CN114627415A - Ship detection method and system based on adaptive data enhancement - Google Patents
Ship detection method and system based on adaptive data enhancement Download PDFInfo
- Publication number
- CN114627415A CN114627415A CN202210249886.8A CN202210249886A CN114627415A CN 114627415 A CN114627415 A CN 114627415A CN 202210249886 A CN202210249886 A CN 202210249886A CN 114627415 A CN114627415 A CN 114627415A
- Authority
- CN
- China
- Prior art keywords
- picture
- ship
- feature
- loss
- enhancement
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 91
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 23
- 238000004364 calculation method Methods 0.000 claims abstract description 40
- 238000003062 neural network model Methods 0.000 claims abstract description 27
- 238000000034 method Methods 0.000 claims abstract description 18
- 238000000605 extraction Methods 0.000 claims abstract description 14
- 230000006870 function Effects 0.000 claims description 54
- 239000000463 material Substances 0.000 claims description 52
- 238000010586 diagram Methods 0.000 claims description 22
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 7
- 238000002372 labelling Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 4
- 238000007499 fusion processing Methods 0.000 claims description 3
- 238000007500 overflow downdraw method Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 125000004432 carbon atom Chemical group C* 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a ship detection method and a system based on adaptive data enhancement, wherein the method comprises the following steps: acquiring a picture data set through a camera and performing data enhancement processing; based on the neural network model, carrying out picture feature extraction processing on the enhanced picture; performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions; performing loss calculation on the characteristic picture with multi-scale dimensionality through a loss function to obtain a final loss value; updating the neural network model according to the final loss value, and constructing a ship detection model; and detecting the picture to be detected based on the ship detection model to obtain a detection result. The ship detection method is based on the visible light video detection technology, the detection sensing capability of the ship body is improved by constructing the ship detection model, and the ship detection precision can be improved under a special detection environment. The ship detection method and system based on adaptive data enhancement can be widely applied to the technical field of ship detection.
Description
Technical Field
The invention relates to the technical field of ship detection, in particular to a ship detection method and system based on adaptive data enhancement.
Background
The marine transportation is an important link of international logistics, with the continuous development of economy in China, the shipping flow is continuously increased, and various problems are met, such as behaviors of marine violation, illegal transportation and the like, and further, in order to ensure the continuity and timeliness of supervision of various large ports and navigation channels, a technology of intelligent ship detection must be introduced.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide a ship detection method and system based on adaptive data enhancement, which utilize a visible light video detection technology and introduce multi-scale dimension attention calculation and aspect ratio loss function to improve a ship detection algorithm, so as to improve ship detection accuracy in a special detection environment.
The first technical scheme adopted by the invention is as follows: the ship detection method based on adaptive data enhancement comprises the following steps:
acquiring a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture;
based on the neural network model, carrying out picture feature extraction processing on the enhanced picture to obtain a feature picture;
performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
performing loss calculation on the characteristic picture with multi-scale dimensionality through a loss function to obtain a final loss value;
updating the neural network model according to the final loss value, and constructing a ship detection model;
and detecting the picture to be detected based on the ship detection model to obtain a detection result.
Further, the step of obtaining a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture specifically includes:
acquiring a picture data set through a camera;
extracting pictures according to the picture data set to obtain a first picture and a second picture;
marking and cutting the ship image in the first picture to obtain an enhanced material;
and pasting the enhanced material to the second picture according to a preset rule to obtain an enhanced picture.
Further, the step of labeling and cutting the ship image in the first picture to obtain the enhanced material specifically includes:
labeling the ship image in the first picture according to a ship detection frame to obtain a ship image with a label, wherein the ship detection frame is provided with a detection label;
cutting the marked ship image to obtain a cut image;
judging whether the cut image is cut off or not according to the detection label of the ship detection frame to obtain the cut image which is cut off and the cut image which is not cut off;
and integrating the cut-off image and the non-cut-off image to obtain the enhanced material.
