A method and device for identifying dangerous vessels

Abstract

The present invention discloses a method and device for identifying dangerous vessels. The method includes: obtaining an image of a vessel in a channel to be measured; inputting the image of the vessel in the channel to be measured into a preselected constructed target fully convolutional network to obtain a dangerous vessel identification result; the dangerous vessel identification result determines that the vessel to be measured in the image of the vessel in the channel to be measured can pass through the transmission line in the image of the vessel in the channel to be measured; wherein, the construction method of the target fully convolutional network includes: obtaining a dataset of optical remote sensing images of the channel; dividing the dataset of optical remote sensing images of the channel into an image training set and an image test set; inputting all the remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; inputting all the remote sensing image data in the image test set into the trained fully convolutional network, and combining with the non-maximum suppression algorithm to obtain the target fully convolutional network. The problem that traditional dangerous vessels cannot maintain high-precision identification in different identification scenarios is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method and device for identifying dangerous vessels. Background Art

[0002] At present, due to the difficulty of the mast height of large vessels in adapting to the sag reserve space of the transmission line designed in the initial stage, when a large vessel approaches the transmission line, it may cause line tripping due to insufficient safety distance and discharge, or even hang up the conductor and cause large-area power outages. In the case of a long time required for emergency repair and power restoration and a large amount of manpower and material resources, it will bring huge economic losses and cause negative social impacts.

[0003] To solve the problems caused by insufficient safety distance, it is necessary to evaluate and give early warnings in real time and accurately the safety of target vessels during navigation. In the evaluation process, the identification of dangerous vessels is particularly important. However, traditional dangerous vessel identification has at least the following disadvantages: The image recognition method based on gray-scale statistics has poor results in identifying complex sea backgrounds; The image recognition method based on vessel wakes has poor wake stability and low recognition rate; The image recognition method based on shape features has poor results and low recognition rate in identifying complex sea backgrounds.

[0004] In summary, it is of great significance to provide a method for identifying dangerous vessels that can achieve high precision in different recognition scenarios. Summary of the Invention

[0005] The present invention provides a method and device for identifying dangerous vessels, which are used to solve the problem that traditional dangerous vessels cannot maintain high-precision identification in different recognition scenarios.

[0006] In a first aspect, a method for identifying dangerous vessels provided by the present invention includes:

[0007] Obtain a dataset of optical remote sensing images of waterways;

[0008] Divide the dataset of optical remote sensing images of waterways into an image training set and an image test set;

[0009] Input all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network;

[0010] Input all remote sensing image data in the image test set into the trained fully convolutional network, and combine with the non-maximum suppression algorithm to obtain a target fully convolutional network;

[0011] Obtain an image of a vessel in a waterway to be measured;

[0012] Input the image of the vessel in the channel to be measured into the target fully convolutional network to obtain the identification result of dangerous vessels; the identification result of dangerous vessels is used to determine whether the vessel to be measured in the image of the vessel in the channel to be measured can pass through the transmission line in the image of the vessel in the channel to be measured.

[0013] Optionally, after obtaining the channel optical remote sensing image dataset, it further includes:

[0014] Preprocess the remote sensing image data in the channel optical remote sensing image dataset.

[0015] Optionally, dividing the channel optical remote sensing image dataset into an image training set and an image test set includes:

[0016] Uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain the cropped remote sensing images; the first image set is arbitrarily obtained from the channel optical remote sensing image dataset;

[0017] Mark the dangerous vessels in the cropped remote sensing images, and label the dangerous vessels with the minimum bounding box of the matrix to obtain the image training set;

[0018] Define the image set other than the image training set in the channel optical remote sensing image data as the image test set.

[0019] Optionally, preprocessing the remote sensing image data in the channel optical remote sensing image dataset includes:

[0020] Perform image denoising and gray histogram enhancement on all remote sensing image data in the channel optical remote sensing image dataset;

[0021] For the night remote sensing image data and thick fog remote sensing image data in the channel optical remote sensing image dataset, use the multi-scale Retinex algorithm for image enhancement.

