A Method and System for Occluded Ship Positioning and Recognition for Shore Monitoring

Abstract

The present invention discloses a method and system for occluded ship positioning and recognition for shore monitoring, which relates to the field of ship positioning. Obtain and preprocess the reference dataset images; introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing on the features through wavelet transform convolution. At the same time, add an activation function and a residual connection to each block to complete the construction of the occluded ship positioning and recognition network; use the reference dataset images to train the occluded ship positioning and recognition network; input the test pictures into the trained occluded ship positioning and recognition network to obtain the recognition results. By designing a backbone network with stronger feature extraction ability, combining the information in the time domain and the frequency domain, and using the attention mechanism, the present invention enhances the network feature extraction ability and improves the recognition and positioning ability of the model for occluded ships.

Description

Technical Field

[0001] The present invention relates to the field of ship positioning, and more specifically, to a method and system for identifying and positioning occluded ships for shore monitoring. Background Art

[0002] With the rapid development of computer vision and deep learning technologies, ship target detection technology based on artificial intelligence has provided great convenience for real-time monitoring of ship positions and states, identifying potential threats, and improving ship scheduling efficiency. However, in some busy ports or important shipping routes, due to the dense traffic of ships, the distances between ships or between ships and shore objects are often too close, and occlusion phenomena are very likely to occur.

[0003] When traditional target detection networks deal with such complex scenarios, they often lose or confuse target feature information due to insufficient feature extraction ability and poor feature processing ability, making it difficult for the model to accurately distinguish the boundaries and categories of targets when facing occluded objects, resulting in performance degradation. Therefore, how to solve the above technical problems urgently requires further research by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for identifying and positioning occluded ships for shore monitoring. By designing a backbone network with stronger feature extraction ability, combining information in the time domain and frequency domain, and using an attention mechanism, the feature extraction ability of the network is enhanced, and the recognition and positioning ability of the model for occluded ships is improved.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for identifying and positioning occluded ships for shore monitoring includes the following steps:

[0007] Obtain and preprocess the reference dataset images;

[0008] Introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing on the features through wavelet transform convolution. At the same time, add an activation function and a residual connection to each block to complete the construction of the occluded ship positioning and recognition network;

[0009] Train the occluded ship positioning and recognition network using the reference dataset images;

[0010] Input the test pictures into the trained occluded ship positioning and recognition network to obtain the recognition results.

[0011] Optionally, the normalization layer is used to ensure the stability of the feature distribution, alleviate the problem of gradient vanishing or explosion, thereby improving the training stability of the network; the GELU activation function improves the fitting ability of the network to the occluded environment through its smooth non-linear mapping.

[0012] Optionally, the attention mechanism is introduced into the convolutional neural network to add spatial attention to the model in the spatial domain. Starting from the importance of features within the receptive field, a dynamic weight adjustment mechanism is used to effectively capture target features at different positions within the receptive field under occlusion or complex backgrounds.

[0013] Optionally, the residual connection retains the low-level features during feature fusion.

[0014] Optionally, the loss function adopts the bounding box loss and the classification loss.

[0015] Optionally, the CIoU loss function is adopted for the bounding box loss, and the calculation formula is as follows:

[0016] ;

[0017] where, respectively represent the center points of the predicted box and the ground truth box, represents the Euclidean distance between the center points of the predicted box and the ground truth box, represents the diagonal length of the smallest circumscribed rectangle that can simultaneously contain the predicted box and the ground truth box, is the weight function, used to measure the consistency of the aspect ratio.

[0018] Optionally, the BCE loss function is adopted for the classification loss, and the calculation formula is as follows:

[0019] ;

[0020] where, N represents the batch size, represents the ground truth value, represents the predicted value. When the ground truth label is 1, the loss function becomes . This means that if the probability predicted by the model is closer to 1, the loss is smaller; if the predicted probability is closer to 0, the loss is larger.

[0021] When the ground truth label is 0, the loss function becomes . This means that if the probability predicted by the model is closer to 0, the loss is smaller; if the predicted probability is closer to 1, the loss is larger.

[0022] An occluded ship positioning and recognition system for shore monitoring, comprising:

[0023] Dataset Image Acquisition Module: used to acquire and preprocess the benchmark dataset images;

[0024] Occluded Ship Localization and Recognition Network Construction Module: used to introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing on the features through wavelet transform convolution. At the same time, an activation function and a residual connection are added to each block to complete the construction of the occluded ship localization and recognition network;

[0025] Occluded Ship Localization and Recognition Network Training Module: used to train the occluded ship localization and recognition network using the benchmark dataset images;

[0026] Recognition Result Output Module: used to input the test images into the trained occluded ship localization and recognition network to obtain the recognition results.

