WO2020048183A1

WO2020048183A1 - Vessel type identification method based on coarse-to-fine cascaded convolutional neural network

Info

Publication number: WO2020048183A1
Application number: PCT/CN2019/092016
Authority: WO
Inventors: 陈信强; 杨勇生; 吴华锋; 苌道方; 于泽崴; 张倩楠; 陈晶; 傅俊杰; 赵建森; 陈辉兴; 刘卫; 李军军; 许波桅; 鲜江峰
Original assignee: 上海海事大学
Priority date: 2018-09-04
Filing date: 2019-06-20
Publication date: 2020-03-12
Also published as: CA3084451C; CA3084451A1; CN109299671A

Abstract

The present invention provides a vessel type identification method based on a coarse-to-fine cascaded deep convolutional neural network. The method uses a random heuristic selection mechanism to dynamically adjust structure and parameter settings of the depth network, and the method obtains the deep convolutional neural network capable of identifying the vessel type by means of the steps of coarse-level training and fine-level training; the coarse-level training process is similar to that of a conventional deep convolutional neural network, and an input sample of the training process is a vessel image; the fine-level training process is directed at a merchant vessel image having the lowest vessel type identification precision in the coarse-level training process, and retrains the deep convolutional neural network to improve the overall precision of the vessel type identification. By means of the method of the present invention, better identification precision can be realized for different vessel types, and information support is provided for automatic vessel type identification and vessel intelligent navigation.

Description

A cascaded rough-to-fine convolutional neural network ship type recognition method

Technical field

The invention relates to the technical field of maritime video surveillance, and in particular to a cascaded convolutional neural network ship type recognition method from coarse to fine.

Background technique

At present, Vessel Traffic Service (VTS) and Automatic Identification System (AIS) are the main means to obtain ship type information. After the ship enters the VTS report line, the on-board personnel on the ship report the basic information of the ship, such as the port of destination, the port of departure, and the type of ship, to the maritime supervision department through a VHF phone. In addition, the AIS system will also periodically distribute static and dynamic information of the ship, including the ship's type, position, call sign, ship name, gross tonnage, ship draft and speed, etc. However, AIS users need to manually enter static information such as ship type and ship call sign for the AIS system in advance. From the above analysis, it can be known that both VTS and AIS require human participation to obtain ship type information. With the rapid growth of maritime traffic and the rapid expansion of fleet size, these traditional methods of acquiring ship types require more and more manual intervention. Therefore, it is a very time-consuming task to obtain the ship type information by using traditional technical means. The automatic identification of ship types based on visual data information is one of the important challenges to be addressed in the era of unmanned ships and the era of intelligent navigation.

Disclosure of invention

The purpose of the present invention is to provide a cascaded deep convolutional neural network ship type identification method for overcoming the shortcomings of the prior art described above. For common ships (including container ships, tankers, and chemical tankers) , Bulk carriers, general cargo ships, LNG carriers, other merchant ships). Obtaining ship type information by traditional technical means is time-consuming and not conducive to improving the efficiency of maritime supervision.

The object of the present invention can be achieved by the following technical solutions:

A cascaded deep-convolution neural network ship type recognition method from coarse to fine, the method includes the following steps:

S1: Input pictures of all ship types and corresponding picture tags, and perform coarse-level training on the deep convolutional neural network from coarse to fine, to obtain the setting parameters of the deep convolutional neural network from coarse to fine, and Training recognition accuracy;

S2: Use the training to identify the ship type with the lowest accuracy picture to perform fine-level training on the deep to convolutional neural network from coarse to fine. If the deep convolutional neural network has not reached the preset convergence condition, return to step S1 to continue Training, otherwise execute step S3;

S3: Perform type recognition on the ship in the picture, and output the recognition result of the ship type.

