A ship warning image recognition system

Abstract

The present invention discloses a ship warning image recognition system, belonging to the technical field of image processing. It solves the problem that in the case of thin fog, the observers on the ship can directly observe the water area with the naked eye without the need for defogging processing. However, when the fog is thick, the ship defogging system can only play an appropriate role in assisting the crew and cannot provide early warnings to the crew during the process of the fog appearing and thickening. The system includes an image acquisition module, an image defogging module, a data processing module, and a warning module. The image acquisition module is used to collect the front view of the ship and feed it back to the image defogging module. The image defogging module is used to defog the front view of the ship and then provide the post-fog image and feed it back to the data processing module. The present invention collects the front view of the ship in the fog through the image acquisition module, filters out the parts obscured by the fog, retains the clear content for recognition, calculates the visibility, warns the crew, avoids misjudgment, and ensures navigation safety.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a ship warning image recognition system. Background Art

[0002] Ships are watercraft with a long history. They are supported by the buoyancy of water and sail on the water surface by means of various power methods such as sails, steam engines, diesel engines, or nuclear power. From the wooden ships in ancient Egypt to modern giant cargo ships, ships have undergone tremendous development in terms of design, scale, and function. They are not only used for transporting goods and passengers but also undertake important tasks such as military, scientific research, and exploration. With the progress of technology, modern ships pay more and more attention to environmental protection and efficiency. The application of innovative technologies such as liquefied natural gas (LNG)-powered ships and sail-assisted propulsion systems has enabled the ship industry to continuously move forward on the path of sustainable development.

[0003] Currently, with the increasing prosperity of water transportation, its order has also been improved, which has led to an increasingly heavy burden on the staff responsible for maintaining water transportation order. Ship monitoring is one of the important responsibilities. However, in autumn and winter, whether in the early morning, evening, or daytime, thick water mist often shrouds the water surface, making ships obscured and difficult to monitor effectively, which poses a challenge to the clear detection of ships on the river.

[0004] To address this problem, it is basically necessary to rely on a system. Currently, in practice, a ship de - fogging and restoration method and system with the application number CN202410029086.4 is adopted. Its technical key points include: S1. Obtain clear - airspace ship pictures and ship pictures in water - mist airspaces of different levels. Perform Fourier transform on each airspace ship picture to obtain the corresponding frequency - domain ship picture. Add all the corresponding airspace ship pictures and frequency - domain ship pictures and input them into the original water - mist concentration classification model for multiple rounds of model training to obtain the target water - mist concentration classification model. S2. Obtain the original training set. The original training set is divided into multiple groups, and each group includes clear ship pictures and corresponding randomly - leveled water - mist ship pictures. Add a water - mist concentration level guidance task to the original Unet model to obtain the class Unet model. S3. Randomly add a kind of Gaussian noise to all the clear ship pictures in the original training set and input all the water - mist ship pictures into the target water - mist concentration classification model to judge the water - mist level, and then use them as the first - round training set and input them into the class Unet model for the first - round model training to obtain the first - round noise value and the first - round updated class Unet model. Calculate the first noise loss value based on the noise values of all clear ship pictures, the first - round noise value, and the noise values of all water - mist ship pictures. S4. Randomly add a kind of Gaussian noise and the previous - round noise value to all the clear ship pictures in the original training set, and subtract the previous - round noise value from all the water - mist ship pictures as the current - round training set. Input it into the previous - round updated class Unet model for the current - round model training to obtain the current - round noise value and the current - round updated class Unet model. Calculate the current noise loss value based on the noise values of all clear ship pictures, the current - round noise value, and the noise values of all water - mist ship pictures. If the current noise loss value is within the preset noise loss value range, stop the model training, take the current - round updated class Unet model as the target class Unet model, and save the training rounds and the Gaussian noise added in each round. S5. After inputting the ship picture to be de - fogged into the target water - mist concentration classification model to judge the water - mist level, perform reverse inference de - fogging on it according to the training rounds and the Gaussian noise added in each round by inputting it into the target class Unet model to obtain the target clear ship picture.

[0005] By modifying the original water - mist concentration classification model through the above - mentioned technical solution, the obtained target water - mist concentration classification model classifies water - mist more accurately. Using the ship water - mist concentration level as a guidance in the original Unet model, defogging can be achieved according to different levels. Combining the airspace and frequency - domain for water - mist removal processing, the obtained target clear ship picture is more accurate.

