Edge contour enhanced EA-LPNet monocular depth estimation method

Abstract

The invention relates to an edge contour enhanced EA-LPNet (LPNet integrating EMA and ACmix) monocular depth estimation method, which comprises the following steps of: 1, constructing a multi-scale network structure by utilizing a Laplacian pyramid, and putting forward and using an improved cavity space pyramid pooling module on an encoder side; step 2, processing edge contour information by using a specific edge filtering operator, and fusing the edge contour information into the network; step 3, utilizing an efficient multi-scale attention mechanism EMA module to splice the edge residual image, the potential feature image and a previous-level depth residual image; 4, replacing a convolution module in a traditional decoder with a self-attention convolution module ACmix, and obtaining potential features in the depth residual image subjected to pre-dimensionality reduction in advance; and step 5, proposing and using an edge information loss function, and designing a specific loss function combination strategy for training. According to the method, the small target identification capability and the performance of the algorithm in the aspect of edge contours are effectively improved, and the network precision is improved to a certain extent.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and deep learning technology, and specifically provides an EA-LPNet (LPNet integrating EMA and ACmix) monocular depth estimation method with edge contour enhancement. Background Art

[0002] Depth estimation is a core technology for tasks such as three-dimensional scene reconstruction, obstacle detection, and autonomous driving. Traditional methods rely on active sensors such as LiDAR and millimeter-wave radar, but they suffer from high costs and susceptibility to weather interference. Depth estimation methods based on monocular cameras have become a research hotspot due to their low cost and high flexibility. However, existing depth estimation methods suffer from insufficient recognition of small targets and blurred edges. Traditional convolutional neural networks (CNNs) are prone to losing small target details in deep feature maps, resulting in depth estimation errors. The edge contours of the depth estimation map are often blurred due to insufficient feature fusion, affecting the accuracy of small target positioning. Summary of the Invention

[0003] In order to overcome the shortcomings of existing monocular depth estimation methods in small target recognition and edge detail preservation, the present invention provides an edge contour enhanced EA-LPNet (LPNet integrating EMA and ACmix) monocular depth estimation method, which effectively improves the small target recognition ability and the algorithm's performance in edge contours, and improves the network accuracy to a certain extent.

[0004] The technical solution adopted by the present invention to solve its technical problem is:

[0005] An edge contour enhanced EA-LPNet monocular depth estimation method includes the following steps:

[0006] Step 1: Use the Laplacian pyramid to build a multi-scale network structure. On the encoder side, propose and use the improved atrous spatial pyramid pooling (L-ASPP) module to obtain network features at different scales.

[0007] Step 2: Use a specific edge filter operator to process edge contour information, integrate the edge contour information into the network, and combine it with the potential feature information and the previous level depth residual information;

[0008] Step 3: Leveraging the excellent multi-channel and cross-dimensional feature extraction capabilities of the efficient multi-scale attention mechanism (EMA) module, the edge residual map, the latent feature map, and the previous level depth residual map are spliced together. Through channel reorganization, the connection between the three is deepened, thereby enhancing the extraction of effective information between the three.

[0009] Step 4: Use the self-attention convolution module ACmix to replace the convolution module in the traditional decoder, so that it has better pixel-level attention and obtains the potential features of the pre-dimensionality-reduced depth residual map in advance, so that it can retain more effective information during dimensionality reduction;

[0010] Step 5: Design a monocular depth estimation loss function to increase the network's attention to depth information, gradient information, and edge contour information.

[0011] Furthermore, the process of step one is:

[0012] Step 1.1 uses the Laplace pyramid structure to build a monocular depth estimation network LPNet based on the Laplace pyramid;

[0013] Step 1.2 proposes an improved dilated spatial pyramid pooling method: the L-ASPP structure is used on the encoder side to process the input color image through convolution blocks with different expansion rates to obtain network features at different scales, so that the feature information obtained in the previous layer can be further involved in the operation of the next layer, and the deep feature information can be more comprehensively retained.

[0014] Step 1.3: The L-ASPP output is divided into two branches, which serve as the deep feature information of the current layer decoder and the supplementary information for subsequent operations in the previous layer decoder respectively;

[0015] Furthermore, the process of step 2 is:

[0016] Step 2.1 grayscale the original three-channel RGB image.

[0017] Step 2.2 uses the Laplacian operator to perform edge filtering on the obtained grayscale image.