Preferably, the step S1 includes the following steps:

S11: crop the originally input ship type picture to a fixed size, match the input ship type picture and corresponding picture tag, and obtain a formatted ship image and image tag;

S12: using the formatted ship image and image label to train a cascaded deep convolutional neural network from coarse to fine to obtain the setting parameters of the cascaded deep convolutional neural network from coarse to fine, Extract depth characteristics of different ship types;

S13: Obtain the confidence level distribution of the input ship type picture according to the depth feature, and output the ship type corresponding to the maximum confidence value as a single training recognition result;

S14: According to the recognition result of the single training, compare the true type of the ship corresponding to the picture tag to obtain the training recognition accuracy of different ship types.

Preferably, the step S12 includes the following steps:

S121: The cascaded deep-to-fine convolutional neural network uses convolution layers to extract ship features; the ship features include: low-level ship features, including ship textures, contours, and corner points; advanced ship features, Different types of ships are obtained by correspondingly abstracting the characteristics of low-level ships;

S122: The cascaded deep convolutional neural network from coarse to fine uses the pooling layer to reduce the dimension of the ship features and learn;

S123: The cascaded deep to convolutional neural network uses a local response normalization layer to increase the local response of the ship feature extracted by the convolution layer to randomly assign a larger response value and extract Generalize ship characteristics.

Preferably, the step S13 includes the following steps:

S131: The cascaded deep-to-fine convolutional neural network uses a fully connected layer to map the generalized ship feature to a single ship feature vector, and its expression is as follows:

F _out = Θ × F _in

Among them: F _out is a single ship feature vector output from the fully connected layer, which has a total of n ₁ elements; F _in is the input generalized ship feature, its dimension is n ₂ +1; Θ is the F _in and F _out 's connection matrix with dimensions n ₁ × (n ₂ +1);

S132: The cascaded deep-to-fine convolutional neural network uses a loss layer to generate a probability vector based on the single ship feature vector as an input ship image, and the elements in the vector represent the probability of the type of ship The calculation expression is as follows:

Among them, F _p is a single ship feature vector of the loss layer; v _j is a weight corresponding to the j-th ship type when calculating a ship probability vector.

Preferably, the calculation expression of the training recognition accuracy e1t of the ship type in step S14 is as follows:

Among them: N _s is the total number of ship pictures to be identified; N _er is the total number of ship pictures with incorrect type recognition.

Preferably, the step S2 includes the following steps:

S21: Acquire all the training pictures of the ship type according to the pictures of the ship type with the lowest training recognition accuracy, as the input samples for the fine-level training of the coarse to fine deep convolutional neural network;

S22: Use a random heuristic selection method to select one of the data enhancement, selective discarding method, and selective connection method as a random regularization mechanism;

S23: According to the setting parameters, use the random regularization mechanism to train the coarse to fine deep convolutional neural network, and obtain the refined parameters and the adjusted coarse to fine deep convolution. Neural Networks;

S24: The coarse-to-fine deep convolutional neural network adjusted according to the refinement parameters is used to re-identify the ship type of the picture. If the training type recognition accuracy change rate is less than a preset threshold, the training process ends; if the ship type The type of training recognition accuracy change rate is greater than the preset threshold, the coarse-to-fine deep convolutional neural network completes the current fine-level training, and returns to step S1 to continue training.

Preferably, the data enhancement includes horizontally / vertically flipping the training picture, changing a color, and / or randomly changing the size of the training picture.

Preferably, the selective discarding method sleeps some neurons of the convolution layer with a preset probability, and all neurons of the convolution layer sleep or stop hibernation with the same probability.

Preferably, the selective connection method is to randomly modify the weights of the neurons in the convolution layer, thereby weakening or strengthening the influence of the ship features extracted by the neurons in the layer on the accuracy of ship type recognition.

Preferably, the calculation formula for the rate of change in training recognition accuracy P _ia of the ship type is as follows:

Among them: A _jc is the accuracy of class j ship type recognition in step S1; A _jf is the accuracy of class j ship type recognition in step S2.