[0022] Optionally, input all the remote sensing image data in the image test set into the trained fully convolutional network, and combine the non-maximum suppression algorithm to obtain the target fully convolutional network, including:

[0023] S1. Determine the vessel identification result of the remote sensing image data in the image test set through the trained fully convolutional network;

[0024] S2. Traverse all the remaining minimum bounding boxes except the minimum bounding box with the highest confidence, and calculate the IOU value between any selected target minimum bounding box and the remaining minimum bounding boxes in turn; the minimum bounding box with the highest confidence is obtained by sorting the confidences of all the vessel identification results;

[0025] S3. Eliminate all the remaining minimum bounding boxes corresponding to the IOU values below the preset IOU value;

[0026] S4. Determine whether there are new minimum bounding boxes. If so, return to step S2; if not, obtain the target fully convolutional network.

[0027] In a second aspect, the present invention further provides an identification device for dangerous vessels, including:

[0028] A dataset acquisition module, configured to acquire a dataset of channel optical remote sensing images;

[0029] A dataset partitioning module, configured to partition the dataset of channel optical remote sensing images into an image training set and an image test set;

[0030] A network training module, configured to input all the remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network;

[0031] A target network generation module, configured to input all the remote sensing image data in the image test set into the trained fully convolutional network, and combine with the non-maximum suppression algorithm to obtain the target fully convolutional network;

[0032] A vessel image acquisition module, configured to acquire an image of a vessel in a channel to be measured;

[0033] An identification result determination module, configured to input the image of the vessel in the channel to be measured into a preselected target fully convolutional network to obtain an identification result of a dangerous vessel; the identification result of the dangerous vessel determines whether the vessel to be measured in the image of the vessel in the channel to be measured can pass through the transmission line in the image of the vessel in the channel to be measured.

[0034] Optionally, it further includes:

[0035] A preprocessing module, configured to preprocess the remote sensing image data in the dataset of channel optical remote sensing images.

[0036] Optionally, the dataset partitioning module includes:

[0037] A cropping sub-module, configured to uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain cropped remote sensing images; the first image set is arbitrarily obtained from the dataset of channel optical remote sensing images;

[0038] A labeling sub-module, configured to label the dangerous vessels in the cropped remote sensing images, and label the dangerous vessels with matrix minimum bounding boxes to obtain the image training set;

[0039] A test set determination sub-module, which is used to define, in the optical remote sensing image data of the waterway, the image set other than the image training set as the image test set.

[0040] Optionally, the preprocessing module includes:

[0041] A preprocessing sub-module, which is used to perform image noise reduction and gray histogram enhancement on all remote sensing image data in the optical remote sensing image dataset of the waterway;

[0042] An image enhancement sub-module, which is used to perform image enhancement on the night remote sensing image data and thick fog remote sensing image data in the optical remote sensing image dataset of the waterway by using the multi-scale Retinex algorithm.

[0043] The recognition result determination module includes:

[0044] A recognition result determination sub-module, which is used to determine the ship recognition result of the remote sensing image data in the image test set through the trained fully convolutional network;

[0045] A traversal sub-module, which is used to traverse all the remaining minimum bounding boxes except the minimum bounding box with the highest confidence, and calculate the IOU value between any selected target minimum bounding box and the remaining minimum bounding boxes in turn; the minimum bounding box with the highest confidence is obtained by sorting the confidences of all the ship recognition results;

[0046] An elimination sub-module, which is used to eliminate the remaining minimum bounding boxes corresponding to all IOU values lower than the preset IOU value;

[0047] A judgment sub-module, which is used to judge whether there is a new minimum bounding box. If so, it returns to execute the traversal sub-module. If not, it obtains the target fully convolutional network.

[0048] The third aspect of the present application provides an electronic device, which includes a processor and a memory;

[0049] The memory is used to store program codes and transmit the program codes to the processor;

[0050] The processor is used to execute the method for identifying dangerous ships according to the instructions in the program codes in the first aspect. [[ID=I35]]

[0051] The fourth aspect of the present application provides a computer-readable storage medium, which is used to store program codes, and the program codes are used to execute the method for identifying dangerous ships according to the instructions in the first aspect.

