Intelligent visual detection and identification method for river water surface target

Abstract

The invention discloses an intelligent visual detection and identification method for a riverway water surface target, and relates to the technical field of deep learning. Firstly, the expression ability of a model is enhanced through a structural re-parameterization feature extraction module; then, characteristic information is promoted to flow in a cross-layer mode through a characteristic cross-layer fusion and reconstruction module, characteristic expression of the targets is enhanced, space channel reconstruction is carried out to remove redundant information, and the detection capacity of the targets of different scales under the complex background is enhanced; and finally, improving the loss function to enhance the model recall rate. After the target detection algorithm is compared with a mainstream target detection algorithm, the applicability of the algorithm in a river water surface target detection task is proved, and the generalization ability of the algorithm in different environments is proved through experiments; experiments show that the average recognition rate of 16 types of objects on an actual data set is superior to that of an existing method, and an effective solution is provided for intelligent supervision of a river channel.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method for intelligent visual detection and recognition of river surface targets. Background Art

[0002] With the rapid development of urbanization, river environments have been severely damaged, and effective river monitoring and management methods are urgently needed. River surface target detection plays a vital role in river monitoring systems and forms the foundation of river violation detection technology. Traditional river pollution monitoring methods typically rely on manual inspections or real-time observations at fixed monitoring points. These methods suffer from limited coverage, low efficiency, and high costs. However, using drone aerial photography combined with target detection technology to monitor rivers significantly improves monitoring efficiency and accuracy while reducing costs.

[0003] In drone aerial videos, water surface targets often face problems such as low pixels and unclear feature contours, and are affected by complex environmental factors in the river (such as weather and light intensity). The scales of different types of water surface targets vary greatly. The above problems will lead to unsatisfactory model performance. Therefore, it is extremely important to develop an effective detection method that can adapt to the complex environment of the river and targets of different scales on the water surface.

[0004] Existing deep learning-based object detection algorithms are mainly divided into two-stage object detection models and single-stage object detection models. Classic two-stage object detection models, such as R-CNN, achieve object detection through candidate region generation and feature extraction. Fast R-CNN further improves the detection accuracy of deformed objects by pooling regions of interest (ROIs). Faster R-CNN introduces a region proposal network to increase detection speed. An improved Faster R-CNN model proposed in the prior art uses feature fusion and anchor optimization to improve detection accuracy.

[0005] In recent years, the performance of single-stage object detection models has gradually improved. These models treat object detection as a regression problem, eliminating the need to generate candidate regions. They use the entire image as input for computation, offering the advantages of fast detection speed and wide adaptability.

[0006] However, existing detection methods struggle with large variations in size due to insufficient feature fusion and limited receptive field size, limiting their ability to detect diverse targets. Furthermore, the complex background environment of rivers, characterized by dimly lit cloudy skies, rainy and foggy conditions, can obscure target features, making detection difficult. Small targets are particularly challenging to identify. While a multi-branch structure can reduce information loss during feature extraction within the backbone, achieving good detection accuracy, it also reduces model detection efficiency. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention provides a method for intelligent visual detection and recognition of river surface targets, comprising the following steps:

[0008] S1. Acquire river surface images and perform image preprocessing operations;

[0009] S2. Construct a river surface target detection network, which includes a backbone network, a feature cross-layer fusion and reconstruction module, and a prediction head. The backbone network includes a structure reparameterization feature extraction module.

[0010] S3, extract multi-level feature maps from the input image through the backbone network;

[0011] S4, enhance the feature map through feature cross-layer fusion and reconstruction module;

[0012] S5, the prediction head generates outputs for object localization and classification based on the enhanced feature maps;

[0013] S6. Post-process the output using non-maximum suppression to remove low-scoring predictions.

[0014] The technical solution further defined in the present invention is:

[0015] Furthermore, in step S2, the structure reparameterization feature extraction module adopts a multi-branch structure during training; during the inference process, the multi-branch structure is reparameterized into a single-path structure; the structure reparameterization feature extraction module downsamples the input features through a convolution with a convolution kernel size of 3×3 and a step size of 2, reducing the length and width of the feature map to half of the input.