Compared with the prior art, the present invention has the following advantages: the method of the present invention realizes the ship type in the picture by stepwise training from rough to fine level by cascading from coarse to fine deep convolutional neural network. The automatic and accurate identification of ships effectively realizes the automation and high-precision identification of ship types, and has important practical value for organizing maritime traffic order, ensuring maritime traffic safety, and improving navigation efficiency in the era of intelligent navigation.

Brief description of the drawings

Embodiments of the present invention will be further described below with reference to the accompanying drawings, in which:

FIG. 1 is a schematic diagram of an overall process of the present invention;

FIG. 2A is a schematic flowchart of step S1 in a preferred embodiment; FIG.

FIG. 2B is a schematic flowchart of step S12 in the preferred embodiment; FIG.

FIG. 2C is a schematic flowchart of step S13 in the preferred embodiment; FIG.

FIG. 2D is a schematic flowchart of step S2 in the preferred embodiment; FIG.

FIG. 3 is a picture of a typical ship to be identified according to the present invention;

4A is a deep network recognition error distribution of different batch sample sizes when different parameters are set;

4B is a deep network recognition error distribution with different weight value attenuation when different parameters are set;

4C is a deep network recognition error distribution with different learning rates when different parameters are set;

4D is a deep network recognition error distribution with different training times when different parameters are set;

5A is a recognition result of a container ship;

Figure 5B is the identification result of the general cargo ship;

5C is the recognition result of the tanker;

FIG. 6 is a classification distribution of different ship types based on a cascaded deep to convolutional neural network;

FIG. 7 is a comparison chart of ship type recognition accuracy by different methods.

The best way to implement the invention

The present invention is described in detail below with reference to the drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

This embodiment provides a cascade-type coarse-to-fine deep convolutional neural network ship type recognition method. Referring to the schematic diagram of the overall process shown in FIG. 1, the method includes the following steps:

S1: Input pictures of all ship types and corresponding picture tags, perform rough-level training on the deep convolutional neural network from coarse to fine, obtain the setting parameters of the deep convolutional neural network from coarse to fine, and obtain different ship types Training recognition accuracy;

Further referring to FIG. 2A, in a preferred embodiment, the above step S1 includes the following steps:

S12: Use the formatted ship images and image tags to train the cascaded deep convolutional neural network from coarse to fine, obtain the setting parameters of the cascaded deep convolutional neural network from coarse to fine, and extract Depth characteristics of different ship types;

S13: Obtain the confidence level distribution of the input ship type picture according to the obtained depth characteristics, and output the ship type corresponding to the maximum confidence value as a single training recognition result;

S14: According to the recognition result of a single training, compare the true type of the ship corresponding to the picture label to obtain the training recognition accuracy of different ship types.

As shown in FIG. 2B, the above step S12 further includes the following steps:

S121: A cascaded deep-to-fine convolutional neural network uses convolution layers to extract ship features. The ship features here include low-level ship features and high-level ship features, where low-level ship features include ship texture, contours, and corner points. The characteristics of high-level ships are high-level features of low-level ships, and different types of ships present different high-level features. Then the ship features extracted by the convolutional layer in this step are as follows:

among them:

Is the mth input ship feature map of the r-1th network layer;

Is the connection weight of the ship feature map of the n-th network output layer and the m-th input feature map;

Is the bias value of the n-th feature map of the r-th layer convolutional network; the symbol f represents the activation function that activates the r-th layer of the convolutional network neuron;

Is the n-th output feature map of the r-th layer convolutional network;

S122: The cascaded deep-to-fine convolutional neural network uses the pooling layer to learn the above-mentioned ship features, and abandons the learning of secondary ship features, and retains the learning of important ship features. The secondary ship feature refers to the feature extracted by the neurons whose pooled weight value is less than 75% of the entire layer of the network. The important ship feature refers to the neurons whose pooled weight value is greater than 75% of the whole network. Extracted features. The pooling layer expression here is as follows:

Where: k is the dimension of the pooling kernel; d is the step size;

Is the n-th ship feature map generated by the cascaded convolutional layer of coarse to fine deep convolutional neural network; Pool _{u, v} pooling layer pair feature map

