The increase effect of the encoder U-MonoVIT is the most significant, which shows that the performance of U-MonoVIT is obviously better than that of Densenet-161 as the baseline network.

[0155] The comparison results of the method disclosed in the present application and various monocular depth estimation methods in the prior art are shown in Table 2.

[0156] Table 2 Quantitative comparison of depth estimation performance of the method of the present application and existing schemes

[0157]

[0158] The quantitative performance comparison of the FLSea dataset

[36] is shown in Table 2. The proposed model shows excellent performance in all monocular depth evaluation indicators δ < 1.25, δ < 1.25 2 , δ < 1.25 3 , AbsRel, SqRel, RMSE, RMSElog, SILlog, log 10 , which indicates that the method of the present application performs higher accuracy in the depth estimation task. According to the number of parameters in Table 2, the proposed architecture is compared with the existing excellent methods. It can be seen that the number of parameters of the proposed method is much smaller than that of Adabins, IEbins, PixelFormer, DepthFormer, LapDepth, GLPDepth. The design of the present application can achieve better depth estimation performance than Pixelformer, although only 13.7% of the calculation cost is required. Therefore, the method proposed in the present application realizes satisfactory performance under the premise of keeping fewer parameters, enhances its adaptability and practicality in the environment with limited computing resources.

[0159] In this embodiment, the qualitative comparison of the depth estimation results under low contrast conditions is shown in Figure 6 , and Figure 6 , in which: (a) is the input image, (b) is the Groundtruth, (c) is GLPDepth, (d) is Adabins, (e) is PixelFormer, (f) is DepthFormer, (g) is LapDepth, and (h) is the depth estimation result obtained by the method disclosed in the present application; from Figure 6As can be seen, the LapDepth and DepthFormer methods lose a lot of detailed depth information in low-contrast scenes, because their network structures pay relatively weak attention to local structures. Even in the presence of small-scale objects (such as corals) in a blurred background, the model of the present application can effectively capture their depth information, which shows that the method of the present application still has strong resolution capability under low-contrast conditions.

[0160] In this embodiment, the qualitative comparison results of depth estimation results under different illumination changes are as shown in Figure 7 Figure 7 In the figures: (a) is the input image, (b) is the Groundtruth, (c) is GLPDepth, (d) is Adabins, (e) is PixelFormer, (f) is DepthFormer, (g) is LapDepth, and (h) is the depth estimation result obtained by the method disclosed in the present application; under these extreme lighting conditions, the depth estimation accuracy of GLPDepth has decreased, and it is difficult to accurately capture detailed features, because its network structure lacks understanding of fine-grained depth information. The robustness and accuracy of LapDepth need to be further improved when processing scenes with strong changes in illumination. The model of the present application can accurately estimate the depth of the foreground and background in the scene under low light and overexposure conditions, and achieves superior depth perception in the entire scene.

[0161] In this embodiment, the detailed comparison images of depth estimation are as shown in Figure 8 Figure 8 In the figures: (a) is the input image, (b) is the GT, (c) is GLPDepth, (d) is Adabins, (e) is LapDepth, and (f) is the detailed image of the depth estimation obtained by the method disclosed in the present application. Due to the absorption of light, red light decays most severely underwater, resulting in the loss of detailed information. In Figure 8 In this embodiment, the method disclosed in the present application is visually compared with GLPDepth, Adabins, LapDepth, and the method of the present application on the depth image map. The results show that GLPDepth is good at capturing global features of the image, and this structure has shortcomings when processing detailed information, so it produces inaccurate depth boundaries. LapDepth has an unsatisfactory accuracy when processing distant and blurred targets, because of the limitations brought by the ordinary skip connection of its decoder. In contrast, the method disclosed in the present application is the most robust in processing global structures and local details, as shown by the red and green boxes in Figure 8

[0162] ​​​Another embodiment of the present application discloses a pond monocular image depth estimation system for the pond monocular image depth estimation method described in the present application, comprising: a pond monocular image depth estimation model;

[0163] The pond monocular image depth estimation model is used to obtain the depth map of the pond scene based on the monocular pond image, obtain the distance information of each pixel point to the camera, generate the pond image depth map based on the distance information, and realize the depth estimation of the pond image;

[0164] The pond monocular image depth estimation model comprises: a U-MonoVIT encoder, a local plane guide layer and a jump multi-scale extended self-attention mechanism.

[0165] The U-MonoVIT encoder is used to extract multi-scale image features, and the U-MonoVIT encoder comprises a local perception unit and a parallel feature extraction module.

[0166] The local plane guide layer is used to obtain the local feature information around the pixel point by fitting the local image block through 4D plane coefficients.

[0167] The jump multi-scale extended self-attention mechanism is used to set different expansion rates for different heads, effectively aggregate the semantic information of different scales in the receptive field of the encoder and the decoder.

[0168] Another embodiment of the present application discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the pond monocular image depth estimation method according to the present application.

[0169] Another embodiment of the present application discloses a computer readable storage medium, which stores a computer program, and when the processor executes the computer program, the processor executes the pond monocular image depth estimation method according to the present application.