[0006] However, the ship pictures finally obtained by the above method are ultimately synthesized from all the clear ship pictures obtained after multiple rounds of model training. Its main purpose is still to provide certain reference for the crew. However, at present, such dehazing methods and systems mainly cooperate with the crew to explore the relatively near waters around. Ultimately, it is still the crew who plays a decisive role. Therefore, the current situation of such ship dehazing systems is rather embarrassing. In the case of thin fog, the observers on the ship can directly observe the water conditions with the naked eye without the need for dehazing treatment. While when the fog is thick, the ship dehazing system can only play an appropriate role in assisting the crew and cannot provide early warnings for the crew during the process from the appearance of fog to its thickening.

[0007] Therefore, a ship warning image recognition system is proposed to solve or alleviate the above problems. Summary of the Invention

[0008] The purpose of the present invention is to solve the deficiencies existing in the prior art and propose a ship warning image recognition system.

[0009] To achieve the above purpose, the present invention adopts the following technical solutions:

[0010] A ship warning image recognition system includes an image acquisition module, an image dehazing module, a data processing module, and a warning module. The image acquisition module is used to collect the picture in front of the ship and feedback it to the image dehazing module. The image dehazing module is used to dehaze the picture in front of the ship and then provide a post-fog image and feedback it to the data processing module. The data processing module is used to match the ship visibility according to the post-fog image and control the warning module to give warnings and feedback according to the ship visibility.

[0011] Preferably, it further includes a storage module, which is coupled to the data processing module and is used to store the information of the data processing module.

[0012] Preferably, the image acquisition module includes a standard high-definition camera, which is used to collect the picture in front of the ship and feedback it to the image dehazing module.

[0013] Preferably, the image dehazing module is used to dehaze the picture in front of the ship and then provide a post-fog image and feedback it to the data processing module, including the following steps:

[0014] Obtain the picture in front of the ship from the image acquisition module;

[0015] Preprocess the picture in front of the ship to obtain a denoised image;

[0016] Calculate the minimum value of the three color channels in the neighborhood of each pixel point using the denoised image to obtain a dark channel image where, Ic (x, y) is the pixel value on the color channel C, J dark (x, y) is the dark channel image;

[0017] Estimate the atmospheric light by taking the maximum brightness value in the denoised image where I σ (x, y) is the denoised image;

[0018] According to the dark channel image J dark (x, y) and the atmospheric light A to estimate the transmittance where ω is a constant, usually taking values between 0.95 - 0.99;

[0019] Set the transmittance threshold t min , and distinguish and mark the fog - covered area in the ship's forward view based on the transmittance t(x, y). If t(x, y) < t min , then mask(x, y) = 1;

[0020] Filter the marked area in the ship's forward view and output the post - fog image.

[0021] Preferably, the data processing module is used to match the ship's visibility according to the post - fog image and control the warning module to give warnings and feedback according to the ship's visibility, including the following steps,

[0022] Extract the bounding boxes of multiple objects in the post - fog image through an object detection algorithm. Each bounding box gives the position of the object. Set the bounding box of each object as [x1, x1, y2, y2], where (x1, y1) is the upper - left coordinate and (x2, y2) is the lower - right coordinate;

[0023] Analyze all the detected objects to obtain the center coordinates of the objects

[0024] Set the reference position of the shooting point as (x ship , y ship ), and calculate the Euclidean distance between the object and the reference position of the shooting point

[0025] Calculate and sort the Euclidean distances of all objects to obtain the sequence Sort(D obj1 , D obj2 , D obj3 , …, D objn ) and select the three objects with the largest distances from it;

[0026] In the post - fog image, extract and normalize the pictures of the three objects with the largest distances;

[0027] Feature extraction is performed on the three normalized objects through a convolutional neural network, and the softmax layer is used to predict the object categories, obtaining multiple similar categories;

[0028] Compare the objects with the largest distances with their respective similar categories and output the similarity;

[0029] Set a similarity threshold, and select the object with the highest average similarity among the three objects with the largest distances as the object with the largest distance.

