Ship cable identification method based on improved VGG16 network and bidirectional feature fusion

Abstract

The invention discloses a ship cable identification method based on an improved VGG16 network and bidirectional feature fusion. The method comprises an image acquisition module, a preprocessing module, a deep learning recognition module and an output module, and comprises the following steps: firstly, carrying out gray enhancement and anti-interference filtering on a construction drawing through the image preprocessing module, and constructing a ship cable marking database; then, an improved VGG16 network deep convolutional layer is adopted to extract cable multi-scale features, a full connection layer module of an original network is deleted, a lightweight channel self-attention module is added, and BiFPN is adopted to carry out bidirectional feature fusion on shallow texture features and deep semantic features of the improved VGG16 network, so that the features of the ship cable are captured more effectively; and finally, outputting structured data of cable numbers, specifications and path coordinates. The identification precision mAP, the processing speed and the generalization ability of the method are superior to those of a traditional method, the method can be effectively applied to the field of ship cable identification and the like, and the industrial detection efficiency and the automation level are greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and ship electrical engineering, and particularly relates to a ship cable recognition method based on an improved VGG16 network and bidirectional feature fusion. BACKGROUND

[0002] The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion is applicable to recognition in the field of ship construction drawings, and can improve the design quality of ship electricity and lay a foundation for the intelligent transformation of ship design.

[0003] With the increasing complexity of ship design and manufacturing, the key role of electrical systems is increasingly prominent. As a system blueprint, electrical drawings carry a large number of equipment symbols, line directions and technical parameters. However, traditionally, drawing recognition and cable list compilation mainly rely on manual operation, which is not only inefficient and prone to errors, but its limitations are particularly obvious when facing complex systems and changing needs. These problems lead to prolonged design cycles, rising costs and risks, and seriously restrict the efficiency and accuracy of design. With the rapid development of artificial intelligence technology, especially the introduction of deep learning technology, it has brought revolutionary changes to ship electrical design. Traditional network models are limited in feature extraction, parameter quantity and feature relevance, processing speed, etc. Image classification is one of the basic tasks in computer vision, aiming to classify input images into predefined categories. Traditionally, convolutional neural networks (CNN) dominate in image classification tasks. However, in recent years, Vision Transformer (ViT) based on Transformer architecture has shown strong performance and flexibility in image classification. ViT uses the self-attention mechanism of Transformer to process image data, which can capture global information and achieve higher classification accuracy. The traditional VGG16 network contains 13 convolutional layers and three fully connected layers, which learn complex features through a 16-layer network. The texture features of the shallow layer are used to improve small target detection, and the semantic features of the deep layer provide more accurate spatial positioning ability.

[0004] The above traditional method does not consider the processing of complex data and the extraction of high-level features, as well as the relationship between shallow texture features and deep texture features. The traditional VGG16 network has a huge parameter quantity and computational complexity in the fully connected layer, which leads to high computational complexity and consumes a large amount of computing resources and time in the training process. Therefore, how to improve the structure of the model to improve the efficiency and accuracy of ship cables is a problem that needs to be solved. SUMMARY

[0005] The present application is just for the problem of low recognition accuracy and low recognition efficiency in the existing model, proposes a ship cable recognition method based on improved VGG16 network and bidirectional feature fusion, deletes the full connection layer of the traditional VGG16 network, retains 13 convolutional layers as the feature extraction backbone, adds a channel self-attention mechanism (SEBlock) module after the 13 convolutional layers, and adds a bidirectional feature fusion module after the improved VGG16 output, so that the texture features of the shallow layer and the semantic features of the deep layer are connected across the scale, the semantic information of the deep layer flows to the shallow layer to enhance the feature understanding of the shallow layer, and the high-resolution detail information of the shallow layer flows to the deep layer to enrich the details and positioning ability of the deep layer features.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is: a ship cable recognition method based on improved VGG16 network and bidirectional feature fusion, comprising the following steps:

[0007] S1, obtaining a ship cable construction drawing to be recognized

[0008] S2, ship cable construction drawing pretreatment: performing gray scale enhancement and anti-interference filtering processing on the input cable construction drawing, and extracting the ROI value of a specific area;

[0009] S3, component ship cable labeling database: establishing a structured labeling data set according to the number, specification and path coordinates of the cable;

[0010] S4, feature extraction based on improved VGG16 network: extracting cable multi-scale features by using the convolutional layer of the VGG16 network, and initializing the network weight by using transfer learning;

[0011] S5, feature fusion and optimization: using bidirectional feature pyramid (BiFPN) to adaptively fuse the shallow texture features and deep semantic features;

[0012] S6, output of cable recognition result: generating a structured data file containing cable line number specification parameters.

[0013] As one expression of the present application, the anti-interference filtering processing in step S2 specifically comprises:

[0014] S21: In the processing process of adaptive median filtering, the input image signal x(n) is an initialization signal window size MxN, the output image signal y(n) is a termination window size NxN, and the pixel gray value is Z xy ;

[0015] S22: In the operation process, the maximum value Z max , the minimum value Z min , and the median value Z med of the gray scale of the pixels in the window are calculated, and α1=Z max +Z min +Z med .med -Z min , a2 = Z med -Z max ;

[0016] S23: compare the two calculated values a1, a2 with the expected value d(n), and determine whether the condition a1 > d(n) & a2 < d(n) is met, if the condition is met, go to step S24, if not, increase the window size by 2 until the mean square value of e(n) is minimized, exit the termination condition, and output the value Z xy ;

[0017] S24: record b1 = Z xy -Z min , b2 = Z xy -Z max If the condition b1 > 0 & b2 < 0 is met, the filter output value is Z xy , otherwise the filter output value is Z med .

[0018] As another improvement of the present application, the step S3 of building a labeled database further includes a data enhancement strategy, specifically including: randomly rotating the cable construction drawing by ±15°, scaling the proportion by 0.8-1.2, and adding salt and pepper noise (noise density ≤5%)

[0019] As another improvement of the present application, the improvement of the VGG16 network in step S4 includes:

[0020] S41: delete the original VGG16 network full connection layer, and keep 13 convolutional layers as the feature extraction backbone. The VGG16 network usually consists of 5 convolutional blocks to form the feature extraction backbone, the first and second convolutional blocks each contain two convolutional layers, and the third, fourth and fifth convolutional blocks each contain three convolutional layers. After each convolutional layer, a ReLU activation function and a max-pooling layer are connected to represent it. The improved VGG16 network retains 13 convolutional layers and deletes the subsequent full connection layer;

[0021] S42: embed a channel attention module (SE Block) after the conv5-3 layer. The VGG16 network has a large number of full connection layer parameters, which is prone to overfitting. By deleting the full connection layer and adding a channel self-attention mechanism afterwards, the SE Block is a lightweight module that mainly includes Transformation, global average pooling Squeeze, a small multilayer perceptron Excitation, and a reweighting operation;

[0022] As another improvement of the present application, in step S42, the Transformation method specifically includes: input feature map X is changed to feature map U through a convolution operator, and the formula is specifically:

[0023]

[0024] where X ∈ R ^ (H' × W' × C') is the input feature map; U ∈ R ^ (H × W × C) is the output feature map; where R is a real number space; H is the height of the feature map; W is the width of the feature map; C is the number of channels of the feature map; V is a set of learned filter kernels; V C is the parameter of the Cth filter; is a 2D spatial kernel; * is a convolution operation; X S is the Sth channel of the input feature map X;

[0025] Global Average Pooling Squeeze compresses the global spatial information into a channel descriptor, using the global average pooling of the channel, the feature map of WxHxC containing global information is directly compressed into a 1x1xC feature vector Z, so that the generated channel-level statistical data contains the context information, which alleviates the dependence of the channel. The specific formula is:

[0026]