Important features obtained by pooling; parameters u and v are dimensions of Pool _{u, v} ;

S123: The cascaded deep to convolutional neural network uses the local response normalization layer to increase the local response of the ship features extracted by the convolution layer to randomly assign larger response values to extract generalized ship features. The expression of the local response normalization layer is as follows;

among them:

Is the i-th ship characteristic response of the neural unit in the r-th local response normalization layer;

Is the eigenvalue of the i-th ship feature map in the r-th local response normalization layer unit; parameters a, η,

h is a predetermined parameter; parameter U is a tag of a ship type.

In a further preferred embodiment, as shown in FIG. 2C, the above step S13 includes the following steps:

S131: The cascaded deep to convolutional neural network uses a fully connected layer to map the generalized ship features obtained in the above step S123 to a single ship feature vector, and its expression is as follows:

F _out = Θ × F _in

Among them: F _out is a single ship feature vector output from the fully connected layer, which has a total of n ₁ elements; F _in is the input generalized ship feature, its dimension is n ₂ +1; Θ is the F _in and The connection matrix of F _out , the dimension is n ₁ × (n ₂ +1), where n2 is the output ship feature after generalization;

S132: The cascaded deep-to-fine convolutional neural network uses the loss layer to generate a probability vector based on the above-mentioned single ship feature vector as an input ship image, and the elements in the vector represent the probability of the type to which the ship belongs, and calculates the expression. as follows:

In another preferred embodiment, the calculation expression of the training type recognition accuracy e _1t of the ship type in step S14 is as follows:

In another preferred embodiment, as shown in FIG. 2D, the above-mentioned step S2 includes the following steps:

S22: Use a random heuristic selection method to select one as a random regularization mechanism from data enhancement, selective discarding, and selective connection. The calculation expression of the random heuristic selection method is as follows:

P _is = Max {ω _i1 × θ _ir + ω _i2 × θ _ih }

Among them: P _is the probability of choosing the i-th regularization method, the data is enhanced, and the dropout and dropconnect mechanisms are respectively labeled as 1, 2, and 3. P _1s indicates that data enhancement is selected as the current regularization method, P _2s indicates that dropout is selected as the current regularization method, and P _3s indicates that dropout is selected as the current regularization method; θ _ir is a random factor; θ _ih is heuristic factor; ω _i1 of representation θ _ir weight; ω _i2 representative of θ _ih weight; N _i ship picture number of class i to be trained; P _ia after the session, rate of change of the depth of the network of the ship recognition accuracy;

S23: According to the setting parameters of step S1, use a random regularization mechanism to train the deep convolutional neural network from coarse to fine, and obtain the refined parameters and the adjusted deep convolutional neural network from coarse to fine;

S24: The coarse-to-fine deep convolutional neural network adjusted according to the refined parameters re-recognizes the ship type of the picture. If the training type ’s recognition accuracy change rate is less than a preset threshold, the training process ends; according to multiple tests As a result, a better network performance can be obtained when the preset threshold is set to 0.01; if the rate of change in training recognition accuracy of the ship type is greater than the preset threshold, the current fine-level training is completed by a coarse to fine deep convolutional neural network And return to step S1 to continue training.

The above-mentioned data enhancement mechanism specifically includes horizontal / vertical flipping, color changing, and / or randomly changing the training picture size of the original training picture.

The selective discarding method described above sleeps some neurons of the convolutional layer with a preset probability, and all neurons of the convolutional layer sleep or stop dormant with the same probability.

The above selective connection method is to randomly modify the weights of the neurons in the convolution layer, thereby weakening or strengthening the influence of the ship features extracted by the neurons in this layer on the accuracy of ship type recognition.