[0018] Step 2.3 Obtain the edge contour map E after edge filtering K Multiply it by appropriate weight parameters and add them to the network to enhance the contour recognition ability.

[0019] Furthermore, the process of step three is:

[0020] Step 3.1: Concatenate the edge residual map, potential feature map, and previous level depth residual map and input them into the EMA module.

[0021] Step 3.2: Divide the sub-feature groups and input the feature maps into the three parallel branches of EMA.

[0022] Step 3.3 Generate a spatial attention map through information aggregation method.

[0023] Step 3.4 combines channel attention with spatial attention and outputs an enhanced feature map.

[0024] Furthermore, the process of step 4 is:

[0025] Step 4.1 performs three 1×1 convolution operations on the input feature map to obtain N sets of projected feature maps C / N×H×W shared by the self-attention module and the convolution module, and inputs them into the corresponding path.

[0026] Step 4.2 performs unidirectional displacement calculation on the output feature tensor.

[0027] Step 4.3 The feature map processed by the ACmix module is concatenated with the feature map of the previous layer to obtain the output D of this layer. k , the depth map D output by each layer k The final depth prediction image D1 is obtained by splicing.

[0028] Furthermore, the process of step five is:

[0029] Step 5.1 defines the depth loss function, taking the difference between the estimated value and the true value to reduce the impact of imbalance.

[0030] Step 5.2 defines the gradient loss function. For the sparse distribution problem of depth label values, the interpolation method is used to obtain the depth gradient value.

[0031] Step 5.3 adds an edge information loss function to enhance contour details. This function uses edge contour information as weights to increase the network's focus on target edges. The edge contour information extracted by the Sobel edge detection operator is used as weights to increase the algorithm's focus on contour boundaries.

[0032] Step 5.4: Sum the above depth loss function, gradient loss function, and edge information loss function to obtain the monocular depth estimation loss function actually used in the algorithm training process.

[0033] Step 5.5 Obtain the KITTI dataset for model training and testing.

[0034] The beneficial effects of the present invention are mainly manifested in:

[0035] 1. By integrating the efficient multi-scale attention (EMA) module with the self-attention convolution module (ACmix) to replace traditional convolution, the cross-dimensional connection between channels and space is enhanced, the network's ability to extract contextual semantic information is improved, and the network's ability to extract local information in the depth residual map is strengthened, effectively improving the ability to recognize small objects and, to a certain extent, improving network accuracy;

[0036] 2. By using the edge contour information extracted by a specific edge filtering operator to replace the Laplace residual and using the edge contour information as the weight parameter of the loss function, the edge contour information of the object in the depth feature map is enriched. The network pays more attention to the edge contour of the object, effectively improving the edge contour details, optimizing the depth estimation edge contour, and effectively improving the performance of the algorithm in terms of edge contour. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the overall technical route of the present invention.

[0038] Figure 2 This is the network structure diagram of the present invention.

[0039] Figure 3 This is the improved atrous spatial pyramid pooling (L-ASPP) structure used in the present invention.

[0040] Figure 4 The self-attention convolution (ACmix) module used in this invention.

[0041] Figure 5 The efficient multi-scale attention (EMA) module used in this paper. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings.

[0043] Reference Figures 1 to 5 , an edge contour enhanced EA-LPNet monocular depth estimation method, comprising the following steps:

[0044] Step 1: Use the Laplacian pyramid to build a multi-scale network structure. On the encoder side, propose and use the improved atrous spatial pyramid pooling (L-ASPP) module to obtain network features at different scales.

[0045] Step 2: Use a specific edge filter operator to process edge contour information, integrate the edge contour information into the network, and combine it with the potential feature information and the previous level depth residual information; the process is:

[0046] Step 3: Leveraging the excellent multi-channel and cross-dimensional feature extraction capabilities of the efficient multi-scale attention mechanism (EMA) module, the edge residual map, the latent feature map, and the previous level depth residual map are spliced together. Through channel reorganization, the connection between the three is deepened, thereby enhancing the extraction of effective information between the three.

[0047] Step 4: Use the self-attention convolution module ACmix to replace the convolution module in the traditional decoder, so that it has better pixel-level attention and obtains the potential features in the pre-dimensionality-reduced depth residual map in advance, so that it can retain more effective information during dimensionality reduction.

[0048] Step 5: Design a monocular depth estimation loss function to increase the network's attention to depth information, gradient information, and edge contour information.