[0027] where Z C is the Cth element of Z; u C (i,j) represents the pixel value at coordinate (i,j) in the Cth channel;

[0028] Small Multi-Layer Perceptron Excitation adopts a two-layer fully connected gating mechanism. The first fully connected layer compresses C channels into C / r channels to reduce the computational load, and then passes through a nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and finally obtains the weight s through the Sigmoid activation. The specific formula is:

[0029] s = σ (g (z, K) ) = σ (K2δ (K1z) ) (3)

[0030] where r is the compression ratio; σ (·) is the Sigmoid function; g (z, K) is a gating generation function, where z is the input feature, K is the weight matrix, K1 is the dimension reduction weight matrix, and K2 is the dimension increase weight matrix; δ (·) is the activation function;

[0031] Finally, the attention weight of the previous operation is added to the feature of each channel through the operation of the weighting Scale. The specific formula is:

[0032]

[0033] where is the Cth channel of the output feature map; F scale is a scaling function; u C is the Cth channel of the input feature map; s C is a scaling factor;

[0034] As another improvement of the present application, the specific implementation of the bidirectional feature pyramid (BiFPN) in step S5 is as follows:

[0035] S51: Cross-scale connection is performed on the output feature maps of the conv3-3, conv4-3 and conv5-3 layers of VGG16.

[0036] S52: The weighted feature fusion formula P out =∑ i w i ·P i is used to realize the fusion of bidirectional features, where w i is a learnable weight parameter; P i is the input multi-scale feature map.

[0037] Further, in step S6, an Excel table file containing the cable number specifications and parameter structured data is finally generated.

[0038] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion when executing the program.

[0039] A computer readable storage medium has computer instructions stored thereon, and the computer instructions are executed by a processor to implement the ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion.

[0040] Compared with the prior art, the present application has the beneficial effects: the present application comprehensively considers the identification accuracy and identification efficiency of the model, improves the traditional VGG16 network, deletes the original model full connection layer, adds a channel self-attention mechanism module (SE Block), retains the original convolution layer, and replaces it with a lightweight channel self-attention module, mainly including a global average pooling and a small multilayer perceptron, so that the volume of the model is significantly reduced, the inference speed is faster, and the required memory for training is smaller, then the improved VGG16 network model is added with a bidirectional feature fusion module, so that the texture features of the shallow layer and the semantic features of the deep layer are bidirectionally fused, so that the strong semantic information of the deep layer features can be more effectively injected into the shallow layer features, which makes the edges, corner points, textures and other details of the original components in the ship cable drawing can understand which category the details belong to, the traditional model is not accurate for identifying some small target individuals in the ship cable drawing, and the small target individuals almost disappear or become very fuzzy in the deep feature map, the shallow feature map has high resolution and can effectively detect small target individuals, so through bidirectional feature fusion, not only the semantic of the shallow layer is enhanced, but also the deep layer features can make correct decisions. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 FIG. 1 is a diagram of the improved VGG16 network and bidirectional feature fusion ship cable recognition system of the present application,

[0042] Figure 2 FIG. 4 is a data set comparison chart before and after image enhancement of the present application,

[0043] Figure 3 FIG. 5 is a structure diagram of the improved VGG16 network of the present application,

[0044] Figure 4 FIG. 6 is a structure diagram of the improved bidirectional feature fusion of the present application,

[0045] Figure 5 FIG. 7 is a structured cable list chart output by the present application,

[0046] Figure 6 FIG. 8 is an mAP comparison chart of the improved VGG16 network and bidirectional feature fusion of the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Embodiment 1

[0049] As Figure 1 shown, the overall structure of the ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion includes an image processing module. The input image is subjected to image enhancement processing. The improved VGG16 network retains the original convolution module of the network structure and uses a lightweight channel self-attention module to replace the original fully connected layer module, thereby solving the defects of the original model such as complex calculation and low efficiency. The BiFPN bidirectional feature fusion module is added after the improved VGG16 network structure. The weighted feature bidirectional fusion is used after the output of the second convolution block and the fifth convolution block of the VGG16 network, which effectively extracts the feature details of the ship cable drawing. Finally, the structured cable list is output.