[0052] From the above technical solutions, it can be seen that the present invention has the following advantages:

[0053] The present invention obtains an image of a vessel in a channel to be measured; inputs the image of the vessel in the channel to be measured into a preselected constructed target fully convolutional network to obtain a recognition result of a dangerous vessel; the recognition result of the dangerous vessel determines that the vessel to be measured in the image of the vessel in the channel to be measured can pass through a transmission line in the image of the vessel in the channel to be measured; wherein, the construction method of the target fully convolutional network includes: obtaining a dataset of channel optical remote sensing images; dividing the dataset of channel optical remote sensing images into an image training set and an image test set; inputting all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; inputting all remote sensing image data in the image test set into the trained fully convolutional network, and combining with a non-maximum suppression algorithm to obtain the target fully convolutional network. It solves the problem that traditional dangerous vessels cannot maintain high-precision recognition in different recognition scenarios. Description of the Drawings

[0054] 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 use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a flowchart of the steps of Embodiment 1 of a method for recognizing a dangerous vessel according to the present invention.

[0056] Figure 2 It is a flowchart of the steps of Embodiment 2 of a method for recognizing a dangerous vessel according to the present invention.

[0057] Figure 3 It is a schematic diagram of the reference residual network ResNet in the embodiments of the present invention.

[0058] Figure 4 It is a schematic diagram of a new structure of the residual network in the embodiments of the present invention.

[0059] Figure 5 It is a structural flowchart of the fully convolutional neural network in the embodiments of the present invention.

[0060] Figure 6 It is a structural block diagram of an embodiment of a device for recognizing a dangerous vessel according to the present invention. Detailed Embodiments

[0061] The embodiments of the present invention provide a method and a device for recognizing a dangerous vessel, which are used to solve the problem that traditional dangerous vessels cannot maintain high-precision recognition in different recognition scenarios.

[0062] In order to make the objectives, features, and advantages of the present invention more obvious and understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described below are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0063] Please refer to Figure 1 , Figure 1 which is a flowchart of the steps of Embodiment 1 of a method for identifying dangerous vessels according to the present invention, and specifically may include the following steps:

[0064] Step S101, obtaining an image of a vessel in the channel to be measured;

[0065] Step S102, inputting the image of the vessel in the channel to be measured into a pre-constructed target fully convolutional network to obtain a dangerous vessel identification result; the dangerous vessel identification result determines that the vessel to be measured in the image of the vessel in the channel to be measured can pass through the transmission line in the image of the vessel in the channel to be measured; wherein, the method for constructing the target fully convolutional network includes: obtaining a dataset of channel optical remote sensing images; dividing the dataset of channel optical remote sensing images into an image training set and an image test set; inputting all the remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; inputting all the remote sensing image data in the image test set into the trained fully convolutional network, and combining with the non-maximum suppression algorithm to obtain the target fully convolutional network.

[0066] In the embodiment of the present invention, by obtaining an image of a vessel in the channel to be measured; inputting the image of the vessel in the channel to be measured into a pre-constructed target fully convolutional network to obtain a dangerous vessel identification result; the dangerous vessel identification result determines that the vessel to be measured in the image of the vessel in the channel to be measured can pass through the transmission line in the image of the vessel in the channel to be measured; wherein, the method for constructing the target fully convolutional network includes: obtaining a dataset of channel optical remote sensing images; dividing the dataset of channel optical remote sensing images into an image training set and an image test set; inputting all the remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; inputting all the remote sensing image data in the image test set into the trained fully convolutional network, and combining with the non-maximum suppression algorithm to obtain the target fully convolutional network. The problem that traditional dangerous vessels cannot maintain high-precision identification in different identification scenarios is solved.

[0067] Please refer to Figure 2 , which is a flowchart of the steps of Embodiment 2 of a method for identifying dangerous vessels according to the present invention, and specifically includes:

[0068] Step S201, obtain a dataset of optical remote sensing images of the waterway;

[0069] Step S202, preprocess the remote sensing image data in the dataset of optical remote sensing images of the waterway;

[0070] In an optional embodiment, preprocessing the remote sensing image data in the dataset of optical remote sensing images of the waterway includes:

[0071] Perform image denoising and gray histogram enhancement on all remote sensing image data in the dataset of optical remote sensing images of the waterway;

[0072] For the night remote sensing image data and thick fog remote sensing image data in the dataset of optical remote sensing images of the waterway, use the multi-scale Retinex algorithm for image enhancement.