[0170] In summary, the application discloses a kind of for pond monocular image depth estimation method and system, based on convolution exploration local feature Vision Transformer and jump multi-scale context fusion, proposes a kind of named U-MonoVIT encoder framework.In each VIT layer, U-MonoVIT designs a feature extraction module in parallel with multi-head self-attention module, to fuse features and feed into feedforward network.Through the novel local plane guide layer of decoding stage, the dense encoded features are guided to depth prediction.Finally, the jump multi-scale extended self-attention mechanism (SMDSA) is used to aggregate the semantic information of different scales in the encoder and decoder.The method effectively solves the problem of reduced depth estimation accuracy caused by blurred, low contrast, overexposed or dim pond images, and accurately and efficiently realizes monocular depth estimation in pond scenes, providing key support for comprehensive exploration of pond environment and biomass estimation.

Claims

1. A method for depth estimation of pond monocular images, characterized in that, The pond monocular image depth estimation model is constructed, the pond image is input into the pond monocular image depth estimation model, the depth map of the pond scene is obtained, the distance information from each pixel point to the camera is obtained, the pond image depth map is generated based on the distance information, and the depth estimation of the pond image is realized; The pond monocular image depth estimation model is constructed, the pond image is input into the pond monocular image depth estimation model, the depth map of the pond scene is obtained, the distance information from each pixel point to the camera is obtained, the pond image depth map is generated based on the distance information, and the depth estimation of the pond image is realized; In the encoding stage, given the current input image, the U-MonoVIT encoder serving as the backbone network extracts multi-scale image features; the U-MonoVIT encoder comprises a local perception unit and a parallel feature extraction module; In the decoding stage, a local plane guide layer is adopted to obtain the local feature information around the pixel point by fitting the local image block through 4D plane coefficients; Through the jump multi-scale expansion self-attention mechanism, different expansion rates are set for different heads, and the semantic information of different scales in the receptive field participated by the encoder and the decoder is effectively aggregated; The U-MonoVIT encoder further comprises: In the encoding stage, the given current input image is sequentially input into a convolution layer and four Transformer blocks; the one convolution layer sequentially comprises a first convolution, a second convolution and a third convolution; the convolution kernel sizes of the three convolutions are 3*3, the step lengths of the first convolution and the second convolution are 1, and the step length of the third convolution is 2; the four Transformer blocks correspond to stages 1, 2, 3 and 4, and the four stages correspond to four dimensions [C1, C2, C3, C4], which respectively correspond to 1 / 4, 1 / 8, 1 / 16 and 1 / 32 proportional features, and the depths of each stage are [3, 4, 6, 3]; the Transformer block comprises a patch embedding and a VIT layer; the patch embedding converts the picture into a sequence; the VIT layer comprises a parallel first branch and a second branch; the first branch comprises a local perception unit, a first normalization layer and a multi-head self-attention module; the second branch comprises a parallel feature extraction module; the outputs of the first branch and the second branch are added and input into a second normalization layer and a multi-layer perception mechanism; The parallel feature extraction module comprises: A feature extraction link, a batch normalization layer BN and a convolution layer Conv; the feature extraction link comprises a convolution layer Conv, a batch normalization layer BN and a SILU activation function; The working process of the parallel feature extraction module is that: the feature map processed by the local perception unit is input into the parallel feature extraction module, first passes through a 3x3 convolution layer Conv, the convolution kernel size of the 3x3 convolution layer can effectively capture local details and global information, and enhance the feature expression ability; then, the feature map passes through a batch normalization layer BN, the batch normalization layer BN can accelerate network convergence, improve training stability, and reduce internal covariate shift; at the same time, a SILU (Sigmoid Linear Unit) activation function is used to further improve the feature extraction ability of the model for underwater blurred areas and enhance the perception effect of low contrast and unevenly illuminated areas; after the above process is executed twice, a 3x3 convolution layer Conv and a normalization layer BN are executed respectively; The local perception unit comprises a 3X3 deep convolution; The working process of the local perception unit is that: first, the input image is divided into small blocks (Patches) and converted into a sequence representation; then, the Patches pass through a 3x3 deep convolution layer to extract local features while preserving spatial information; in order to enhance the feature expression ability, the feature map processed by the deep convolution is added to the feature map without deep convolution to realize feature complementation; the fused feature map is further processed by a convolution layer block to capture local details and global dependencies at the same time; The local plane guide layer uses 4D plane coefficients to fit the local image block to obtain the local feature information around the pixel point, and specifically comprises: Given a feature map with spatial resolution H / k, the proposed layer estimates a 4D plane coefficient for each spatial cell; to guide the features with a local plane hypothesis, each estimated four-dimensional plane coefficient is converted to a k x k local depth cue c using ray-plane intersection i : where n1n2n3n4denote the estimated planar coefficients, u i , v i k x k block normalized coordinates of pixel i; Through a stack of 1x1 convolutions, the channel number is repeatedly reduced by 2 times until the channel number reaches 3; the local plane coefficient estimation is obtained by transmitting the feature mapping through the first method and the second method; the first method is to convert into a unit normal vector, and the second method is to define a sigmoid function of the vertical distance of the plane to the origin; after the sigmoid function, the output is multiplied by the maximum distance k to obtain the actual depth value; The first two channels of the given feature mapping are regarded as angles, and they are converted into unit normal vectors using the following formula: The unit normal vector includes two degrees of freedom: polar angle θ and azimuth angle φ The up-sampling results of the feature mapping reduced to H / 8 and H / 4 stages are spliced together after passing through a 3*3 convolution and a bn layer at the same time as the down-sampling results recovered to full resolution using a local plane guide layer, to obtain local features Given global features and local features Perform 3x3 convolution on and , and input into the jump multi-scale expansion self-attention mechanism; by normalizing the global features and local features , the distribution of the input is more stable: In the formula, LayerNorm is layer normalization, global feature normalized feature, local feature normalized feature.