[0049] Figure 1 This is the overall technical route of the present invention. In order to solve the problem of difficulty in obtaining small target information, the encoder backbone network is improved with the help of the self-attention convolution module, which improves the network accuracy while reducing the number of parameters, and uses the attention mechanism to enhance the ability of local information extraction and small target recognition. In order to solve the problem of simple multi-scale information fusion and poor small target recognition effect, based on the feature pyramid encoder-decoder network, the EMA attention mechanism and the hole convolution pyramid structure are integrated to obtain multi-scale information, enhance the attention to small target features, and improve the accuracy of the algorithm. In order to solve the problem of blurred contour output by depth estimation, the edge contour map extracted by the edge operator is supplemented into the model as residual information, and integrated into the loss function to improve the contour details of the depth estimation and further enhance the small target recognition effect.

[0050] Figure 2 This is the network structure diagram of the present invention. Based on LP-Net, by using edge contour information E k Replacing the Laplace residual provides the network with more local details and makes the network pay more attention to edge contours to improve the recognition effect of small target categories. k , potential feature information X k , the previous level depth residual map R k After the fusion of +1, the EMA module is used to establish the attention relationship between channels and spaces to enhance the fusion effect of different information. And by replacing the traditional convolution operation with the ACmix module, the depth residual R after fusion is enhanced. k The extraction of local information further strengthens the retention of effective information in the depth residual map. At the same time, the L-ASPP structure suitable for depth estimation is used to obtain multi-scale information, ensuring that low-scale local information is more involved, avoiding the problem of missing small target categories.

[0051] Figure 3 This paper presents the improved L-ASPP atrous spatial pyramid pooling structure used in this paper. Based on the traditional atrous spatial pyramid pooling (ASPP), a new and improved atrous spatial pyramid pooling method is proposed. This allows feature information obtained from the previous layer to be further incorporated into the calculations of the next layer, preserving deeper feature information more comprehensively. For monocular depth estimation algorithms with relatively sparse depth labels, this L-ASPP structure is more effective in improving both network performance and computational accuracy.

[0052] Figure 4The self-attention convolution (ACmix) module used in this invention. The ACmix model combines self-attention and traditional convolution through two stages and two branches. In the first stage, the input feature map is convolved with a 1×1 convolution to obtain a shared projection feature map. In the second stage, the self-attention branch calculates the projection matrix and applies the self-attention mechanism, while the traditional convolution branch processes the feature map through a fully connected layer and shift summation. To improve efficiency, the shift operation is replaced by a learnable convolution kernel to enhance the model capacity. Finally, the outputs of the two branches are added according to the weights to obtain a fused feature map.

[0053] Figure 5 This is the efficient multi-scale attention EMA module used in this paper. It fuses the outputs of different branches to generate an output of the same size as the input feature map, thereby integrating contextual information at different scales and achieving pixel-level attention. EMA uses a cross-spatial information aggregation method to model long-range dependencies while retaining precise location information, enabling the network to generate better pixel-level attention to high-level feature maps, effectively improving feature expression capabilities.

[0054] The implementation process of this embodiment is as follows:

[0055] Step 1: Use the Laplacian pyramid to build a multi-scale network structure. On the encoder side, propose and use the improved atrous spatial pyramid pooling (L-ASPP) module to obtain network features at different scales. The process is as follows:

[0056] Step 1.1 uses the Laplace pyramid structure to construct a Gaussian pyramid, and on this basis constructs a differential pyramid DOG, and finally completes the construction of the Laplace pyramid LP, and builds a monocular depth estimation network LPNet based on the Laplace pyramid. Figure 2 As shown, X k The potential deep feature information extracted by the encoder network. The deeper the network layer, the smaller the feature information graph is and the denser the feature information is.

[0057] Step 1.2 starts from the deepest fifth layer decoder. The input X5 first passes through the L-ASPP structure to fuse the context information, so that low-scale information can be more involved in feature extraction. Through more hierarchical level-by-level splicing, more and more comprehensive depth information is retained. The convolution block of the encoder has four expansion rates, namely 3, 6, 12, and 18. Features are spliced layer by layer. The number of channels is reduced through 1×1 convolution operations to reduce computational overhead. Using this method, while obtaining multi-scale information, the convolution receptive field is expanded, ensuring that low-scale local information is more involved in subsequent calculations, avoiding the problem of missing small target categories;

[0058] The calculation results of the L-ASPP structure are divided into two branches. One branch is directly output as the output depth map D5 of the fifth layer decoder after the convolution operation, which is used to generate the final depth map D1. The other branch is input into the upper layer decoder as supplementary information for subsequent operations after passing through the up-convolution module.