[0027] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a method and system for occluded ship localization and recognition for shore monitoring, which pays more attention to the establishment of a lightweight backbone network, fully considers high-order semantic information and low-order feature information, and utilizes the feature information in the time domain and frequency domain; at the same time, it fully considers feature mining and feature diversity, and can mine more discriminative feature regions. Description of the Drawings

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0029] Figure 1 It is a schematic diagram of the overall architecture of the present invention;

[0030] Figure 2 It is a diagram of the backbone network architecture of the present invention. Detailed Embodiments

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0032] An embodiment of the present invention discloses a method for positioning and identifying occluded ships for shore monitoring. Aiming at the problems of feature loss and insufficient detection accuracy in the positioning and identification of occluded ships, the present invention proposes an occluded ship positioning and identification network based on the improved YOLOv11. By introducing the improved InceptionNeXt lightweight backbone network, while effectively reducing the computational complexity of the model, the feature extraction ability is improved, especially in occluded scenarios, key feature regions can be better captured. In terms of the loss function, the CIoU loss function enhances the sensitivity to the position and shape of the bounding box, and the BCE classification loss function alleviates the impact of class imbalance on the detection performance by adjusting the attention to rare classes. The specific steps are as follows:

[0033] Step 1: Obtain and preprocess the images of the benchmark dataset;

[0034] Step 2: Introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing on the features through wavelet transform convolution. At the same time, add an activation function and a residual connection to each block to complete the construction of the occluded ship positioning and identification network;

[0035] Step 3: Use the images of the benchmark dataset to train the occluded ship positioning and identification network;

[0036] Step 4: Input the test pictures into the trained occluded ship positioning and identification network to obtain the recognition results.

[0037] Further, in Step 1, obtain the publicly available images of the benchmark dataset and preprocess the images. In the training stage, referring to the masking idea of MAE, random masking processing is added to all training pictures. For a training picture of 640*640 pixels, randomly place shadow blocks of 32*32 pixels on it, and control the final masking ratio to 10%, that is, 40 masking blocks of 32*32 are placed on each 640*640 picture. By performing masking operations on the training pictures, the model can be forced to learn more ship information, achieving higher detection accuracy in occluded situations. In the validation and inference stages, these masks will not be added to the pictures.

[0038] Further, in Step 2, the specific steps of network construction are as follows:

[0039] Build an improved detection network based on the YOLOv11 network, as Figure 1 shown.

[0040] First, the original backbone of YOLOv11 is improved by adopting the lightweight new backbone network InceptionNeXt and modifying it to enhance the feature extraction ability of the backbone network. As Figure 2 shown, the main improvement parts include the following:

[0041] 1. Improve the downsampling ability. Aiming at the problem of feature loss easily caused by each downsampling stage of InceptionNeXt, a normalization layer and a GELU activation function are introduced for each downsampling stage. The normalization layer can ensure the stability of the feature distribution, alleviate the problem of gradient disappearance or explosion, and thus improve the training stability of the network. The GELU activation function can improve the fitting ability of the network to the occluded environment through its smooth non-linear mapping. These two improvements can enable the model to better retain effective features in the downsampling stage and reduce the loss of occluded detail information.

[0042] 2. Introduce wavelet transform convolution. Through wavelet transform convolution, a large receptive field can be obtained without over-parameterization, and the low frequencies in the input can be better captured, enabling the model to perform more fine-grained multi-scale processing on features, thereby improving the model's ability to locate target regions and express edge features, and thus enhancing the detection ability for occluded ships.

[0043] 3. Introducing the attention mechanism into the convolutional neural network can improve the performance of the convolutional neural network, which has been widely recognized. Therefore, aiming at the problem of insufficient feature extraction ability of the traditional convolutional network, the Receptive-Filed Attention Convolution is introduced to add spatial attention to the model in the spatial domain. Starting from the importance of features within the receptive field, a dynamic weight adjustment mechanism is used to effectively capture target features at different positions within the receptive field under occlusion or complex backgrounds.

[0044] 4. Add an activation function and a residual connection to each block to improve the overall stability of the module. In addition, the residual connection retains the low-level features during feature fusion, enhances the network's ability to capture detail features in complex scenarios, and improves the model's expression ability and training efficiency.