The calculation formula for the change rate P _ia of the training recognition accuracy of the above ship types is as follows:

The experiment and analysis process for identifying the ship type in a specific application example of the method of the present invention are described in detail below:

The experimental platform for ship type identification in this application example is Windows 10 operating system, 16G RAM, the main frequency of the CPU processor is 3.4GHz, and the simulation platform is MATLAB (R2016 version). The experimental test ship of the present invention includes 7 types of ships, including container ships, oil tankers, chemical ships, liquefied natural gas ships (LNG), general cargo ships and bulk carriers. These 6 types of ships are common types of merchant ships. Category 7 ships are a collection of uncommon merchant ships, including timber ships, refrigerated ships and barges. In the training and test of this application example, the names of the first six ship types are recorded as the tags of various ship types, and the tag of the seventh type is "other ship types". The training and test sets include 11,760 pictures, of which 2,720 are container ship pictures, 1,320 are tanker pictures, 1600 are chemical tanker ships, 1,200 are LNG ship pictures, 2,850 are general ship pictures, and 2,070 are bulk ship pictures. A total of 1,560 pictures of 7 types of ships. Figure 3 shows a picture of a typical ship type.

Improper parameter settings will seriously affect the accuracy of the ship type recognition of the cascaded coarse-to-fine deep convolutional neural network (Coarse-to-Fine Cascaded Convolutional Neural Network, CFCCNN). Therefore, this application example first Tuning the parameters of the cascaded deep convolutional neural network from coarse to fine, the initial settings of the parameters are shown in Table 1. Batch sample size (batch size), weight decay rate (learning rate), learning rate (epoch) and network training times (epoch) are the key parameters of a cascaded deep convolutional neural network from coarse to fine. Here, the 1-type error rate is used to obtain the recognition accuracy of the ship type, and the parameter setting when the ship type has the highest recognition accuracy is set to the network, as the optimal cascade type from coarse to fine deep convolutional neural network parameter setting -The calculation expression for the class error rate is as follows:

among them:

N _s is the total number of ship images to be identified;

N _er is the total number of misidentified ship pictures.

Table 1.Initial tuning settings for cascaded deep convolutional neural network parameters from coarse to fine

参数parameter	初始值Initial value	步长Stride
N _bz* N _bz *	55	55
N _ep N _ep	5050	5050
σ _wd σ _wd	1×10 ^-4 1 × 10 ^-4	1×10 ^-4 1 × 10 ^-4
σ _lr σ _lr	2×10 ^-1 2 × 10 ^-1	1×10 ^-1 1 × 10 ^-1
ω _i1(i＝1，2，3) ω _i1 (i = 1, 2, 3)	0.50.5	__
ω _i2(i＝1，2，3) ω _i2 (i = 1, 2, 3)	0.50.5	__
(i＝1，2，3)(i = 1, 2, 3)	00	__
N _cl N _cl	200200	__
θ _c θ _c	0.10.1	__
N _d N _d	33	__
f _cl f _cl	3×33 × 3	__
f _pl f _pl	2×22 × 2	__

*: N _bz is the sample batch capacity; N _ep is the number of network trainings; σ _wd is the weight decay rate; σ _lr is the learning rate, and the parameters f _cl and f _pl are the convolution kernels of the convolution layer and the pooling layer, respectively. size.

Figures 4A to 4D show the distribution of the 1-type error rate for ship type identification when different parameters are set. Figures 4A to 4D show that irrational parameter settings will reduce the recognition accuracy of the network. FIG. 4A shows that the optimal size of the batch sample is 15, that is, the cascaded deep convolutional neural network of the present invention selects 15 pictures from the test set as a training set for training. When the batch size increased from 5 to 15 in equal steps, the 1-type error rate showed a significant downward trend. However, as the batch size increased from 15 to 50, the 1-type error rate increased rapidly. When the batch size is set to 50, the 1-type error rate reaches almost 40%. Based on the above analysis, the default batch size of the cascaded deep convolutional neural network from coarse to fine of the present invention is set to 15. FIG. 4B shows the change of the ship type recognition accuracy corresponding to different weight attenuation rates. When the weight attenuation rate is 5 × 10 ^-4 , the error rate of ship type identification is the smallest. In fact, when the weight attenuation rate is 5 × 10 ^-4 , the 1-type error rate of the cascaded deep convolutional neural network of the present invention is only 10%. Therefore, the default value of the weight decay rate is set to 5 × 10 ^-4 .