[0059] Step 2: Use a specific edge filter operator to process edge contour information, integrate the edge contour information into the network, and combine it with the potential feature information and the previous level depth residual information; the process is:

[0060] Step 2.1 First, grayscale the original three-channel RGB image. The formula is as follows:

[0061] Gray(i,j)=0.299*R(i,j)+0.587G(i,j)+0.114B(i,j);

[0062] Step 2.2 uses the Laplacian operator to perform edge filtering on the obtained grayscale image. The core idea is to compare the grayscale values of a single pixel with the pixels in other directions around it. When the grayscale value of the pixel is higher than the grayscale values in other directions, the grayscale value of the pixel is increased; when the grayscale value of the pixel is lower than the grayscale value in other directions, the grayscale value of the pixel is reduced. In the actual algorithm implementation process, the comparison of pixel values is achieved by calculating the gradient of its surrounding pixels. The formula is as follows:

[0063]

[0064] Step 2.3 The present invention uses the four-domain Laplacian operator, takes the target pixel as the center, and calculates the gradient in the four directions around it. The convolution template is as follows:

[0065]

[0066] The Laplacian filter operator of the image is calculated by the following formula:

[0067] H=4f(x,y)-f(x-1,y)-f(x+1,y)-f(x,y-1)-f(x,y+1)

[0068] The Laplacian operator calculation formula shows that when the grayscale of other pixels in the four directions of the pixel is the same, the calculation result is 0; when the grayscale value of the pixel is higher than the average grayscale of other pixels in the four directions, the calculation result is greater than 0; when the grayscale value of the pixel is lower than the average grayscale of other pixels in the four directions, the calculation result is less than 0. The grayscale image is filtered by the Laplacian operator to obtain the edge contour image E. k, E k Multiplying by appropriate weight parameters and adding them into the network to replace the edge residual map can enhance the recognition of target edge contours.

[0069] Step 3: Leveraging the excellent multi-channel and cross-dimensional feature extraction capabilities of the efficient multi-scale attention mechanism (EMA) module, we concatenate the edge residual map, the potential feature map, and the previous-level depth residual map. By reorganizing the channels, we deepen the connection between the three and enhance the extraction of effective information between them. The process is as follows:

[0070] Step 3.1: The output of the fourth-layer decoder X4 is concatenated with the output of the L-ASPP module and then passes through a four-layer convolution block. The concatenation is then concatenated with the output of the fifth-layer decoder D5 and the edge contour map E4 and then input into the efficient multi-scale attention EMA module.

[0071] Step 3.2EMA divides the input feature map into G sub-feature groups along the channel dimension and uses three parallel branches (two 1×1 convolution branches and one 3×3 convolution branch) to extract the attention weights of the grouped feature maps. The 1×1 branch captures cross-channel information interactions through average pooling along the spatial direction and shared 1×1 convolution to generate a channel attention map; the 3×3 branch captures global feature information through local convolution;

[0072] Step 3.3EMA further uses the cross-spatial information aggregation method Softmax to model long-range dependencies in the channel and spatial dimensions and generate a spatial attention map;

[0073] Step 3.4 Finally, EMA combines channel attention and spatial attention, and outputs an enhanced feature map of the same size as the input through nonlinear activation (Sigmoid) and feature reweighting, thereby achieving the fusion of multi-scale contextual information and pixel-level global contextual attention.

[0074] Step 4: Use the self-attention convolution module ACmix to replace the convolution module in the traditional decoder, so that it has better pixel-level attention and obtains the potential features of the pre-dimensionality-reduced depth residual map in advance, so that it can retain more effective information during dimensionality reduction. The process is as follows:

[0075] Step 4.1 Since both the self-attention module and the convolution module can be divided into two stages: feature learning module and feature aggregation, although there are differences in the learning modes of the two stages, both use 1×1 convolution to project features into a deeper space in terms of execution methods, so these convolution operations can be merged to reduce computational overhead. Based on this idea, the present invention uses the ACmix module to replace the traditional convolution module used in the decoder. In the first stage, three 1×1 convolution operations are performed on the input feature map to obtain N groups of projected feature maps C / N×H×W shared by the two paths. For the self-attention path, N represents the number of perception heads of the multi-head attention. In the second stage, the three groups of shared projection features are input into their respective branches. For the self-attention path, the projection matrices W of the query, key, and value are calculated respectively. q (l) , W k (l) , W v (l) , and then calculate the output feature map based on the self-attention mechanism according to the following formula:

[0076]

[0077] For the traditional convolution branch with kernel k, the input 3 N groups of feature maps C / N×H×W are passed through a lightweight fully connected layer to obtain k 2 The feature maps are C / N×H×W, which is equivalent to performing a channel dimensionality reduction operation on the input feature map, which can reduce some computational overhead. The feature maps are then shifted and summed, and finally the number of output channels is restored through convolution. Finally, the output maps obtained from the two paths are added together with a certain weight to obtain the output feature map that combines the self-attention mechanism with traditional convolution. Its expression is as follows:

[0078] F out =αF att +βF conv ;

[0079] Step 4.2 replaces the unidirectional shift calculation in ACmix with a 3×3 convolution with only one bit. The convolution kernel is shown as follows:

[0080]

[0081] For any output feature tensor g i,j You can get:

[0082]

[0083] Corresponding shift operations in different directions can be replaced by this depthwise convolution with a fixed kernel. The convolution kernels corresponding to the shift operations in the nine directions are grouped together. This modification makes ACmix more efficient when performing shift operations. This fixed kernel convolution is changed to a convolution kernel with learnable weights, with the offset kernel as the initial weight parameter. In this way, the capacity of the model is increased while retaining the shift and sum operations. The final output channel dimension size of the multiple sets of convolution kernel matching algorithms is also set;

[0084] Step 4.3 The feature map processed by the ACmix module is then convolved and concatenated with the feature map of the previous layer to obtain the depth map output D of this layer of decoder. k The operations in the subsequent layer decoders are similar to those in the fourth layer, and the depth map D output by each layer is finally converted into k The final depth prediction image D1 is obtained by splicing.

[0085] Step 5: Design a monocular depth estimation loss function to increase the focus of network depth information, gradient information, and edge contour information. The process is as follows:

[0086] Step 5.1 Due to the sparse distribution of depth label values, directly using the difference between the depth estimate and the true depth value as the loss function may cause an imbalance problem where the pixel does not have a depth label. Therefore, the logarithm of the estimated value and the true value is taken as the difference to minimize the impact of the imbalance. Therefore, the depth value loss function is defined as:

[0087]

[0088] where y i Predict the depth value for pixel i, y i * is the actual depth value of pixel i. There are n pixels in total, and λ is the balancing weight parameter, which is generally set to 0.85;

[0089] Step 5.2 The gradient loss function can make the network pay more attention to the gradient changes in the depth map, making the target depth gradient changes smoother. For the sparse distribution of depth label values, the interpolation method is used to obtain the depth gradient value. Therefore, the gradient loss function is defined as:

[0090]

[0091] where g h (y i ) represents the predicted gradient value of pixel i in the h direction, i.e., the height direction, g w (y i ) represents the predicted depth gradient value of pixel i in the w direction, i.e., the width direction. Similarly, g h (y i *), gw (y i *) represents the true depth gradient value of pixel i in the h direction and w direction respectively;

[0092] Step 5.3 Use the Sobel edge detection operator to extract the edge contour information e i Assigning the logarithmic difference between the estimated value and the true value as a weight to strengthen the algorithm's attention to the contour boundary can further amplify the depth error of the depth prediction image at the edge contour, thereby achieving the effect of further learning the contour boundary. Therefore, the edge information loss function that emphasizes contour details is defined as:

[0093]

[0094] where e i Represents edge contour information and also assigns weights. The extraction method is the edge filter operator. Referring to the definition of the depth loss function, weights are also assigned to the logarithmic difference between the estimated value and the true value to reduce the impact of imbalance.

[0095] Step 5.4: Sum the above depth loss function, gradient loss function, and edge information loss function to obtain the monocular depth estimation loss function Lt(y,y*) actually used in the algorithm training process. Therefore, the monocular depth estimation loss function actually used in the algorithm training process is defined as:

[0096]

[0097] Where α and β are balancing weights, typically set to 10 and 0.1. μ is the balancing weight corresponding to the edge information loss function. Experiments have shown that a value of 0.01 works well. Due to the sparsity of the ground-truth depth values in the KITTI dataset, using gradient loss for calculations at the beginning of training is unstable. Therefore, as defined above, the gradient loss function is not added to the calculations until after 25 epochs.