[0045] When a 640*640*3 image enters the backbone feature extraction network, the image first passes through a two-fold downsampling convolution, followed by normalization and the SiLU activation function, to obtain initial features of 320*320*32. Then the initial features pass through four Stages, named Stage1, Stage2, Stage3, and Stage4. For each feature entering a Stage, the feature is represented as [B, C, H, W], where B, C, W, and H are the batch number, number of channels, width, and height respectively. First, the feature input to a Stage passes through an improved downsampling module to achieve two-fold downsampling to obtain features of the corresponding number of channels, represented by Then this downsampled feature is input to the Block stage in that Stage. In the Block stage, the input feature is divided into 8 parts according to the number of channels, i.e., . For the first part of the feature, no processing is done; for the second part of the feature, a 3*3 depthwise separable convolution is applied, with convolution performed independently for each channel, reducing the computational cost compared to standard convolution; for the third and fourth parts of the feature, by borrowing the improved idea of Inception, the 11*11 large kernel convolution is decomposed into 1*11 and 11*1 convolutional kernels, greatly reducing the huge computational cost brought by the large kernel convolution; for the fifth and sixth parts of the feature, a wavelet transform convolution is applied, which is a scheme that combines traditional wavelet transform and convolution, enabling a larger receptive field, better capturing of low frequencies in the input, and spatial mixing of the global receptive field configured in a pure convolution manner. For a given feature , it is first divided into four channels through wavelet transform, with the resolution of each channel being half of the original, including the low-frequency component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component. Then small kernel depth convolutions are performed on these four frequency maps, as shown in the following formula,

[0046]

[0047] where X is the input tensor, W is the weight tensor of the k*k depthwise kernel with four times the number of input channels of X, and finally the resulting tensor Y is subjected to inverse wavelet transform (IWT), and a residual connection is added to obtain the final output.

[0048] For the seventh and eighth parts of the feature, a Receptive-FiledAttention Convolution (RFAConv) is applied. By using the Unfold method for extracting receptive field spatial features, it emphasizes guiding attention to receptive field spatial features to enhance network performance. The specific convolution calculation formula is as follows:

[0049]

[0050] Among them, g represents grouped convolution, k represents the convolution kernel, Norm represents normalization, X represents the input feature map, and F is obtained by multiplying the attention map A rf with the transformed receptive field spatial feature F rf multiplied. RFAConv can completely solve the problem that the performance of convolutional neural networks is restricted by standard convolution operations by emphasizing the importance of different features within the receptive field slider and giving priority to the receptive field spatial features. This is because convolution operations rely on shared parameters and are insensitive to information differences brought about by position changes.

[0051] After obtaining these 8 processed features, the Concat method is used to aggregate them into a tensor of and passed through the GELU activation function, and then a residual connection is made with the original input to obtain the final Block output.

[0052] For each Stage, the downsampling stage is performed once, while the Block feature extraction stage is repeated multiple times. In the improved backbone network, the Block repeats in these 4 Stages in the order of [2, 2, 2, 2].

[0053] After obtaining the features extracted by the backbone network, the Fast Spatial Pyramid Pooling (SPPF) and the C2PSA attention module are used to enhance the features of the backbone features. At the same time, the Feature Pyramid Network (FPN) technology is used to obtain three final processed features with sizes of 80*80*128, 40*40*256, and 20*20*512 respectively. These are passed into the improved depthwise separable convolution lightweight detection head to obtain the final detection result.

[0054] Furthermore, in step three, the loss function is designed as follows:

[0055] For two common loss functions in object detection: bounding box loss and classification loss. For the bounding box loss, the commonly used loss is the Intersection over Union (IoU) function:

[0056] ;

[0057] However, this method is not sensitive enough to the change of the loss value when dealing with bounding boxes with small or no overlap, and it cannot effectively distinguish the direction deviation and scale difference of the bounding boxes. Therefore, a new CIoU is adopted. By introducing a center point distance penalty term, adding an aspect ratio penalty term, and optimizing the comprehensive loss in three parts, compared with the traditional IoU, CIoU can provide effective gradient information in scenarios with no overlap or small overlap, and improve the robustness of bounding box regression.

[0058]

[0059] Among them, respectively represent the center points of the predicted box and the ground truth box, represents the Euclidean distance between the center points of the predicted box and the ground truth box, represents the diagonal length of the smallest circumscribed rectangle that can contain both the predicted box and the ground truth box, is a weight function, used to measure the consistency of the aspect ratio.

[0060]

[0061]

[0062] For the classification loss, the BCE classification loss function, also known as Binary Cross-Entropy, is adopted. Binary Cross-Entropy is a commonly used classification loss function in deep learning, mainly used for binary classification tasks or multi-label classification tasks where each class is processed independently. It optimizes the classification performance of the model by measuring the difference between the model's predicted value and the target value. For a dataset containing N samples, the BCE loss function is:

[0063]

[0064] Among them, N represents the batch size, represents the ground truth value, represents the predicted value. When the true label is 1, the loss function becomes . This means that if the probability predicted by the model is closer to 1, the loss is smaller; if the predicted probability is closer to 0, the loss is larger.