Further, the step of pasting the enhancement material to the second picture according to a preset rule to obtain the enhanced picture specifically includes:
the preset rule comprises selecting a pasting area in the second picture, carrying out scaling processing on the enhancement materials and judging the pasting position of the enhancement materials;
judging that the enhancement material is a cut image which is cut off, and selecting the edge which is pasted on the second picture to obtain the pasting position of the edge target;
judging that the enhancement material is an uncut cut image, and randomly selecting a position on the second picture to obtain a target pasted central point;
judging that the size of the enhanced material is larger than a preset size, and reducing the enhanced material;
and when the cross ratio value of the pasting position of the enhancement material is judged to be larger than or equal to the preset value, reselecting the pasting position of the enhancement material, and pasting the enhancement material to the second picture to obtain the enhanced picture until the cross ratio value of the pasting position of the enhancement material is judged to be smaller than the preset value.
Further, the step of performing image feature extraction processing on the enhanced image based on the neural network model to obtain a feature image specifically includes:
the neural network model comprises four stage layers, each stage layer comprises a plurality of residual blocks, and each residual block comprises a first convolution layer, a second convolution layer, a first batch of normalization layers, a second batch of normalization layers, a linear rectification layer and a down-sampling layer;
and performing feature extraction processing on the enhanced picture through convolution calculation to obtain a feature picture.
Further, the step of performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions specifically includes:
performing channel dimension lifting processing on the feature picture based on an extended dimension method to obtain a multi-dimensional feature picture;
calculating an attention value of the multi-dimensional feature picture based on a channel weighting method to obtain a multi-dimensional feature picture;
and performing fusion processing on the multi-scale feature picture based on a dimension fusion method to obtain the feature picture with multi-scale dimensions.
Further, the step of performing loss calculation on the feature picture with multi-scale dimensions through a loss function to obtain a final loss value specifically includes:
the loss functions include a thermodynamic loss function, a center point offset value loss function, a length-width prediction loss function, and an aspect ratio loss function;
respectively performing loss calculation on the characteristic picture with multi-scale dimensionality through a thermodynamic diagram loss function, a central point deviation value loss function, a length-width prediction loss function and an aspect ratio loss function to obtain a thermodynamic diagram loss value, a central point deviation loss value, an aspect ratio prediction loss value and an aspect ratio loss value;
and integrating the thermodynamic diagram loss value, the central point offset loss value, the aspect ratio prediction loss value and the aspect ratio loss value to obtain a final loss value.
Further, the length and width prediction loss function is expressed as follows:
in the above formula, LsizeThe length and width predicted loss value is represented,
indicates the predicted length and width, skRepresenting a genuine label.
Further, the aspect ratio loss function is represented as follows:
in the above formula, LRatio、
Represents the aspect ratio loss value, x, y represent coordinate points on the thermodynamic diagram, c represents the target class,
indicates the predicted length of the detected object,
indicates the predicted width of the detected object, HxyzRepresenting the true length of the detected object, WxycRepresenting the true width of the detection target.
The second technical scheme adopted by the invention is as follows: a ship detection system based on adaptive data enhancement comprises:
the enhancement module is used for acquiring a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture;
the characteristic extraction module is used for extracting the image characteristics of the enhanced image based on the neural network model to obtain a characteristic image;
the multi-scale dimension calculation module is used for carrying out multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
the loss calculation module is used for performing loss calculation on the characteristic picture with the multi-scale dimensionality through a loss function to obtain a final loss value;
the updating module is used for updating the neural network model according to the final loss value and constructing a ship detection model;
and the detection module is used for detecting the picture to be detected based on the ship detection model to obtain a detection result.
The method and the system have the beneficial effects that: the method applies a visible light video detection technology, can enhance the perception capability of the neural network model to the ship detection target through multi-scale dimension attention calculation, and can predict the length-width ratio information of the ship to be detected by using the loss function based on the length-width ratio consistency so as to further improve the prediction precision capability of the neural network model, thereby improving the ship detection precision under a special detection environment.