[0098] Step 5.5: Train the model using the defined loss function strategy in the KITTI dataset. Use the trained model to train the test set and obtain the depth value information predicted by the network.

[0099] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. An edge contour enhanced EA-LPNet monocular depth estimation method, characterized in that: The method comprises the following steps: Step 1: Use the Laplacian pyramid to build a multi-scale network structure. On the encoder side, propose and use the improved atrous spatial pyramid pooling (L-ASPP) module to obtain network features at different scales. Step 2: Use a specific edge filter operator to process edge contour information, integrate the edge contour information into the network, and combine it with the potential feature information and the previous level depth residual information; Step 3: Use the efficient multi-scale attention mechanism (EMA) module to concatenate the edge residual map, the potential feature map, and the previous level depth residual map to enhance the extraction of effective information between the three. Step 4: Use the self-attention convolution module ACmix to replace the convolution module in the traditional decoder to obtain the potential features in the pre-dimensionality-reduced depth residual map in advance and retain more effective information; Step 5: Design a monocular depth estimation loss function to increase the network's attention to depth information, gradient information, and edge contour information.

2. The edge contour enhanced EA-LPNet monocular depth estimation method according to claim 1, characterized in that The process of step one is: Step 1.1 uses the Laplace pyramid structure to build a monocular depth estimation network LPNet based on the Laplace pyramid; Step 1.2 proposes an improved dilated spatial pyramid pooling method: the L-ASPP structure is used on the encoder side to process the input color image through convolution blocks with different expansion rates to obtain network features at different scales. This allows the feature information obtained in the previous layer to further participate in the calculation of the next layer, and more comprehensively retains deep feature information. Step 1.3 The L-ASPP output is divided into two branches, which serve as the deep feature information of the current layer decoder and the supplementary information for subsequent operations in the previous layer decoder respectively.

3. The EA-LPNet monocular depth estimation method with edge contour enhancement according to claim 1 or 2, wherein: The process of step 2 is: Step 2.1 grayscale the original three-channel RGB image; Step 2.2 uses the Laplacian operator to perform edge filtering on the obtained grayscale image; Step 2.3 Obtain the edge contour map E after edge filtering K Multiply it by appropriate weight parameters and add them to the network to enhance the contour recognition ability.

4. The EA-LPNet monocular depth estimation method with edge contour enhancement according to claim 1 or 2, wherein: The process of step three is: Step 3.1: Concatenate the edge residual map, potential feature map, and previous level depth residual map and input them into the EMA module. Step 3.2 Divide the sub-feature groups and input the feature graph into the three parallel branches of EMA; Step 3.3 Generate a spatial attention map through information aggregation method; Step 3.4 combines channel attention with spatial attention and outputs an enhanced feature map.

5. The EA-LPNet monocular depth estimation method with edge contour enhancement according to claim 1 or 2, wherein: The process of step 4 is: Step 4.1 performs three 1×1 convolution operations on the input feature map to obtain the projected N sets of feature maps C / N×H×W shared by the self-attention module and the convolution module, and inputs them into the corresponding path; Step 4.2 performs unidirectional displacement calculation on the output feature tensor; Step 4.3 The feature map processed by the ACmix module is concatenated with the feature map of the previous layer to obtain the output D of this layer. k , the depth map D output by each layer k The final depth prediction image D1 is obtained by splicing.

6. The edge contour enhanced EA-LPNet monocular depth estimation method according to claim 1 or 2, characterized in that The process of step five is: Step 5.1 defines the depth loss function, taking the difference between the estimated value and the true value to reduce the impact of imbalance; Step 5.2 defines the gradient loss function. For the sparse distribution problem of depth label values, the interpolation method is used to obtain the depth gradient value; Step 5.3 adds an edge information loss function to enhance contour details. The edge contour information is used as a weight to enhance the network's attention to the target edge. The edge contour information extracted by the Sobel edge detection operator is used as a weight to enhance the method's attention to the contour boundary. Step 5.4: Sum the above depth loss function, gradient loss function, and edge information loss function to obtain the monocular depth estimation loss function actually used in the method training process; Step 5.5 Obtain the KITTI dataset for model training and testing.

Images