[0030] Preferably, the feature extraction of the three normalized objects through a convolutional neural network includes the following steps:

[0031] Use a convolutional neural network to extract the features of the images of the three objects with the largest distances through a convolution kernel;

[0032] Use the ReLU activation function for non-linear transformation ReLU(x) = max(0, x);

[0033] Use a pooling layer for downsampling where pooling window is the pooling window, and i and j are the offsets within the window;

[0034] Flatten and connect the outputs of the convolutional and pooling layers to a fully connected layer to obtain the feature vector V = W·F + b, where W is the weight matrix, b is the bias, and F is the feature of the images of the three objects with the largest distances.

[0035] Preferably, the use of the softmax layer to predict the object categories and obtain multiple similar categories includes the following steps:

[0036] Using the feature vector V output by the fully connected layer, use the softmax layer to predict the object categories. For multi-classification tasks, the category prediction formula is: where V c is the score corresponding to the feature vector V for category c, and p(y = c|V) is the probability of predicting to belong to category c.

[0037] Preferably, the comparison of the objects with the largest distances with their respective similar categories and the output of the similarity includes the following steps:

[0038] Use cosine similarity to calculate the similarity between the objects with the largest distances and multiple similar categories, where ||V1|| and ||V2|| are the L2 norms of the feature vectors V1 and V2 respectively, and · represents the dot product of the vectors.

[0039] Preferably, the data processing module is used to match the ship visibility according to the post-fog image and control the warning module to give warnings and feedback according to the ship visibility. The method further includes the following steps:

[0040] Determine the object with the largest distance, and take the central coordinates (x center , y center ) of the object with the largest distance, determine the focal length of the image acquisition module, and determine that there is a bow in the front view of the ship and the actual width of the bow;

[0041] Determine the pixel width of the bow through the front view of the ship and calculate the proportional relationship = actual width of the bow / pixel width of the bow;

[0042] Calculate the actual width W of the central coordinates of the object with the largest distance p = w p × proportional relationship, where w p is the pixel width of the central coordinates of the object with the largest distance;

[0043] Apply the collinear equation to calculate the distance between the central coordinates of the object with the largest distance and the image acquisition unit

[0044] Preferably, the data processing module is used to match the ship visibility according to the post-fog image and control the warning module to give warnings according to the ship visibility. The method further includes the following steps:

[0045] Determine the distance between the central coordinates of the object with the largest distance and the image acquisition unit as the visibility. If the visibility ≥ 200 meters, no warning is required. If the visibility < 200 meters, control the warning module to give a warning to the crew.

[0046] The present invention has the following beneficial effects:

[0047] The present invention can collect the front view of the ship through the image acquisition module during the ship's voyage. When there is fog in the front view, filter the part of the front view that is completely obscured by the fog, and only retain the content in the front view that has not been completely blocked by the fog. At the same time, identify and judge based on the remaining filtered content, and obtain the distance between the position with the highest clarity, recognition, and finally verified similarity and the present ship as the visibility. By this method, a relatively actual ship visibility can be obtained to give a warning to the crew, avoiding the crew's misjudgment of the ship visibility due to the image obtained by the conventional defogging system. At the same time, when the visibility is lower than a certain value, give a clear warning to the crew, which can enable the crew to turn on devices such as radar in advance, and arrange more experienced crew to drive the ship or issue more accurate instructions separately, avoiding situations such as collisions during the ship's voyage. Description of the Drawings

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0049] Figure 1 It is the structural block diagram of the present invention.

[0050] 1. Image acquisition module; 2. Image dehazing module; 3. Data processing module; 4. Early warning module; 5. Storage module. Specific embodiments

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings here can be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0053] It should be noted that: similar reference numerals and letters indicate similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0054] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product of the present invention is normally placed, or the orientation or positional relationship commonly understood by those skilled in the art. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0055] In addition, the terms "first", "second", "third", etc. are only used for descriptive distinction and cannot be understood as indicating or implying relative importance.

[0056] In the description of the present invention, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "installed", "connected", and "coupled" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0057] A ship warning image recognition system, as Figure 1 shown, includes an image acquisition module 1, an image defogging module 2, a data processing module 3, a warning module 4, and a storage module 5. The image acquisition module 1 is used to acquire the front view of the ship and feed it back to the image defogging module 2. The image defogging module 2 is used to defog the front view of the ship and provide a post-defogging image and feed it back to the data processing module 3. The data processing module 3 is used to match the ship visibility according to the post-defogging image and control the warning module 4 to give a warning and feedback according to the ship visibility. The storage module 5 is coupled to the data processing module 3, and the storage module 5 is used to store the information of the data processing module 3.