[0016] As described above, in a method for intelligent visual detection and recognition of river surface targets, in step S3, the feature map is split into multiple parts for processing, and feature information is extracted through multiple SR modules. During the training process, the SR module adopts a multi-branch structure, including a 3×3 convolutional layer branch, a 1×1 convolutional layer branch, and a direct mapping of the front layer output; in the structural reparameterization process, the 3×3 convolutional layer and the 1×1 convolutional layer in the multi-branch structure are first batch normalized, and then the normalized 1×1 convolutional kernel is converted into a 3×3 convolutional kernel with only the center having a valid weight and the rest of the weights being 0; the direct mapping of the front layer output is regarded as a 3×3 convolutional kernel with a center weight of 1 and the rest of the weights being 0; finally, the 1×1 convolutional layer branch and the front layer output are respectively added to the convolution kernel weights of the normalized 3×3 convolutional layer; during the inference process, it is equivalent to extracting features using a 3×3 convolution kernel.

[0017] As described above, in the intelligent visual detection and recognition method for river surface targets, in step S4, after the backbone network extracts feature maps of different levels, the feature cross-layer fusion and reconstruction module performs weighted fusion on the original input of the same level and the feature maps of different levels, and then reconstructs the spatial information and channel information through the spatial channel reconstruction unit.

[0018] As described above, in the method for intelligent visual detection and recognition of river surface targets, in step S4, the feature cross-layer fusion and reconstruction module further deepens the feature information of the feature map extracted by the backbone network through a convolution layer with a convolution kernel size of 1×1 and a step size of 1. The upsampled information of the high-level feature map, the downsampled information of the low-level feature map, and the deepened information are fused according to different weights, as shown in the following formula:

[0019]

[0020] Among them, F i-1 、F i and F i+1 They represent the features after downsampling of the low-level feature map, the features of this level, and the features after upsampling of the high-level feature map, respectively. i-1 、w i and w i+1 They represent the corresponding weights of each level that can be trained; Concat represents the operation of splicing features of different scales; the fused feature F c After being processed by the spatial channel reconstruction unit, the result F is output by refining the weights and separating the reconstruction. out .

[0021] As described above, in the intelligent visual detection and recognition method for river surface targets, the spatial channel reconstruction unit includes a Weight unit, a convolution layer, a Concat layer, a Transl module, a Trans2 module, and a separation and reconstruction layer.

[0022] As mentioned above, the intelligent visual detection and recognition method for river surface targets, the Weight unit algorithm is shown as follows:

[0023]

[0024] Among them, W represents weight information; γ i ∈R N , γ j ∈R N Represents the input feature F in The parameters learned by group normalization, R N represents an N-dimensional real vector; i and j are both positive integers from 1 to N; μ and σ represent F in The mean and standard deviation of ;ε represents a normal number; represents element-wise multiplication; γ and β represent trainable affine transformations.

[0025] As described above, the process of the spatial channel reconstruction unit of the intelligent visual detection and recognition method for river surface targets includes the following steps:

[0026] S4.1, weight information W is gated with a threshold of 0.5 to obtain the informative feature weight W1 and the non-informative weight W2;

[0027] S4.2, input feature Fi n The feature map after point convolution is multiplied element-wise with the informative feature weight W1 and the non-informative weight W2 to obtain the informative feature F1 and the less informative feature F2;

[0028] S4.3. Split features F1 and F2 along the channels respectively, and obtain four channels with only half of the original feature F. 11 、F 12 、F 21 and F 22 , and cross-add them item by item to get features F3 and F4, and then connect features F3 and F4 along the channel dimension to restore them to the same size as the input feature Fi n Same number of channels;

[0029] S4.4. Re-split the feature map into half the number of channels along the channel, and compress the number of channels of the feature map through point convolution to obtain features F5 and F6 respectively;

[0030] S4.5. The features are processed by group convolution and point-by-point convolution in the Trans1 module and then added item by item to obtain feature F7, as shown in the following formula: Among them, GWC and PWC represent the learnable weight matrices, represents element-wise summation;

[0031] S4.6. The point-wise convolution processed features are fused with the unprocessed features through the Trans2 module to obtain feature F8, as shown in the following formula: F8 = PWC(F6)UF6;

[0032] S4.7, connect features F7 and F8 along the channel dimension, and output the result F through weight refinement and separation reconstruction. out .

[0033] As described above, in the method for intelligent visual detection and recognition of river surface targets, in step S6, the bounding box similarity comparison metric MPDIoU based on minimum point distance is used as the loss function for the model bounding box regression.