We gradually reduce the learning rate from 2 × 10 ^-1 to 2 × 10 ^-7 . Figure 4C shows that when the learning rate decreases from 2 × 10 ^-1 to 2 × 10 ^-3 , the 1-type error rate goes from 22%. Down to 15%. When the learning rate is reduced from 2 × 10 ^-3 to 2 × 10 ^-7 , the 1-type error rate remains basically unchanged, but the convergence time of CFCCNN increases significantly. Therefore, the optimal learning rate is set to 2 × 10 ^-3 . Referring to FIG. 4D, it can be seen that compared with the above three parameters, the change in the number of network training epochs has little effect on the accuracy of ship type recognition. In fact, when the network training number is equal to or more than 200, the recognition accuracy of the ship remains basically unchanged, so the default value of the network training number is 200.

Container ships, general cargo ships and oil tankers are common types of ships in water transportation. Correct identification of these three types of ships has important practical significance for the visual perception of intelligent ships and ensuring the safety of maritime traffic. Figures 5A-5C show the recognition results of these three types of typical ships. Figure 5A shows a typical container ship image and recognition results. The right side of FIG. 5A shows that the cascaded deep convolutional neural network according to the present invention has a probability of thinking that the ship in the image is a container ship and the probability of being a general cargo ship is 97.8%. The cascaded deep convolutional neural network of the present invention determines the ship type with the highest probability value as the recognized ship type. Therefore, it is determined that the input image is a container ship, which shows that the cascaded deep convolutional neural network of the present invention fully extracts and learns important and significant features of the container ship.

FIG. 5B shows that the cascaded deep convolutional neural network of the present invention tests a ship image as a general cargo ship. The recognition result of the cascaded deep convolutional neural network of the present invention is as follows. 5B right sub-picture. It is obvious that the ship in the test picture was equipped with a crane. It is only possible for general cargo ships and small bulk carriers to be equipped with restraint cranes. In addition, from the perspective of the ship, bulk carriers are basically equipped with hatch covers, while general cargo ships are not equipped with such facilities. It can be seen from the left sub-picture of FIG. 5B that the ship in the picture does not have a hatch. The above characteristics of the general cargo ship and the bulk cargo ship make the cascaded deep convolutional neural network of the present invention to easily distinguish between the general cargo ship and the bulk cargo ship. It can be known from FIG. 5B that the probability that the cascaded deep convolutional neural network of the present invention considers the ship to be a general cargo ship is 99.6%, and the probability that it is a bulk carrier is 0.4%. Therefore, it is determined that the ship type of the input image is a general cargo ship. This shows that the cascaded deep convolutional neural network of the present invention can correctly extract the characteristics of general cargo ships and bulk cargo ships.

The last sub-picture of FIG. 5C shows the recognition result of the oil tanker by the cascaded deep convolutional neural network from coarse to fine according to the present invention. Although the appearance of tankers and chemical tankers is similar, we can perceptually recognize that the pipelines on the decks of chemical tankers are more complicated than the pipelines on tanker decks, and the quantitative complexity cannot be used to describe the complexity of the pipelines of the two types of ships. However, the good generalization ability of the cascaded deep convolutional neural network from coarse to fine according to the present invention enables it to effectively grasp the complexity of two kinds of ship pipelines. In FIG. 5C, the cascaded deep convolutional neural network of the present invention considers that the probability that the ship belongs to an oil tanker is 96.4%, and the probability that it belongs to a chemical tanker is 3.6%. The above-mentioned CFCCNN ship confidence level distribution also validates our analysis. Therefore, it is determined that the ship type of the input image is a tanker. This shows that the cascaded deep convolutional neural network of the present invention can correctly extract the characteristics of general cargo ships and bulk cargo ships.