[0058] The image acquisition module 1 includes a standard high-definition camera, and the standard high-definition camera is used to acquire the front view of the ship and feed it back to the image defogging module 2. The storage module 5 includes a memory. The image defogging module 2 and the data processing module 3 include a processor and a computer program stored in the memory and running on the processor. The warning module 4 includes an alarm.

[0059] A system for judging visibility and giving a warning during ship navigation, through the image acquisition unit carried on the ship, can capture the front view ahead of the bow.

[0060] When affected by fog, the system will automatically identify and filter out the areas completely blocked by fog, and only retain those partially blocked or unblocked images.

[0061] Then, using image processing technology, analyze the remaining visible content, identify the target objects with the highest clarity and recognition, and calculate the actual distance between these target objects and the ship, which is used as a measure of visibility. In this way, it can effectively provide more real visibility data, help the crew avoid misjudgment due to the images of traditional defogging systems. When the system detects that the visibility drops to a preset threshold, it will give a warning in time, prompt the crew to activate auxiliary equipment such as radar, and recommend that a more experienced crew take over the driving or issue navigation instructions to prevent possible maritime collision accidents.

[0062] Preferably, the image dehazing module 2 is used to dehaze the image of the front of the ship and provide the dehazed image to the data processing module 3, including the following steps:

[0063] Obtain the image of the front of the ship from the image acquisition module 1;

[0064] Preprocess the image of the front of the ship to obtain the denoised image;

[0065] Use the denoised image to calculate the minimum value of the three color channels in the neighborhood of each pixel to obtain the dark channel image where I c (x, y) is the pixel value on the color channel C, and J dark (x, y) is the dark channel image;

[0066] Estimate the atmospheric light by taking the maximum brightness value in the denoised image where I σ (x, y) is the denoised image;

[0067] Estimate the transmittance according to the dark channel image J dark (x, y) and the atmospheric light A where ω is a constant, usually taking values between 0.95 and 0.99;

[0068] Set the transmittance threshold t min , and distinguish and mark the fog-covered area in the image of the front of the ship based on the transmittance t(x, y). If t(x, y) < t min , then mask(x, y) = 1;

[0069] Filter the marked area in the image of the front of the ship and output the dehazed image.

[0070] Through the above method steps, the system can process the image of the front of the ship, so that when the image of the front of the ship is blocked by fog, the completely fog-blocked part of the image of the front of the ship can be filtered, and only the part that can still be seen in the image of the front of the ship is retained. In subsequent steps, these visible parts can be used to calculate the visibility of the ship.

[0071] Preferably, the data processing module 3 is used to match the ship visibility according to the dehazed image and control the warning module 4 to give warnings and feedback according to the ship visibility, including the following steps:

[0072] Extract the bounding boxes of multiple objects in the dehazed image through the object detection algorithm. Each bounding box gives the position of the object. Set the bounding box of each object as [x1, x1, y2, y2], where (x1, y1) is the upper left coordinate and (x2, y2) is the lower right coordinate;

[0073] Analyze all detected objects to obtain the center coordinates of the objects

[0074] Set the reference position of the shooting point as (x ship , y ship ), and calculate the Euclidean distance between the object and the reference position of the shooting point

[0075] Calculate and sort the Euclidean distances of all objects to obtain the sequence Sort(D obj1 , D obj2 , D obj3 , …, D objn ). Select the three objects with the largest distances from it

[0076] In the post - fog image, extract and normalize the pictures of the three objects with the largest distances

[0077] Extract features of the three normalized objects through a convolutional neural network, and use the softmax layer to predict the object categories to obtain multiple similar categories

[0078] Compare each object with the largest distance with its respective similar category and output the similarity

[0079] Set a similarity threshold, and select the object with the highest average similarity among the three objects with the largest distances as the object with the largest distance

[0080] Through the above method steps, the system can perform target recognition on the visible part of the post - fog image obtained after defogging, and then identify several objects with the farthest Euclidean distances among these target objects. Then, through the convolutional neural network for matching among these farthest objects, it is required to select an object with the highest clarity and the farthest distance as a benchmark, so that the system can calculate the distance between this object and the ship as the visibility for feedback