[0050] The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion includes the following steps:

[0051] S1, obtaining a ship cable construction drawing to be recognized

[0052] S2, ship cable construction drawing preprocessing: performing gray scale enhancement and anti-interference filtering processing on the input cable construction drawing, and extracting the ROI value of a specific region;

[0053] In the processing process of the adaptive median filter, the input image signal x(n) is the initialization signal window size MxN, the output image signal y(n) is the termination window size NxN, and the pixel gray value is Z xy .

[0054] In the operation process, the maximum value Z max , the minimum value Z min , and the median value Z med of the gray scale of the pixels in the window are calculated, and α1=Z med -Z min , α2=Z med -Z max .

[0055] Compare α1 and α2 with the expected value d(n) to determine whether the condition α1>d(n)&α2<d(n) is met. If the condition is met, go to step S24. If not, increase the window size by 2 until the mean square value of e(n) is minimized, jump out of the termination condition, and output the value Z xy .

[0056] b1=Z xy -Z min , b2=Z xy -Z maxIf condition b1>0 & b2<0 is met, the filter output value is Z xy , otherwise the filter output value is Z med .

[0057] S3, component ship cable label database: according to the number, specification and path coordinates of the cable, a structured label data set is established;

[0058] The cable construction drawing is randomly rotated by ± 15°, scaled by 0.8-1.2, and salt and pepper noise (noise density ≤5%) is added, and the data set before and after image enhancement is compared as shown in Figure 2 ;

[0059] S4, feature extraction based on improved VGG16 network: the convolutional layers of VGG16 network are used to extract multi-scale features of the cable, and the network weights are initialized by transfer learning, and the structure diagram of the improved VGG16 network is as shown in Figure 3 ;

[0060] The original VGG16 network full connection layer is deleted, and 13 convolutional layers are retained as the feature extraction backbone. VGG16 network usually consists of 5 convolutional blocks to form the feature extraction backbone, the first and second convolutional blocks contain two convolutional layers respectively, and the third, fourth and fifth convolutional blocks contain three convolutional layers respectively. After each convolutional layer, a ReLU activation function and a maximum pooling layer are connected to represent it. The improved VGG16 network retains 13 convolutional layers and deletes the subsequent full connection layer;

[0061] A channel attention module (SE Block) is embedded after the conv5-3 layer. The parameter quantity of the full connection layer of the VGG16 network is large, which is easy to cause overfitting. By deleting the full connection layer, a channel self-attention mechanism is added behind. SE Block is a lightweight module, mainly including Transformation, global average pooling Squeeze, small multilayer perceptron Excitation and reweighting operation;

[0062] The Transformation method is specifically: the input feature map X is changed into a feature map U through a convolution operator, and the formula is specifically:

[0063]

[0064] In the formula, X∈R ^ (H'×W'×C') is the input feature map; U∈R ^ (H×W×C) is the output feature map; where R is a real number space; H is the height of the feature map; W is the width of the feature map; C is the number of channels of the feature map; V is a set of learned filter kernels; V C is the parameter of the Cth filter; is a 2D spatial kernel; * is a convolution operation; X S is the S-th channel of the input feature map X;

[0065] Global Average Pooling Squeeze compresses global spatial information into a channel descriptor, using global average pooling of channels, a feature map of WxHxC containing global information is directly compressed into a 1x1xC feature vector Z, so that the generated channel-level statistics contain contextual information, alleviating the dependence of channels. The specific formula is:

[0066]

[0067] In the formula, Z C is the C-th element of Z; H and W represent the height and width of the feature map, respectively; u C (i,j) represents the pixel value at coordinate (i,j) in the c-th channel;