[0034] As mentioned above, in the intelligent visual detection and recognition method for river surface targets, MPDIoU is calculated based on the minimum point distance between the predicted bounding box and the true bounding box, as shown in the following formula:

[0035]

[0036] Among them, w and h represent the width and height of the input image respectively. and Represent the coordinates of the upper left corner and lower right corner of the predicted bounding box, and They represent the coordinates of the upper left corner and lower right corner of the true bounding box respectively, and IoU represents the intersection over union ratio of the predicted bounding box and the true bounding box.

[0037] The beneficial effects of the present invention are:

[0038] (1) In this paper, a structural reparameterization feature extraction module (SRFE) is proposed. During the training process, a multi-branch structure is used to obtain sufficient information. During the inference process, it is reparameterized into a single-path structure, thereby maintaining high efficiency and low memory consumption, and effectively improving the backbone feature extraction capability.

[0039] (2) In this invention, a feature cross-layer fusion and reconstruction module FCFR is proposed. FCFR fuses the original input with features from different layers to expand the receptive field. Then, through feature reconstruction, it improves the performance of identifying targets of different sizes and eliminates the influence of the river background environment. It can not only strengthen the association of features at different levels and improve the recognition ability of targets of different sizes, but also reduce the calculation of redundant features of the dynamic background of the river environment through spatial channel reconstruction, thereby improving information utilization.

[0040] (3) In this invention, the loss function is improved to improve the ability to detect dense targets and the convergence speed. A large-scale unmanned aerial vehicle river surface target detection dataset (UARODD) is also established. This dataset collects 16 types of marine targets and marks 9,500 images, which has a certain driving effect on the research of multi-target detection on the river surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0042] Figure 2 Schematic diagram of the structure of the river surface target detection network in an embodiment of the present invention;

[0043] Figure 3 Schematic diagram of the structure of the structural re-parameterization feature extraction module in an embodiment of the present invention;

[0044] Figure 4 Schematic diagram of the structure of a spatial channel reconstruction unit in an embodiment of the present invention;

[0045] Figure 5 Schematic diagram of bounding box regression when the predicted bounding box and the true bounding box have the same intersection-over-union ratio but different heights and widths in an embodiment of the present invention, where (a) is a schematic diagram of the predicted bounding box being twice the size of the true bounding box, and (b) is a schematic diagram of the predicted bounding box being half the size of the true bounding box. DETAILED DESCRIPTION

[0046] This embodiment provides a method for intelligent visual detection and recognition of river surface targets, such as Figure 1 As shown, the following steps are included:

[0047] S1. Obtain river surface images and perform image preprocessing operations such as image enhancement and denoising.

[0048] S2, build a river surface target detection network, such as Figure 2 As shown in FIG, the network includes a backbone network, a feature cross-layer fusion and reconstruction module (FCFR) and a prediction head. The backbone network includes a structured reparameterized feature extraction module (SRFE).

[0049] Structural reparameterization feature extraction is a feature extraction method that ensures the efficiency of backbone feature extraction while improving the accuracy of backbone feature extraction. During the training process, SRFE adopts a multi-branch structure to extract more image feature information and enhance the expressive ability of the model; during the inference process, the multi-branch structure is reparameterized into a single-path structure, obtaining the advantages of high efficiency and low memory consumption of the single-path model.

[0050] S3. For an RGB image with an input size of 640×640×3, a backbone network is used to extract multi-level feature maps to fully preserve the spatial and semantic information of the target.

[0051] like Figure 3 As shown in the figure, SRFE downsamples the input features through convolution with a kernel size of 3×3 and a stride of 2, reducing the length and width of the feature map to half of the input; point convolution can increase the depth of the network; the feature map is split into multiple parts for processing, and feature information is extracted through multiple structured reparameterized modules (SR). During the training process, the SR module adopts a multi-branch structure, including 3×3 convolution layer branches, 1×1 convolution layer branches, and direct mapping of the previous layer output.