[0081] Preferably, extracting features of the three normalized objects through a convolutional neural network includes the following steps

[0082] Use the convolutional neural network to extract features of the pictures of the three objects with the largest distances through convolutional kernels

[0083] Use the ReLU activation function for non - linear transformation ReLU(x) = max(0, x)

[0084] Use the pooling layer for downsampling where pooling window is the pooling window, and i and j are the offsets within the window

[0085] Flatten the output of the convolutional and pooling layers and connect it to the fully connected layer to obtain the feature vector V of the object, where V = W·F + b, W is the weight matrix, b is the bias, and F is the feature of the pictures of the three objects with the largest distance.

[0086] Preferably, use the softmax layer to predict the object category, obtaining multiple similar categories, including the following steps.

[0087] Using the feature vector V output by the fully connected layer, use the softmax layer to predict the object category. For the multi-classification task, the category prediction formula is: where, V c is the score corresponding to the category c of the feature vector V, and p(y = c∣V) is the probability of predicting to belong to the category c.

[0088] Preferably, compare each object with the largest distance with its respective similar category and output the similarity, including the following steps.

[0089] Use the cosine similarity to calculate the similarity between the object with the largest distance and multiple similar categories. where, ||V1|| and ||V2|| are the L2 norms of the feature vectors V1 and V2 respectively, and · represents the dot product of vectors.

[0090] Preferably, the data processing module 3 is used to match the ship visibility according to the post-fog image and control the warning module 4 to give warnings and feedback according to the ship visibility, and further includes the following steps.

[0091] Determine the object with the largest distance, and take the center coordinates (x center , y center ) of this object with the largest distance, determine the focal length of the image acquisition module 1, and determine that there is a ship's bow in the forward view of the ship and the actual width of the bow.

[0092] Determine the pixel width of the bow through the forward view of the ship and calculate the proportional relationship = actual width of the bow / pixel width of the bow.

[0093] Calculate the actual width W of the center coordinates of the object with the largest distance p = w p × proportional relationship, where w p is the pixel width of the center coordinates of the object with the largest distance.

[0094] Apply the collinear equation to calculate the distance between the center coordinates of the object with the largest distance and the image acquisition unit.

[0095] Preferably, the data processing module 3 is used to match the ship visibility according to the post-fog image and control the warning module 4 to give a warning according to the ship visibility. The following steps are further included:

[0096] Determine the distance between the central coordinates of the object with the largest distance and the image acquisition unit as the visibility. If the visibility ≥ 200 meters, no warning is required. If the visibility < 200 meters, control the warning module 4 to give a warning to the crew.

[0097] Through the above steps and methods, the system can calculate the approximate distance between the object and the ship based on the obtained target object, and then obtain the corresponding visibility. This visibility can be used as a reference for the crew and can also be compared with the subsequent set threshold to achieve the action of automatic warning.

[0098] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A ship warning image recognition system, characterized in that, It includes an image acquisition module (1), an image dehazing module (2), a data processing module (3), and an early warning module (4). The image acquisition module (1) is used to acquire the front view of the ship and feed it back to the image dehazing module (2). The image dehazing module (2) is used to dehaze the front view of the ship and provide a dehazed image, which is then fed back to the data processing module (3). The data processing module (3) is used to match the ship visibility based on the dehazed image and control the early warning module (4) to give an early warning and provide feedback according to the ship visibility. Extract the bounding boxes of multiple objects in the dehazed image through an object detection algorithm. Each bounding box gives the position of the object. Set the bounding box of each object as [x1, x1, y2, y2], where (x1, y1) is the upper left corner coordinate and (x2, y2) is the lower right corner coordinate. Analyze all detected objects to obtain the center coordinates of the objects Set the reference position of the shooting point as (x ship , y ship ), and calculate the Euclidean distance between the object and the reference position of the shooting point Calculate and sort the Euclidean distances for all objects to obtain the sequence Sort(D obj1 , D obj2 , D obj3 , …, D objn ). Select the three objects with the largest distances from it; In the dehazed image, extract and normalize the images of the three objects with the largest distances. Extract features of the three normalized objects through a convolutional neural network, and use the softmax layer to predict the object categories to obtain multiple similar categories. Compare each object with the largest distance with its respective similar category and output the similarity. Set a similarity threshold, and take the object with the highest average similarity among the three objects with the largest distances as the object with the largest distance.