[0073] In the present invention, the optical remote sensing images of the waterway are sampled respectively in the daytime, at night, and in heavy fog by an image recognition device installed at the pole tower to form a dataset of optical remote sensing images of the waterway, and the dataset of optical remote sensing images of the waterway is denoised and gray histogram enhanced. At the same time, for the night images and heavy fog images in the dataset of optical remote sensing images of the waterway, the multi-scale Retinex algorithm is used for enhancement, that is, a Gaussian is used to filter an image at different scales, and then the filtered results at different scales are averaged and weighted to obtain the estimated illumination image, and its formula is as follows:

[0074]

[0075] Among them, N is the number of scale parameters, and when N takes 3, the original image is filtered by three Gaussian filters with different scales; ω

[0080] ,

[0076] ,

[0079] ,

[0075] , k , , , , , , , k , , ,

[0078] ,

[0077] is the weight coefficient when the k-th scale is weighted, and it needs to satisfy the following formula:

[0076]

[0077] Finally, F k (x, y) is the Gaussian filtering function at the k-th scale, that is:

[0078]

[0079] Step S203, uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain the cropped remote sensing images; the first image set is arbitrarily obtained from the dataset of optical remote sensing images of the waterway;

[0080] Step S204, identify the dangerous vessels in the cropped remote sensing image, and label the dangerous vessels with the minimum matrix bounding box to obtain the image training set;

[0081] In the embodiment of the present invention, part of the images in the waterway optical remote sensing image dataset are divided into the training set.

[0082] In a specific implementation, the remote sensing images in the training set are cropped into a unified size so that there are only target objects in the image without other interfering objects. Then, the dangerous vessels in the cropped remote sensing image are identified, the ultra-high vessels such as sand dredgers are identified as dangerous vessels, and the vessels are labeled with the minimum matrix bounding box.

[0083] Step S205, define the image set other than the image training set in the waterway optical remote sensing image data as the image test set;

[0084] In the embodiment of the present invention, the ratio of the number of remote sensing images in the image training set to the number of remote sensing images in the image test set is 3:1.

[0085] Step S206, input all the remote sensing image data in the image training set into the initial fully convolutional network for training to obtain the trained fully convolutional network;

[0086] Please refer to Figure 3 , Figure 3 which is a schematic diagram of the reference residual network ResNet in the embodiment of the present invention. In the embodiment of the present invention, the residual block mechanism has two layers, and the expression is as follows:

[0087] F = W2σ(W1x),

[0088] Then, the output y is obtained through Figure 2 the bypass shortcut and the second relu:

[0089] y = F(x, {W i}) + x;

[0090] To save the computing cost, the convolutional layer is replaced with 1×1 + 3×3 + 1×1, as Figure 4 shown, Figure 4 which is a schematic diagram of the new structure of the residual network in the embodiment of the present invention. The middle 3×3 convolutional layer in this structure first reduces the calculation based on the activation function under a dimensionality reduction 1×1 convolutional layer, and then restores it under another 1×1 convolutional layer based on another activation function, which not only maintains the accuracy but also reduces the amount of calculation.

[0091] In the embodiments of the present invention, in order to improve the effect of feature extraction, the constructed initial fully convolutional network adopts a region-based fully convolutional network (R-FCN), extracts image features through a ResNet residual network, then calculates ship candidate boxes based on the obtained multi-scale feature maps, and then obtains the saliency feature map within the candidate boxes, and takes the minimum rectangular bounding of the saliency region as the detection result.

[0092] Please refer to Figure 5 , Figure 5 which is the structural flowchart of the fully convolutional neural network in the embodiments of the present invention. In specific implementation, the image first passes through the ResNet residual network to extract image features, and the output of the convolutional layer conv is used as the input of the RPN to extract the regions of interest Rols. At the same time, the output of conv will generate a position-sensitive score map with a dimension of k×2(c + 1) through the feature image, and then pooling is performed in combination with the previously extracted Rols, and then a classification result is obtained for each Rol.

[0093] At the same time, in order to more easily obtain an end-to-end R-CFN network, during the training process, a pre-trained ResNet network and pre-computed candidate regions are used.

[0094] In addition, in order to obtain a fully convolutional network with high generalization ability, a series of image augmentation operations can be performed on the remote sensing images in the training set, including random scaling and random flipping, etc. In specific implementation, key hyperparameters such as an appropriate batch size and learning rate can be preset in advance. When the hardware permits, generally try to set a larger batch-size as much as possible, but the larger the batch-size, the more likely the model will fall into a local minimum or a saddle point. If the learning rate is too large, the modification range of the network parameters after each iteration will be too large, which will cause the loss function to oscillate and not converge; while if the learning rate is set too small, the convergence speed of the model will be too slow, significantly increasing the training time.