Drawings
FIG. 1 is a flow chart of the steps of the adaptive data enhancement based ship detection method of the present invention;
FIG. 2 is a block diagram of the structure of the ship detection system based on adaptive data enhancement;
FIG. 3 is a schematic illustration of the adaptive data enhancement of the present invention;
FIG. 4 is a schematic diagram of the deep neural network model structure of the present invention;
FIG. 5 is a schematic diagram of a backbone network architecture to which the present invention is applied;
FIG. 6 is a block diagram of a multi-scale dimensional attention calculation module of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments. The step numbers in the following embodiments are provided only for convenience of illustration, the order between the steps is not limited at all, and the execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
The method relies on a visible light video detection technology, improves a ship detection algorithm by introducing multi-scale dimension attention calculation and an aspect ratio loss function, and can improve the ship detection precision in a special detection environment.
Referring to fig. 1, the invention provides a ship detection method based on adaptive data enhancement, which comprises the following steps:
s1, acquiring a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture;
s11, acquiring a picture data set through a camera;
s12, extracting pictures according to the picture data set to obtain a first picture and a second picture;
specifically, referring to fig. 3, first two pictures, I for each picture, are randomly extracted from the picture data set1And I2And further from picture I1To obtain the enhancement material.
S13, labeling and cutting the ship image in the first picture to obtain an enhanced material;
s131, performing annotation processing on the ship image in the first picture according to a ship detection frame to obtain a ship image with an annotation, wherein the ship detection frame is provided with a detection label;
s132, cutting the marked ship image to obtain a cut image;
s133, judging whether the cut image is cut off or not according to the detection label of the ship detection frame to obtain the cut image which is cut off and the cut image which is not cut off;
and S134, integrating the cut-off image and the non-cut-off image to obtain the enhanced material.
In particular, Picture I1Wherein the marked ship detection frame G ═ G1,g2,…,gnWith 4 variables per box representing gi＝ (xi,min,yi,min,xi,max,yi,max) Wherein x isi,min,yi,minIndicates the upper left corner, x, of the corresponding detection boxi,max,yi,maxRepresenting the lower right hand corner of the corresponding detection box. Then randomly selecting one of all the objects in the picture, and directly cutting the selected object according to the marked coordinates to obtain an image I2Then, whether the target is cut off is judged, and the judgment rule is as follows: if (x)i,min,yi,min) Or (x)i,max,yi,max) And on the periphery, the target is considered to be truncated, otherwise, the target is not truncated, and the truncated cut image and the non-truncated cut image are integrated to obtain the enhanced material.
And S14, pasting the enhancement material to the second picture according to a preset rule to obtain an enhanced picture.
S141, the preset rule comprises selecting a pasting area in the second picture, zooming the enhancement material and judging the pasting position of the enhancement material;
s142, judging that the enhancement material is a cut image which is cut off, and selecting the edge which is pasted on the second picture to obtain the pasting position of the edge target;
s143, judging that the enhancement material is an uncut cut image, and randomly selecting a position on the second picture to obtain a target pasted central point;
specifically, the center point of the target paste target is randomly selected, and the center-lower part of the whole picture, i.e., between 1/3 and 3/4, is generally selected. For a truncated object, we paste the object to the corresponding edge, otherwise to an area that is not at random.
S144, judging that the size of the enhanced material is larger than the preset size, and reducing the enhanced material;
specifically, the size of the enhancement material is judged, the target with the smaller size of the enhancement material is not reduced before pasting, the target characteristics are seriously lost due to the too small target, the training effect is influenced, and when the size of the enhancement material is judged to be larger than the preset size, the enhancement material is reduced.
And S145, when the cross ratio value of the pasting position of the enhancement material is judged to be larger than or equal to the preset value, reselecting the pasting position of the enhancement material, and pasting the enhancement material to the second picture to obtain the enhanced picture until the cross ratio value of the pasting position of the enhancement material is judged to be smaller than the preset value.