2. The ship warning image recognition system according to claim 1, characterized in that, It further includes a storage module (5). The storage module (5) is coupled to the data processing module (3), and the storage module (5) is used to store the information of the data processing module (3).

3. The ship warning image recognition system according to claim 1, characterized in that, The image acquisition module (1) includes a standard high-definition camera, and the standard high-definition camera is used to acquire the front view of the ship and feed it back to the image dehazing module (2).

4. A ship warning image recognition system according to claim 1, characterized in that, The image dehazing module (2) is used to dehaze the front view of the ship and provide a dehazed image, which is then fed back to the data processing module (3). It includes the following steps: Obtain the front view of the ship from the image acquisition module (1). Preprocess the front view of the ship to obtain a denoised image. Calculate the dark channel image by obtaining the minimum value of the three color channels in the neighborhood of each pixel using the denoised image where I c (x, y) is the pixel value on the color channel C, and J dark (x, y) is the dark channel image; Estimate the atmospheric light by taking the maximum brightness value in the denoised image where I σ (x, y) is the denoised image; Estimate the transmittance according to the dark channel image J dark (x, y) and the atmospheric light A where ω is a constant with a value ranging from 0.95 to 0.99; Set the transmittance threshold t min , and distinguish and mark the fog-shaded area in the front view of the ship based on the transmittance t(x, y). If the transmittance is less than the transmittance threshold, set the coordinate mask to 1; Filter the marked area in the front view of the ship and output the dehazed image.

5. A ship warning image recognition system according to claim 1, characterized in that, The step of extracting features of the three normalized objects through a convolutional neural network includes the following steps: Use a convolutional neural network to extract the features of the images of the three objects with the largest distances through a convolution kernel. Use the ReLU activation function for non-linear transformation ReLU(x) = max(0, x). Downsampling using a pooling layer where poolingwindow is the pooling window, and i and j are the offsets within the window; Flatten the output of the convolution and pooling layers and connect it to the fully connected layer to obtain the feature vector of the object V = W·F + b, where W is the weight matrix, b is the bias, and F is the feature of the images of the three objects with the largest distances.

6. The ship warning image recognition system according to claim 1, characterized in that, The step of using the softmax layer to predict the object categories to obtain multiple similar categories includes the following steps: The feature vector V output by the fully connected layer is used by the softmax layer to predict the object category. For a multi-classification task, the category prediction formula is: where V c is the score of the feature vector V corresponding to category c, and p(y = c|V) is the probability of predicting to belong to category c.

7. The ship warning image recognition system according to claim 1, characterized in that, The step of comparing each object with the largest distance with its respective similar category and outputting the similarity includes the following steps: Calculate the similarity between the object with the maximum distance and multiple similar classes using cosine similarity, where ||V1|| and ||V2|| are the L2 norms of feature vectors V1 and V2 respectively, and · represents the dot product of vectors.

8. A ship warning image recognition system according to claim 1, characterized in that, The data processing module (3) is used to match the ship visibility based on the dehazed image and control the early warning module (4) to give an early warning and provide feedback according to the ship visibility. It further includes the following steps: Determine the object with the maximum distance, and take the x coordinate of the center of the object with the maximum distance center ,y center ), determine the focal length of the image acquisition module (1), determine whether there is a bow in the image in front of the ship and the actual width of the bow; Determine the pixel width of the bow through the front view of the ship and calculate the proportional relationship = actual width of the bow / pixel width of the bow. Calculate the actual width W of the center coordinates of the object with the maximum distance p = w p × proportional relationship, where w p is the pixel width of the center coordinates of the object with the maximum distance; Calculate the distance between the center coordinates of the object with the maximum distance and the image acquisition unit using the collinearity equation 9. The ship warning image recognition system according to claim 8, characterized in that, The data processing module (3) is used to match the ship visibility according to the image after fog and control the warning module (4) to give a warning, and further includes the following steps, Determine the distance between the central coordinates of the object with the largest distance and the image acquisition unit as the visibility. If the visibility ≥ 200 meters, no warning is required. If the visibility < 200 meters, control the warning module (4) to give a warning to the crew.

Images