[0095] Step S207, determine the ship recognition result of the remote sensing image data in the image test set through the trained fully convolutional network;

[0096] Step S208, traverse all the remaining minimum bounding boxes except the minimum bounding box with the highest confidence, and calculate the IOU value between any selected target minimum bounding box and the remaining minimum bounding boxes in turn; the minimum bounding box with the highest confidence is obtained by sorting the confidences of all the ship recognition results;

[0097] Step S209, eliminate all the remaining minimum bounding boxes corresponding to the IOU values lower than the preset IOU value;

[0098] Step S210, determine whether there are new minimum bounding boxes. If so, return to Step S208. If not, obtain the target fully convolutional network;

[0099] In the embodiment of the present invention, when inputting the test set image into the fully convolutional network in Step 206, the confidence level of the ship recognized by the network and the minimum bounding box corrected by the saliency feature map can be obtained. The confidence level threshold is taken as 0.6, that is, the recognition results with a confidence level less than 0.6 are discarded.

[0100] In specific implementation, in order to avoid multiple detections of the same target by the network, the non-maximum suppression algorithm is adopted to ensure that only one detection is obtained for each object. The steps of the non-maximum suppression algorithm are as follows: (1) Sort the confidence levels of all ship recognition results, and select the minimum bounding box with the highest confidence level; (2) Traverse the remaining minimum bounding boxes, calculate the intersection over union (IOU) between the selected minimum bounding box and each of them in turn, and delete the minimum bounding boxes with an IOU greater than the set threshold; (3) Continue to select the minimum bounding box in descending order of confidence level, and loop Steps (1) and (2) until no new minimum bounding box is deleted and the loop stops.

[0101] Step S211, obtain the ship image of the channel to be measured;

[0102] Step S212, input the ship image of the channel to be measured into the preselected target fully convolutional network to obtain the recognition result of dangerous ships; the recognition result of dangerous ships determines whether the ship to be measured in the ship image of the channel to be measured can pass through the transmission line in the ship image of the channel to be measured.

[0103] In an identification method for dangerous vessels provided by an embodiment of the present invention, a vessel image of a channel to be measured is obtained; the vessel image of the channel to be measured is input into a pre-selected constructed target fully convolutional network to obtain a dangerous vessel identification result; the dangerous vessel identification result determines that the vessel to be measured in the vessel image of the channel to be measured can pass through the power transmission line in the vessel image of the channel to be measured; wherein, the construction method of the target fully convolutional network includes: obtaining a dataset of channel optical remote sensing images; dividing the dataset of channel optical remote sensing images into an image training set and an image test set; inputting all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; inputting all remote sensing image data in the image test set into the trained fully convolutional network, and combining with a non-maximum suppression algorithm to obtain the target fully convolutional network. This solves the problem that traditional dangerous vessels cannot maintain high-precision identification in different identification scenarios. The target full convolution network involved in the embodiment of the present invention adopts the R-FCN network structure and uses ResNet as the feature extraction network, which speeds up and improves the learning speed, and can also ensure that the situation of gradient explosion is avoided when the network is deepened, and at the same time, network degradation is also avoided; at the same time, in order to avoid multiple detections of the same target by the network, the non-maximum suppression algorithm is used to ensure that only one detection is obtained for each object; in addition, the multi-scale Retinex algorithm is used to enhance the images in the case of night and heavy fog, which greatly improves the recognition rate of dangerous vessels in the case of night and heavy fog, and solves the problem that traditional dangerous vessels cannot maintain high-precision identification in different identification scenarios.

[0104] Please refer to Figure 6 , which shows a structural block diagram of an embodiment of an identification device for dangerous vessels, including the following modules:

[0105] The dataset acquisition module 401 is used to acquire a dataset of channel optical remote sensing images;

[0106] The dataset division module 402 is used to divide the dataset of channel optical remote sensing images into an image training set and an image test set;

[0107] The network training module 403 is used to input all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network;

[0108] The target network generation module 404 is used to input all remote sensing image data in the image test set into the trained fully convolutional network, and combine with a non-maximum suppression algorithm to obtain the target fully convolutional network;

[0109] The vessel image acquisition module 405 is used to acquire a vessel image of a channel to be measured;

[0110] An identification result determination module 406 is configured to input the image of the vessel in the to-be-detected waterway into a preselected target fully convolutional network to obtain a dangerous vessel identification result; the dangerous vessel identification result determines whether the vessel to be detected in the image of the vessel in the to-be-detected waterway can pass through the transmission line in the image of the vessel in the to-be-detected waterway.