Specifically, since the enhancement material itself may carry some background information of the picture, random pasting may affect the original target of the picture to be trained, before pasting, the intersection ratio of the target pasting position and all target positions of the picture to be detected is calculated respectively, if the sum of the values is less than a certain threshold, the next operation is performed, otherwise, the pasting position is reselected, when the selection frequency reaches the preset threshold, the maximum reselection frequency is exceeded, the data enhancement is given up to the picture to be detected, and when the intersection ratio of the pasting position of the enhancement material is less than the preset threshold, the data enhancement is given up to the picture to be detectedWhen setting the value, pasting the enhancement material to the second picture to obtain an enhanced picture which is marked as Iinput。
S2, based on the neural network model, carrying out picture feature extraction processing on the enhanced picture to obtain a feature picture;
s21, the neural network model comprises four stage layers, each stage layer comprises a plurality of residual blocks, and each residual block comprises a first convolution layer, a second convolution layer, a first batch normalization layer, a second batch normalization layer, a linear rectification layer and a down-sampling layer;
and S22, performing feature extraction processing on the enhanced picture through convolution calculation to obtain a feature picture.
Specifically, referring to fig. 4 and 5, the enhanced picture is subjected to feature extraction processing using ResNet-34 as a model backbone network, and a neural network model is represented by S0,S1,S2,S3Each stage consists of N residual blocks, and one residual block consists of two convolution layers, two batch normalization layers, one linear rectification layer and one down-sampling layer. Enhanced picture IinputEach characteristic obtained by performing characteristic extraction after calculating the convolution layer of each stage after the ResNet-34 network convolution is respectively fe0＝S0(Iinput),fe1＝S1(fe0),fe2＝S2(fe0),fe3＝S3(fe2) Each characteristic diagram is fe0、fe1、fe2And fe3。
S3, performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
s31, carrying out channel dimension lifting processing on the feature picture based on an extended dimension method to obtain a multi-dimension feature picture;
specifically, the feature map f is input into a multi-scale attention module after being calculated by a convolution layer and a linear rectification layer respectivelye0,fe1,fe2,fe3Extracting the channel dimension by a convolution layerAnd performing matrix fusion with two adjacent characteristic pictures to obtain a multi-dimensional characteristic picture of the next stage for input, wherein the multi-dimensional characteristic pictures are fd0,fd1,fd2,fd3The process can be expressed by the following formula:
in the above formula, fd0,fd1,fd2,fd3The multi-dimensional feature picture is represented,
representing a merge operation in the matrix channel dimension, P (e) representing a process through a convolutional layer and a linear rectifying layer, fd0、fd1、fd2And fd3Feature pictures respectively representing multi-scale dimensions, fe0、fe1、fe2And fe3Each representing a feature picture.
S32, calculating an attention value of the multi-dimensional feature picture based on a channel weighting method to obtain a multi-dimensional feature picture;
specifically, referring to fig. 6, let the feature picture f of multi-scale dimensiond0,fd1,fd2,fd3Respectively pass through four attention modules, each of which is respectively denoted as A0,A1,A2,A3The attention module firstly globally pools the multi-scale-dimension feature picture into a one-dimensional vector of (c,1) by adopting a channel weighting mode, wherein c represents the number of channels, and then the multi-scale feature picture is obtained by multiplying the original feature picture on the channels after full-connection layer transformation, and the specific process can be represented by the following formula:
m0＝A0(d0),m1＝A1(d1),m2＝A2(d2),m3＝A3(d3)
in the above formula, m0, m1, m2 and m3 respectively represent multi-scale feature pictures, ai(d) The characteristic diagram is shown to go through the ith attentionAnd (6) force module processing.
And S33, carrying out fusion processing on the multi-scale feature picture based on a dimension fusion method to obtain the feature picture with multi-scale dimensions.
Specifically, after the multi-scale feature maps respectively pass through the attention module, the multi-scale feature maps are fused on the channel dimension, and finally the feature picture f with the multi-scale dimension of the downstream task is obtainedinputThe specific process can be represented by the following formula:
in the above formula, the first and second carbon atoms are,
representing a fusion operation in the dimension of the matrix channel, finputRepresenting a feature picture with multi-scale dimensions.