[0052] In the process of structural reparameterization, the 3×3 convolutional layer and 1×1 convolutional layer in the multi-branch structure are first batch normalized (BN), and then the normalized 1×1 convolution kernel is converted into a 3×3 convolution kernel with only valid weights in the center and the rest of the weights are 0; the direct mapping of the front layer output is regarded as a 3×3 convolution kernel with a center weight of 1 and the rest of the weights are 0; finally, the 1×1 convolutional layer branch and the front layer output are respectively added to the normalized 3×3 convolution kernel weights of the convolution layer; in the inference process, it is equivalent to extracting features with a 3×3 convolution kernel; compared with the use of normally trained 3×3 convolution kernels, SR has the excellent feature extraction capability of the multi-branch structure while also having the efficient detection speed of the single-path structure.

[0053] S4. Enhance the feature map through cross-layer feature fusion and reconstruction module, integrate feature maps of different scales, improve the perception ability of multi-scale targets, and reconstruct the fused feature map through spatial channels to further enhance the key area features.

[0054] After the backbone network extracts feature maps at different levels, the feature cross-layer fusion and reconstruction module performs weighted fusion on the original input at the same level and the feature maps at different levels. This promotes the contextual propagation of feature information, obtains a larger receptive field and richer semantic information, and is conducive to detecting aerial photography targets of rivers at different scales. The Spatial-Channel Reconstruction Unit (SCRU) then reconstructs spatial information such as shape and structure and channel information such as color and depth, reducing the redundant information generated after fusion. This module performs well in the complex background of constantly changing river lighting and weather.

[0055] like Figure 2 As shown in Figure 2, FCFR further deepens the feature information of the feature map extracted by the backbone network through a convolution layer with a convolution kernel size of 1×1 and a step size of 1, and then fuses the upsampled information of the high-level feature map, the downsampled information of the low-level feature map, and the deepened information according to different weights, as shown in the following formula:

[0056]

[0057] Among them, F i-1 、F i and F i+1 They represent the features after downsampling of the low-level feature map, the features of this level, and the features after upsampling of the high-level feature map, respectively. i-1 、w i and w i+1 They represent the corresponding weights of each level that can be trained; Concat represents the operation of splicing features of different scales; the fused feature F cAfter SCRU processing, the background redundant information is reduced and the result F is output by refining the weights and separating and reconstructing. out .

[0058] like Figure 4 As shown in the figure, the spatial channel reconstruction unit includes the Weight unit, convolution layer, Concat layer, Transl module, Trans2 module and separation reconstruction layer; the Weight unit algorithm is shown in the following formula:

[0059]

[0060] Among them, W represents weight information; γ i ∈R N , γ j ∈R N Represents the parameters learned by group normalization of the input feature Fin, R N represents an N-dimensional real vector; i and j are both positive integers from 1 to N; μ and σ represent F in The mean and standard deviation of ;ε represents a small positive constant to ensure the stability of division; represents element-wise multiplication; γ and β represent trainable affine transformations.

[0061] The SCRU process reduces redundant feature information and improves information utilization during detection. The process includes the following steps:

[0062] S4.1, weight information W is gated with a threshold of 0.5 to obtain the informative feature weight W1 and the non-informative weight W2;

[0063] S4.2, input feature F in The feature map after point convolution is multiplied element-wise with the informative feature weight W1 and the non-informative weight W2 to obtain the informative feature F1 and the less informative feature F2;

[0064] S4.3. Split features F1 and F2 along the channels respectively, and obtain four channels with only half of the original feature F. 11 、F 12 、F 21 and F 22 , and cross-add them item by item to get features F3 and F4, and then connect features F3 and F4 along the channel dimension to restore them to the same size as the input feature F in Same number of channels;

[0065] S4.4. Re-split the feature map into half the number of channels along the channel, and compress the number of channels of the feature map through point convolution to obtain features F5 and F6 respectively;

[0066] S4.5. The features are processed by the Transl module using Group-Wise Convolution (GWC) and Point-Wise Convolution (PWC) respectively, and then added item by item to obtain feature F7, as shown in the following formula: Among them, GWC and PWC represent the learnable weight matrices, represents element-wise summation;

[0067] S4.6. The point-wise convolution processed features are fused with the unprocessed features through the Trans2 module to obtain feature F8, as shown in the following formula: F8 = PWC(F6)UF6;

[0068] S4.7, connect features F7 and F8 along the channel dimension, and output the result F through weight refinement and separation reconstruction. out .

[0069] S5. With the help of the detection head, the water surface targets are identified and located, and the type of target and its specific location in the image are output, thereby achieving accurate detection and intelligent perception of river surface targets.