[0111] In an alternative embodiment, it further includes:

[0112] A preprocessing module is configured to preprocess the remote sensing image data in the waterway optical remote sensing image dataset.

[0113] In an alternative embodiment, the dataset division module 402 includes:

[0114] A cropping sub-module is configured to uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain the cropped remote sensing images; the first image set is arbitrarily obtained from the waterway optical remote sensing image set;

[0115] A labeling sub-module is configured to identify the dangerous vessels in the cropped remote sensing images and label the dangerous vessels with the minimum bounding box of the matrix to obtain the image training set;

[0116] A test set determination sub-module is configured to define the image set other than the image training set in the waterway optical remote sensing image data as the image test set.

[0117] In an alternative embodiment, the preprocessing module includes:

[0118] A preprocessing sub-module is configured to perform image denoising and gray histogram enhancement on all the remote sensing image data in the waterway optical remote sensing image dataset;

[0119] An image enhancement sub-module is configured to perform image enhancement on the night remote sensing image data and the thick fog remote sensing image data in the waterway optical remote sensing image dataset by using a multi-scale Retinex algorithm.

[0120] In an alternative embodiment, the identification result determination module 406 includes:

[0121] An identification result determination sub-module is configured to determine the vessel identification result of the remote sensing image data in the image test set through the trained fully convolutional network;

[0122] A traversal sub-module is configured to traverse all the remaining minimum bounding boxes except the minimum bounding box with the highest confidence, and calculate the IOU value between any selected target minimum bounding box and the remaining minimum bounding boxes in sequence; the minimum bounding box with the highest confidence is obtained by sorting the confidences of all the vessel identification results;

[0123] A culling sub-module, configured to cull all remaining minimum bounding boxes corresponding to IOU values lower than a preset IOU value;

[0124] A determination sub-module, configured to determine whether there is a new minimum bounding box. If so, it returns to execute the traversal sub-module. If not, it obtains the target fully convolutional network.

[0125] This application also provides an electronic device, which includes a processor and a memory;

[0126] The memory is used to store program codes and transmit the program codes to the processor;

[0127] The processor is configured to execute the method for identifying dangerous vessels in the above method embodiments according to the instructions in the program codes.

[0128] This application also provides a computer-readable storage medium, which is used to store program codes, and the program codes are used to execute the method for identifying dangerous vessels in the above method embodiments.

[0129] In several embodiments provided by this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical or other forms.

[0130] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0131] In addition, in each embodiment of this application, the functional units can be integrated in one processing unit, or each unit exists physically alone, or two or more units can be integrated in one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0132] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (English full name: Read-Only Memory, English abbreviation: ROM), random access memories (English full name: Random Access Memory, English abbreviation: RAM), magnetic disks, or optical discs.

[0133] As described above, the above embodiments are only used to illustrate the technical solution of this application, rather than to limit it; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of this application.

Claims

1. A method for identifying dangerous vessels, characterized in that Including: Obtain a dataset of optical remote sensing images of waterways; Divide the dataset of optical remote sensing images of waterways into an image training set and an image test set; Input all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; wherein, the initial fully convolutional network adopts the structure of a region-based fully convolutional network R-FCN and uses ResNet as the feature extraction network; Input all remote sensing image data in the image test set into the trained fully convolutional network, and combine with the non-maximum suppression algorithm to obtain a target fully convolutional network; Obtain an image of a ship in a waterway to be measured; Input the image of the ship in the waterway to be measured into the target fully convolutional network to obtain a dangerous ship recognition result; the dangerous ship recognition result is used to determine whether the ship to be measured in the image of the ship in the waterway to be measured can pass through the transmission line in the image of the ship in the waterway to be measured; After obtaining the dataset of optical remote sensing images of waterways, it further includes: Preprocess the remote sensing image data in the dataset of optical remote sensing images of waterways, specifically including: performing image denoising and gray histogram enhancement on all remote sensing image data in the dataset of optical remote sensing images of waterways; For the night remote sensing image data and thick fog remote sensing image data in the dataset of optical remote sensing images of waterways, use the multi-scale Retinex algorithm for image enhancement.