Fig. 6 shows that the cascaded deep convolutional neural network of the present invention has the highest recognition accuracy for a category 7 ship, and its recognition accuracy is as high as 93.3%. This is because the structural characteristics of Class 7 ships are more obvious. For example, as a member of the seventh category of ships, ro-ro ships usually carry small ships. Therefore, from the perspective of the image, the shape and structure corresponding to the ro-ro ship will be significantly different from other types of ships, and the cascaded deep-convolution neural network of the invention can easily obtain the ro-ro ship's This structural texture features. In fact, the shape and texture characteristics of other types of ships, such as wooden ships and refrigerated ships of the seventh type of ship, are clearly different from those of the other six types of ship. Therefore, the cascaded deep convolutional neural network of the present invention can obtain better accuracy when identifying a type 7 ship.

Figure 6 shows that the cascaded deep convolutional neural network of the present invention has a recognition accuracy of 90.7% for container ships, a recognition accuracy for general cargo ships of 86%, and a tanker recognition accuracy of 84.6%. Compared with the above ship types, the cascaded deep convolutional neural network of the present invention has lower recognition accuracy for chemical tankers and LNG tankers. This is because the cascaded deep convolutional neural network of the present invention recognizes part of the chemical tankers as tankers, and part of the liquefied natural gas ships as class 7 ships. Although the cascaded deep convolutional neural network of the present invention does not have high recognition accuracy for chemical tankers and LNG carriers, the average recognition rate for all ship types reaches 81.4%.

As shown in FIG. 7, this embodiment also uses the existing K-Nearest Neighbor (KNN), artificial neural network (ANN), random forest (RF), and traditional convolution. The neural network (convolutional neural network, CNN) method compares the recognition results of different ship types. The KNN algorithm and the ANN algorithm have the lowest recognition accuracy for chemical tankers, and their recognition accuracy is 29.8% and 28.1%, respectively. The recognition accuracy is the lowest, and its recognition accuracy is only 41.2%. The traditional CNN algorithm has a recognition accuracy of 61.3% and 63.2% for chemical tankers and LNG ships respectively, while the cascaded deep convolutional neural network of the present invention has a recognition accuracy of 65.6 for the above two ship types. % And 66.7%. The above-mentioned traditional methods (KNN, ANN and RF) and the accuracy of ship type recognition based on traditional CNN deep learning methods show that the traditional method cannot well extract the characteristics of different ship types, and the cascaded depth of the present invention ranges from coarse to fine The ship type recognition method of convolutional neural network can better find the depth characteristics of different ship types and obtain reliable ship recognition results.

The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the technical field based on the concept of the present invention through logic analysis, reasoning, or limited experiments based on the prior art should fall within the protection scope determined by the claims.

Claims

A cascaded deep to convolutional neural network ship type recognition method from coarse to fine, which is characterized by including the following steps:

S1: Input pictures of all ship types and corresponding picture tags, and perform coarse-level training on the deep convolutional neural network from coarse to fine, to obtain the setting parameters of the deep convolutional neural network from coarse to fine, and the settings of different ship types. Training recognition accuracy;

S2: Use the training to identify the ship type with the lowest accuracy picture to perform fine-level training on the deep to convolutional neural network from coarse to fine. If the deep convolutional neural network has not reached the preset convergence condition, return to step S1 to continue Training, otherwise execute step S3;

S3: Perform type recognition on the ship in the picture, and output the recognition result of the ship type.
The method according to claim 1, wherein the step S1 comprises the following steps:

S11: crop the originally input ship type picture to a fixed size, match the input ship type picture and corresponding picture tag, and obtain a formatted ship image and image tag;

S12: using the formatted ship image and image label to train a cascaded deep convolutional neural network from coarse to fine to obtain the setting parameters of the cascaded deep convolutional neural network from coarse to fine, Extract depth characteristics of different ship types;

S13: Obtain the confidence level distribution of the input ship type picture according to the depth feature, and output the ship type corresponding to the maximum confidence value as a single training recognition result;