[0070] S6. Post-process the output using non-maximum suppression to remove low-scoring predictions.

[0071] In drone-photographed river surface images, there are a large number of densely packed samples, with multiple targets occluding each other. This reduces target feature information and reduces model detection accuracy. To effectively alleviate this problem and improve detection accuracy, this embodiment uses a bounding box similarity comparison metric based on minimum point distance (MPDIoU) as the loss function for the model's bounding box regression.

[0072] When the predicted bounding box and the true bounding box have the same intersection-over-union ratio, but different heights and widths, such as Figure 5 As shown in the figure, the green box represents the true bounding box and the red box represents the predicted bounding box; the traditional bounding box regression loss function calculates the same value in these two cases, which is regarded as equivalent during the optimization process, hindering the correct adjustment of the model parameters and reducing the convergence speed and accuracy.

[0073] MPDIoU is calculated based on the minimum point distance between the predicted bounding box and the true bounding box, as shown in the following formula:

[0074]

[0075] Among them, w and h represent the width and height of the input image respectively. and Represent the coordinates of the upper left corner and lower right corner of the predicted bounding box, and They represent the coordinates of the upper left corner and lower right corner of the true bounding box respectively, and IoU represents the intersection over union ratio of the predicted bounding box and the true bounding box.

[0076] MPDIoU comprehensively addresses the relevant factors of the bounding box regression problem, such as overlapping areas, changes in height and width, and the distance between center points. It can not only accurately distinguish such situations, but also effectively avoid the problem of gradient vanishing in the process of dense target recognition, thereby improving the robustness of the model.

[0077] Handling multi-scale targets in complex backgrounds has always been a key difficulty in river target detection. This embodiment first enhances the expressiveness of the model through a structural reparameterization feature extraction module; then, through a feature cross-layer fusion and reconstruction module, feature information is promoted to flow across layers, thereby enhancing the feature expression of scale targets; spatial channel reconstruction is then performed to remove redundant information, thereby enhancing the detection capability of targets of different scales in complex backgrounds; finally, the model recall rate is enhanced by improving the loss function; after comparing the target detection algorithm of the present invention with mainstream target detection algorithms, the applicability of the algorithm to the task of river surface target detection is demonstrated, and its generalization ability in different environments is verified through experiments; experiments show that the average recognition rate map@0.5 of 16 types of objects in the actual data set of the present algorithm reaches 79.6%, which is better than the existing methods and provides an effective solution for intelligent river supervision.

[0078] In addition to the above embodiments, the present invention may also have other implementations. Any technical solution formed by equivalent replacement or equivalent transformation falls within the scope of protection required by the present invention.

Claims

1. A method for intelligent visual detection and recognition of river surface targets, characterized by: The following steps are involved: S1. Acquire river surface images and perform image preprocessing operations; S2. Construct a river surface target detection network, which includes a backbone network, a feature cross-layer fusion and reconstruction module, and a prediction head. The backbone network includes a structure reparameterization feature extraction module. S3, extract multi-level feature maps from the input image through the backbone network; S4, enhance the feature map through feature cross-layer fusion and reconstruction module; S5, the prediction head generates outputs for object localization and classification based on the enhanced feature maps; S6. Post-process the output using non-maximum suppression to remove low-scoring predictions.

2. The method for intelligent visual detection and recognition of river surface targets according to claim 1, characterized in that: In step S2, the structure reparameterization feature extraction module adopts a multi-branch structure during the training process; during the inference process, the multi-branch structure is reparameterized into a single-path structure; the structure reparameterization feature extraction module downsamples the input features through a convolution with a convolution kernel size of 3×3 and a step size of 2, reducing the length and width of the feature map to half of the input.

3. The method for intelligent visual detection and recognition of river surface targets according to claim 1, characterized in that: In the step S3, the feature map is split into multiple parts for processing, and feature information is extracted through multiple SR modules. During the training process, the SR module adopts a multi-branch structure, including a 3×3 convolution layer branch, a 1×1 convolution layer branch, and a direct mapping of the front layer output; in the structural reparameterization process, the 3×3 convolution layer and the 1×1 convolution layer in the multi-branch structure are first batch normalized, and then the normalized 1×1 convolution kernel is converted into a 3×3 convolution kernel with only a valid weight at the center and the remaining weights being 0; the direct mapping of the front layer output is regarded as a 3×3 convolution kernel with a center weight of 1 and the remaining weights being 0; finally, the 1×1 convolution layer branch and the front layer output are respectively added to the convolution kernel weight of the normalized 3×3 convolution layer; during the inference process, it is equivalent to extracting features with a 3×3 convolution kernel.