[0068] Small Multilayer Perceptron Excitation adopts a two-layer fully connected gating mechanism. The first fully connected layer compresses C channels into C / r channels to reduce the amount of calculation, and then passes through a nonlinear activation layer. The second fully connected layer restores the channel number to C channels, and finally obtains the weight s through the Sigmoid activation. The specific formula is:

[0069]

[0070] In the formula, r is the compression ratio; σ(·) is the Sigmoid function; g(z,K) is a gating generation function, where z is the input feature, K is the weight matrix, K1 is the dimension reduction weight matrix, and K2 is the dimension increase weight matrix; δ(·) is the activation function;

[0071] Finally, the attention weight is added to the feature of each channel through the operation of weighting Scale. The specific formula is:

[0072]

[0073] In the formula, is the C-th channel of the output feature map; F scale is a scaling function; u C is the C-th channel of the input feature map; s C is a scaling factor;

[0074] S5, feature fusion and optimization: bidirectional feature pyramid (BiFPN) is used to adaptively fuse shallow texture features and deep semantic features. The structure diagram of bidirectional feature fusion based on shallow texture features and deep semantic features is shown in Figure 4 ;

[0075] S51: Cross-scale connection is performed on the output feature maps of the conv2-2, conv3-3, conv4-3 and conv5-3 layers of the VGG16;

[0076] S52: The weighted feature fusion formula P out =∑ i w i ·P i is used to realize fusion of the bidirectional features, where w i is a learnable weight parameter; P i is the input multi-scale feature map;

[0077] S6, output of the cable identification result: a structured data file containing the cable number specification parameters is generated; the finally generated structured cable list is as shown in Figure 5 ;

[0078] Figure 6 is the mAP comparison chart of the improved VGG16 network and the bidirectional feature fusion, and it can be seen from the comparison that, as the number of iterations increases, the accuracy of the improved VGG16 network has reached 60% after 20 iterations, and the final accuracy is above 90%. Compared with the traditional VGG16 network, it can be seen that the improved VGG16 network and the bidirectional feature fusion network have better recognition speed and accuracy.

[0079] In the description of the present specification, the description referring to the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0080] It should be noted that the above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. For ordinary skilled persons in the technical field, without departing from the principles of the present application, a number of improvements and refinements can be made, which fall within the protection scope of the claims of the present application.

Claims

1. A ship cable recognition method based on an improved VGG16 network and bidirectional feature fusion, characterized in that, The method comprises the following steps: S1, obtaining a ship cable construction drawing to be identified, S2, ship cable construction drawing preprocessing: performing gray scale enhancement and anti-interference filtering processing on the input cable construction drawing, and extracting ROI values of a specific region; S3, component ship cable labeling database: establishing a structured labeling dataset according to the number, specification and path coordinates of the cable; S4, feature extraction based on the improved VGG16 network: extracting cable multi-scale features by using the convolutional layers of the VGG16 network, and initializing the network weights through transfer learning; S5, feature fusion and optimization: performing adaptive fusion on shallow texture features and deep semantic features by using a bidirectional feature pyramid (BiFPN); S6, output of the cable identification result: generating a structured data file containing cable line number and specification parameters.

2. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 1, characterized in that, The anti-interference filtering processing in step S2 specifically comprises: S21: in the processing procedure of adaptive median filter, input image signal x(n) is initialization signal window size is MxN, output image signal y(n) is termination window size NxN, and pixel gray value is Z xy ; S22: In the process of operation, the maximum value Z of the gray scale of the pixels in the window needs to be counted max , the minimum value Z min , the median value Z med , and α1=Z med -Z min , α2=Z med -Z max ; S23: Compare the two calculated values a1, a2 with the expected value d(n) to determine whether the condition a1>d(n) & a2d(n) is met. If the condition is met, go to step S24. If the condition is not met, increase the window size by 2 until the mean square value of e(n) is minimized, exit the termination condition, and output the value Z xy ; S24: let b1 = Z xy -Z min , b2 = Z xy -Z max If the condition b1 > 0 & b2 < 0 is met, the filter output value is Z xy , otherwise the filter output value is Z med .

3. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 1, characterized in that, The labeling database construction in step S3 further comprises a data enhancement strategy, specifically including: randomly rotating the cable construction drawing by ±15°, scaling the proportion by 0.8-1.2, and adding pepper and salt noise (noise density ≤5%).

4. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 1, characterized in that, The improvement of the VGG16 network in step S4 comprises: S41: deleting the original VGG16 network full connection layer, retaining 13 convolutional layers as the feature extraction backbone, and the VGG16 network is composed of 5 convolutional blocks as the feature extraction backbone, the first and second convolutional blocks contain two convolutional layers respectively, and the third, fourth and fifth convolutional blocks contain three convolutional layers respectively, followed by ReLU activation function and maximum pooling layer for representation, the improved VGG16 network retains 13 convolutional layers and deletes the subsequent full connection layer; S42: embedding a channel attention module (SE Block) after the conv5-3 layer, the full connection layer of the VGG16 network has a large number of parameters and is prone to overfitting, by deleting the full connection layer and adding a channel self-attention mechanism afterwards, the SE Block is a lightweight module mainly including Transformation, global average pooling Squeeze, small multilayer perceptron Excitation and reweighting operation.

5. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 4, characterized in that, In step S42, the Transformation method specifically comprises: the input feature map X is changed into a feature map U through a convolution operator, and the formula is specifically: where X ∈ R ^ (H' x W' x C') is the input feature map; U ∈ R ^ (H x W x C) is the output feature map; where R is a real number space; H is the height of the feature map; W is the width of the feature map; C is the number of channels of the feature map; V is a set of learned filter kernels; V C is the parameter of the Cth filter; is a 2D spatial kernel; * is the convolution operation; X S is the Sth channel of the input feature map X; The global average pooling Squeeze compresses the global spatial information into a channel descriptor, uses the global average pooling of the channel to directly compress the feature map of WxHxC containing global information into a 1x1xC feature vector Z, so that the generated channel-level statistical data contains context information, and the channel dependency is alleviated, and the specific formula is: where Z C is the cth element of Z; u C (i,j) denotes the pixel value at coordinate (i,j) in the cth channel. The small multilayer perceptron Excitation adopts a two-layer full connection composed of a gate mechanism, the first full connection layer compresses C channels into C / r channels to reduce the calculation amount, and then passes through a nonlinear activation layer, and the second full connection layer restores the channel number to C channels, and finally obtains the weight s through the Sigmoid activation, and the specific formula is: s = σ(g(z, K)) = σ(K2δ(K1z)) (3) where r is the compression ratio; σ(·) is a Sigmoid function; g(z, K) is a gating generation function, where z is the input feature, K is the weight matrix, K1 is the dimension reduction weight matrix, and K2 is the dimension increase weight matrix; and δ() is an activation function; Finally, the previous attention weight is added to the feature of each channel by the operation of weighted Scale, and the specific formula is: wherein is the Cth channel of the output feature map; F scale is a scaling function; u C is the Cth channel of the input feature map; s C is a scaling factor.

6. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 1, characterized in that, The specific implementation of the bidirectional feature pyramid (BiFPN) in the step S5 is as follows: S51: cross-scale connection is performed on the output feature maps of the conv2-2, conv3-3, conv4-3 and conv5-3 layers of the VGG16. S52: Perform feature fusion by a weighted feature fusion formula P out =∑ i w i ·P i Implement bidirectional feature fusion, where w i is a learnable weight parameter, and P i is the input multi-scale feature map.

7. The ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to claim 1, characterized in that, In the step S6, an Excel table containing the cable line number specification and parameter structured data is finally generated.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to any one of claims 1-7 when executing the program.

9. A computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions implement the ship cable recognition method based on the improved VGG16 network and bidirectional feature fusion according to any one of claims 1-7 when executed by the processor.

Images