[0065] When the true label is 0, the loss function becomes . This means that if the probability predicted by the model is closer to 0, the loss is smaller; if the predicted probability is closer to 1, the loss is larger.

[0066] This embodiment also discloses an occluded ship positioning and recognition system for shore monitoring, including:

[0067] Dataset Image Acquisition Module: Used to acquire and preprocess the benchmark dataset images;

[0068] Occluded Ship Localization and Recognition Network Construction Module: Used to introduce a normalization layer and GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing on the features through wavelet transform convolution. At the same time, add activation functions and residual connections to each block to complete the construction of the occluded ship localization and recognition network;

[0069] Occluded Ship Localization and Recognition Network Training Module: Used to train the occluded ship localization and recognition network using the benchmark dataset images;

[0070] Recognition Result Output Module: Used to input the test pictures into the trained occluded ship localization and recognition network to obtain the recognition results.

[0071] To verify the effectiveness of the algorithm, the present invention will be compared with the models of other advanced works. The comparative experiment is carried out on the test set of the MVDD13 public dataset. The detailed experimental results are shown in Table 1.

[0072] Table 1 Comparison of Experimental Results between Other Advanced Models and the Method of the Present Invention

[0073]

[0074] In summary, these results show that the method of the present invention has achieved good results on the MVDD13 public dataset. Compared with the results of recent related works, the method of the present invention has higher accuracy and lower computational complexity.

[0075] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, please refer to the description in the method section.

[0076] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for positioning and identifying occluded ships for shore monitoring, characterized in that, It includes the following steps: Obtain and preprocess the reference dataset images; Introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing of features through wavelet transform convolution. At the same time, add an activation function and a residual connection to each block to complete the construction of the occluded ship localization and recognition network; Train the occluded ship localization and recognition network using the reference dataset images; Input the test images into the trained occluded ship localization and recognition network to obtain the recognition results.

2. The method for positioning and identifying an occluded ship for shore monitoring according to claim 1, wherein The normalization layer is used to ensure the stability of the feature distribution, reduce the problem of gradient vanishing or explosion, thereby improving the training stability of the network; the GELU activation function improves the fitting ability of the network to the occluded environment through its smooth non-linear mapping.

3. A method for positioning and identifying an occluded ship for shore monitoring according to claim 1, characterized in that Introduce a convolutional neural network in the attention mechanism, which is used to add spatial attention to the model in the spatial domain. Starting from the importance of features within the receptive field, a dynamic weight adjustment mechanism is used to effectively capture target features at different positions within the receptive field under occlusion or complex backgrounds.

4. A method for positioning and identifying an occluded ship for shore monitoring according to claim 1, characterized in that, The residual connection retains the low-level features during feature fusion.

5. A method for positioning and identifying an occluded ship for shore monitoring according to claim 1, characterized in that, The loss function adopts a bounding box loss and a classification loss.

6. The method for positioning and identifying an occluded ship for shore monitoring according to claim 5, wherein The bounding box loss adopts the CIoU loss function, and the calculation formula is as follows: ； Among them, respectively represent the center points of the predicted box and the ground truth box, represents the Euclidean distance between the center points of the predicted box and the ground truth box, represents the diagonal length of the smallest bounding rectangle that can simultaneously contain the predicted box and the ground truth box, is the weight function, used to measure the consistency of the aspect ratio.

7. A method for positioning and identifying an occluded ship for shore monitoring according to claim 5, characterized in that, The classification loss adopts the BCE loss function, and the calculation formula is as follows: ； Among them, N represents the number of batches, represents the true value, represents the predicted value.

8. An occlusion ship positioning and recognition system for shore monitoring, characterized in that, It includes: Dataset image acquisition module: used to obtain and preprocess the reference dataset images; Occluded ship localization and recognition network construction module: used to introduce a normalization layer and a GELU activation function in the downsampling stage of the lightweight backbone network InceptionNeXt, introduce a convolutional neural network in the attention mechanism, and perform more fine-grained multi-scale processing of features through wavelet transform convolution. At the same time, add an activation function and a residual connection to each block to complete the construction of the occluded ship localization and recognition network; Occluded ship localization and recognition network training module: used to train the occluded ship localization and recognition network using the reference dataset images; Recognition result output module: used to input the test images into the trained occluded ship localization and recognition network to obtain the recognition results.

Images