4. The method for intelligent visual detection and recognition of river surface targets according to claim 1, characterized in that: In step S4, after the backbone network extracts feature maps of different levels, the feature cross-layer fusion and reconstruction module performs weighted fusion on the original input of the same level and the feature maps of different levels, and then reconstructs the spatial information and channel information through the spatial channel reconstruction unit.

5. The method for intelligent visual detection and recognition of river surface targets according to claim 1, characterized in that: In step S4, the feature cross-layer fusion and reconstruction module further deepens the feature information of the feature map extracted by the backbone network through a convolution layer with a convolution kernel size of 1×1 and a step size of 1, and then fuses the upsampled information of the high-level feature map, the downsampled information of the low-level feature map, and the deepened information according to different weights, as shown in the following formula: Among them, F i-1 、F i and F i+1 They represent the features after downsampling of the low-level feature map, the features of this level, and the features after upsampling of the high-level feature map, respectively. i-1 、w i and w i+1 They represent the corresponding weights of each level that can be trained; Concat represents the operation of splicing features of different scales; the fused feature F c After being processed by the spatial channel reconstruction unit, the result F is output by refining the weights and separating the reconstruction. out .

6. The method for intelligent visual detection and recognition of river surface targets according to claim 4, characterized in that: The spatial channel reconstruction unit includes a Weight unit, a convolution layer, a Concat layer, a Trans 1 module, a Trans 2 module and a separation reconstruction layer.

7. The method for intelligent visual detection and recognition of river surface targets according to claim 6, characterized in that: The Weight unit algorithm is shown below: Among them, W represents weight information; γ i ∈R N , γ j ∈R N Represents the input feature F in The parameters learned by group normalization, R N represents an N-dimensional real vector; i and j are both positive integers from 1 to N; μ and σ represent F in The mean and standard deviation of ;ε represents a normal number; represents element-wise multiplication; γ and β represent trainable affine transformations.

8. The method for intelligent visual detection and recognition of river surface targets according to claim 7, characterized in that: The process of the spatial channel reconstruction unit specifically includes the following steps: S4.1, weight information W is gated with a threshold of 0.5 to obtain the informative feature weight W and the non-informative weight W2; S4.2, input feature F in The feature map after point convolution is multiplied element-wise with the informative feature weight W1 and the non-informative weight W2 to obtain the informative feature F1 and the less informative feature F2; S4.

3. Split features F1 and F2 along the channels respectively, and obtain four channels with only half of the original feature F. 11 、F 12 、F 21 and F 22 , and cross-add them item by item to get features F3 and F4, and then connect features F3 and F4 along the channel dimension to restore them to the same size as the input feature F in Same number of channels; S4.

4. Re-split the feature map into half the number of channels along the channel, and compress the number of channels of the feature map through point convolution to obtain features F5 and F6 respectively; s4.

5. The features are processed by group convolution and point-by-point convolution in the Trans1 module and then added item by item to obtain feature F7, as shown in the following formula: Among them, GWC and PWC represent the learnable weight matrices, represents element-wise summation; S4.

6. The point-wise convolution processed features are fused with the unprocessed features through the Trans2 module to obtain feature F8, as shown in the following formula: F8 = PWC(F6)UF6; S4.7, connect features F7 and F8 along the channel dimension, and output the result F through weight refinement and separation reconstruction. out .

9. The method for intelligent visual detection and recognition of river surface targets according to claim 1, characterized in that: In step S6, the bounding box similarity comparison metric MPDIoU based on minimum point distance is used as the loss function of the model bounding box regression.

10. The method for intelligent visual detection and recognition of river surface targets according to claim 9, characterized in that: The MPDIoU is calculated based on the minimum point distance between the predicted bounding box and the true bounding box, as shown in the following formula: Among them, w and h represent the width and height of the input image respectively. and Represent the coordinates of the upper left corner and lower right corner of the predicted bounding box, and They represent the coordinates of the upper left corner and lower right corner of the true bounding box respectively, and IoU represents the intersection over union ratio of the predicted bounding box and the true bounding box.

Images