2. The method for depth estimation from pond monocular images according to claim 1, wherein, The U-MonoVIT encoder extracts multi-scale image features, which comprises: The definition stage i is: In the formula, DWConv(x) denotes a depthwise convolution; The local context is embedded in the visual marker using the parallel feature extraction module, and the fusion mode of the parallel feature extraction module is represented as: In the formula, is the feature map processed by the parallel feature extraction module of stage i, PFEM(x) is the parallel feature extraction module, is a multi-head attention mechanism; The local relationship between the features in the neighborhood is constructed based on the local perception unit and the parallel feature extraction module; Multi-head self-attention module transforms the original input to query key and value where n = H x W is the number of visual markers: MHA(Q, K, V) = Concat(head0,..., head N )W O (3) head i = Attention(QW i Q ,KW i K ,VW i V ) (4) where MHA(Q, K, V) represents the output of multi-head attention, head N represents the output of the Nth attention head; W O represents the projection matrix for concatenating the projected dimensions; Q represents the query, K T represents the key, d k represents the scaling factor, V represents the value matrix; W i Q represents the Q-learning weight matrix, W i K represents the K-learning weight matrix, W i V represents the V-learning weight matrix; Softmax represents the normalization weight The fused features are fed into a feedforward network, which is represented as: Y = MHSA(Norm(X)) + X (6) Z = MLP(Norm(Y)) + Y (7) In the formula, MLP represents a multi-layer perception machine, Norm represents Layer Normalization on Y, which ensures numerical stability, X represents the input of the encoder layer, Y represents the output after multi-head self-attention calculation, and Z represents the final output feature.

3. The method for depth estimation from pond monocular images according to claim 1, wherein, The different expansion rates are set for different heads through the skip multi-scale expansion self-attention mechanism, which specifically comprises: The up-sampling results of the feature mapping reduced to H / 8 and H / 4 stages are spliced together with the down-sampling results recovered to full resolution by using a local plane guide layer, and then sent into a jump expansion context aggregation module after passing through a 3*3 convolution and a bn layer, and aggregated with global features, and through sparse selection, the keys and values centered on the query are queried, the locality and sparsity are explicitly met, and the model of effectively establishing long-range dependencies is established; The jump multi-scale expansion self-attention mechanism SMDSA comprises: In the formula, The formula represents the feature map after SMDSA processing, r represents the expansion rate, the sparsity is controlled, Q, K, and V represent the query matrix, the key matrix, and the value matrix generated by the local feature respectively, and each row of the three matrices represents a query / key / value feature vector, which is specifically defined as follows: where W Q , and are parameters of linear mapping, for the original feature map at position (i, j), SMDSA sparse selection key and value are self-attention in the sliding window of size w x w centered at i, the output of the jump extended context aggregation module operation is defined as follows: The output will use different dilated rate heads Padded together, sent to linear layer for feature aggregation: By setting different expansion rates for different heads, the jump expansion context aggregation module can effectively aggregate semantic information of different scales of the encoder and the decoder in the receptive field; different expansion rates are used for each head, so that the module can extract features at multiple scales; The output of the linear layer is added to the initial global feature and local feature ​ 4. The pond monocular image depth estimation system for the method of depth estimation of pond monocular images according to any one of claims 1 to 3, characterized in that, Comprise: A pond monocular image depth estimation model; The pond monocular image depth estimation model is used to obtain the depth map of the pond scene based on the monocular pond image, obtain the distance information from each pixel point to the camera, generate the pond image depth map based on the distance information, and realize the depth estimation of the pond image; The pond monocular image depth estimation model comprises a U-MonoVIT encoder, a local plane guide layer and a jump multi-scale expansion self-attention mechanism. The U-MonoVIT encoder is used to extract multi-scale image features; the U-MonoVIT encoder comprises a local perception unit and a parallel feature extraction module. The local plane guide layer is used to obtain the local feature information around the pixel point by fitting the local image block through 4D plane coefficients. The jump multi-scale expansion self-attention mechanism is used to set different expansion rates for different heads, effectively aggregating semantic information of different scales in the receptive field of the encoder and the decoder.

5. A computer device, comprising: The computer program is executed by the processor, and the processor executes the depth estimation method for the pond monocular image according to any one of claims 1-3.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor, and the processor executes the depth estimation method for the pond monocular image according to any one of claims 1-3.

Images