S14: According to the recognition result of the single training, compare the true type of the ship corresponding to the picture tag to obtain the training recognition accuracy of different ship types.
The cascaded deep to convolutional neural network ship type recognition method according to claim 2, wherein the step S12 comprises the following steps:

S121: The cascaded deep-to-fine convolutional neural network uses convolution layers to extract ship features; the ship features include: low-level ship features, including ship textures, contours, and corner points; advanced ship features, according to Different types of ships are obtained by correspondingly abstracting the characteristics of low-level ships;

S122: The cascaded deep convolutional neural network from coarse to fine uses the pooling layer to reduce the dimension of the ship features and learn;

S123: The cascaded deep to convolutional neural network uses a local response normalization layer to increase the local response of the ship feature extracted by the convolution layer to randomly assign a larger response value and extract Generalize ship characteristics.
The cascaded deep to convolutional neural network ship type recognition method according to claim 3, wherein the step S13 comprises the following steps:

S131: The cascaded deep-to-fine convolutional neural network uses a fully connected layer to map the generalized ship feature to a single ship feature vector, and its expression is as follows:

F out = Θ × F in

Among them: F out is a single ship feature vector output from the fully connected layer, which has a total of n 1 elements; F in is the input generalized ship feature, its dimension is n 2 +1; Θ is the F in and F out 's connection matrix with dimensions n 1 × (n 2 +1);

S132: The cascaded deep-to-fine convolutional neural network uses a loss layer to generate a probability vector based on the single ship feature vector as an input ship image, and the elements in the vector represent the probability of the type of ship The calculation expression is as follows:

Among them, F p is a single ship feature vector of the loss layer; v j is a weight corresponding to the j-th ship type when calculating a ship probability vector.
The method for identifying a ship type according to a cascaded deep convolutional neural network from coarse to fine according to claim 1 or 2 or 3 or 4, characterized in that, in step S14, the training type recognition accuracy e 1t The calculation expression is as follows:

Among them: N s is the total number of ship pictures to be identified; N er is the total number of ship pictures with incorrect type recognition.
The method for identifying a ship type of a cascaded deep convolutional neural network from coarse to fine according to claim 1, wherein the step S2 comprises the following steps:

S21: Acquire all the training pictures of the ship type according to the pictures of the ship type with the lowest training recognition accuracy, as the input samples for the fine-level training of the coarse to fine deep convolutional neural network;

S22: Use a random heuristic selection method to select one of the data enhancement, selective discarding method, and selective connection method as a random regularization mechanism;

S23: According to the setting parameters, use the random regularization mechanism to train the coarse to fine deep convolutional neural network, and obtain the refined parameters and the adjusted coarse to fine deep convolution. Neural Networks;

S24: The coarse-to-fine deep convolutional neural network adjusted according to the refinement parameters is used to re-identify the ship type of the picture. If the training type recognition accuracy change rate is less than a preset threshold, the training process ends; if the ship type The type of training recognition accuracy change rate is greater than the preset threshold, the coarse-to-fine deep convolutional neural network completes the current fine-level training, and returns to step S1 to continue training.
The method according to claim 6, wherein the data enhancement includes horizontal / vertical flipping of the training picture, color change, and / Or randomly change the size of the training picture.
The method according to claim 6, wherein the selective discarding method sleeps some neurons of the convolutional layer with a preset probability, and All neurons in the convolutional layer sleep or cease to sleep with the same probability.
The method according to claim 6, wherein the selective connection method is to randomly modify the weights of the neurons in the convolutional layer, so that Weaken or strengthen the influence of ship features extracted by neurons in this layer on the accuracy of ship type recognition.
The cascaded deep to convolutional neural network ship type recognition method according to claim 6, wherein the calculation rate of the change rate P ia of the training recognition accuracy of the ship type is as follows:

Among them: A jc is the accuracy of class j ship type recognition in step S1; A jf is the accuracy of class j ship type recognition in step S2.