2. The identification method of a dangerous vessel according to claim 1, wherein Dividing the dataset of optical remote sensing images of waterways into an image training set and an image test set includes: Uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain cropped remote sensing images; the first image set is arbitrarily obtained from the dataset of optical remote sensing images of waterways; Mark the dangerous ships in the cropped remote sensing images, and label the dangerous ships with the smallest matrix bounding box to obtain the image training set; Define the image set other than the image training set in the optical remote sensing image data of the waterway as the image test set.

3. The method for identifying a dangerous vessel according to claim 1, characterized in that Input all remote sensing image data in the image test set into the trained fully convolutional network, and combine with the non-maximum suppression algorithm to obtain the target fully convolutional network, including: S1. Determine the ship recognition result of the remote sensing image data in the image test set through the trained fully convolutional network; S2. Traverse all the remaining smallest bounding boxes except the smallest bounding box with the highest confidence, and calculate the IOU value of any selected target smallest bounding box and the remaining smallest bounding boxes in turn; the smallest bounding box with the highest confidence is obtained by sorting the confidences of all the ship recognition results; S3. Eliminate all the remaining smallest bounding boxes corresponding to the IOU values lower than the preset IOU value; S4. Determine whether there is a new smallest bounding box. If so, return to step S2. If not, obtain the target fully convolutional network.

4. An identification device for dangerous vessels, characterized in that, Including: A dataset acquisition module for obtaining a dataset of optical remote sensing images of waterways; A dataset division module for dividing the dataset of optical remote sensing images of waterways into an image training set and an image test set; A network training module, which is used to input all remote sensing image data in the image training set into an initial fully convolutional network for training to obtain a trained fully convolutional network; wherein, the initial fully convolutional network adopts the structure of a region-based fully convolutional network R-FCN and uses ResNet as the feature extraction network; A target network generation module, which is used to input all remote sensing image data in the image test set into the trained fully convolutional network and combine it with the non-maximum suppression algorithm to obtain a target fully convolutional network; A ship image acquisition module, which is used to acquire ship images of the channel to be measured; An identification result determination module, which is used to input the ship image of the channel to be measured into the target fully convolutional network to obtain a dangerous ship identification result; the dangerous ship identification result determines whether the ship to be measured in the ship image of the channel to be measured can pass through the transmission line in the ship image of the channel to be measured; It further includes: A preprocessing module, which is used to preprocess the remote sensing image data in the channel optical remote sensing image dataset; The preprocessing module includes: A preprocessing sub-module, which is used to perform image denoising and gray histogram enhancement on all remote sensing image data in the channel optical remote sensing image dataset; An image enhancement sub-module, which is used to perform image enhancement on the night remote sensing image data and thick fog remote sensing image data in the channel optical remote sensing image dataset by using the multi-scale Retinex algorithm.

5. The identification device for dangerous vessels according to claim 4, characterized in that, The dataset division module includes: A cropping sub-module, which is used to uniformly crop the remote sensing images in the first image set according to a predetermined size to obtain cropped remote sensing images; the first image set is arbitrarily obtained from the channel optical remote sensing image set; A labeling sub-module, which is used to label the dangerous ships in the cropped remote sensing images and label the dangerous ships with the smallest matrix bounding box to obtain the image training set; A test set determination sub-module, which is used to define the image set other than the image training set in the channel optical remote sensing image data as the image test set.

6. The identification device for dangerous vessels according to claim 4, characterized in that The identification result determination module includes: An identification result determination sub-module, which is used to determine the ship identification result of the remote sensing image data in the image test set through the trained fully convolutional network; A traversal sub-module, which is used to traverse all the remaining smallest bounding boxes except the smallest bounding box with the highest confidence, and calculate the IOU value between any selected target smallest bounding box and the remaining smallest bounding boxes in turn; the smallest bounding box with the highest confidence is obtained by sorting the confidences of all the ship identification results; An elimination sub-module, which is used to eliminate all the remaining smallest bounding boxes corresponding to the IOU values lower than the preset IOU value; A judgment sub-module, which is used to judge whether there is a new smallest bounding box. If so, it returns to execute the traversal sub-module. If not, it obtains the target fully convolutional network.

Images