S4, performing loss calculation on the characteristic picture with multi-scale dimensionality through a loss function to obtain a final loss value;
s41, the loss functions comprise a thermodynamic loss function, a center point offset value loss function, a length-width prediction loss function and an aspect ratio loss function;
s42, respectively performing loss calculation on the feature picture with multi-scale dimensions through a thermodynamic diagram loss function, a central point offset value loss function, a length-width prediction loss function and an aspect ratio loss function to obtain a thermodynamic diagram loss value, a central point offset loss value, an aspect ratio prediction loss value and an aspect ratio loss value;
specifically, firstly, a characteristic picture f of multi-scale dimensions is takeninputReducing the dimension of the first convolutional layer and the second convolutional layer through the first convolutional layer and the second convolutional layer, then respectively inputting the reduced dimension into each module for calculating a loss function, calculating a loss value, wherein the loss function comprises a thermodynamic diagram loss function, a central point deviation value loss function, a length-width prediction loss function and an aspect ratio loss function, and respectively calculating a characteristic picture f of a multi-scale dimensioninputCalculating;
the thermodynamic diagram loss function calculation formula is as follows:
in the above formula, LkRepresenting thermodynamic diagram loss values, N representing the image keypoint number index,
indicates the predicted value, YxycRepresenting a true value, alpha and beta representing a hyper-parameter, x and y representing coordinate points on a thermodynamic diagram, and c representing a target class;
the central point offset value loss function calculation formula is as follows:
in the above formula, LoffRepresenting the center point offset loss value, p represents the coordinates of the target center point,
and R is the scaling proportion of the scaled approximate integer coordinate.
Indicating a predicted deviation;
the calculation formula of the length and width prediction loss function is as follows:
in the above formula, LsizeThe length and width predicted loss value is represented,
indicates the predicted length and width, skA tag representing a real;
the aspect ratio loss function is calculated as follows:
in the above formula, LRatio、
Represents the aspect ratio loss value, x, y represent coordinate points on the thermodynamic diagram, c represents the target class,
indicates the predicted length of the detected object,
indicates the predicted width of the detected object, HxycRepresenting the true length of the detected object, WxycRepresenting the true width of the detection target;
in the course of the calculation process,
and λRatioThe same parameter is represented, namely a directly deleted denominator part, an aspect ratio fraction direct subtraction mode is not directly adopted, loss calculation is carried out by adopting a numerator part after fraction general division, and the purpose of doing so is that the model can predict a few small values in the initial training stage, so that the aspect ratio loss is too large, and the model training effect is influenced.
And S43, integrating the thermodynamic diagram loss value, the central point offset loss value, the aspect ratio prediction loss value and the aspect ratio loss value to obtain a final loss value.
Specifically, the weight of each loss value is set, the final loss value is the weighted sum of all the above loss values, and the weighting function calculation formula is as follows:
Ldet＝Lk+λsizeLsize+λoffLoff+λRationLRatio
in the above formula, LdetRepresents the finalLoss value, λsizeWeight, λ, representing the length-width predicted penalty valueoffWeight, λ, representing the center point offset penalty valueRatioRepresenting the weight of the aspect ratio loss value.
S5, updating the neural network model according to the final loss value, and constructing a ship detection model;
specifically, the final loss value is subjected to gradient reverse updating of the neural network model, loading parameters of the neural network model are updated to obtain an updated neural network model, and the ship detection model is constructed based on the updated neural network model.
And S6, detecting the picture to be detected based on the ship detection model to obtain a detection result.
Specifically, based on a ship detection model, pictures are sequentially selected from pictures to be tested and input into the ship detection model, the ship detection model is responsible for outputting predicted values of the pictures, including categories, central point positions and corresponding length and width values, and finally the overall accuracy of the pictures is calculated.
Referring to fig. 2, the ship detection system based on adaptive data enhancement comprises:
the enhancement module is used for acquiring a picture data set through a camera and carrying out data enhancement processing to obtain an enhanced picture;
the characteristic extraction module is used for extracting the image characteristics of the enhanced image based on the neural network model to obtain a characteristic image;
the multi-scale dimension calculation module is used for carrying out multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
the loss calculation module is used for performing loss calculation on the characteristic picture with the multi-scale dimensionality through a loss function to obtain a final loss value;
the updating module is used for updating the neural network model according to the final loss value and constructing a ship detection model;
and the detection module is used for detecting the picture to be detected based on the ship detection model to obtain a detection result.
The contents in the above method embodiments are all applicable to the present system embodiment, the functions specifically implemented by the present system embodiment are the same as those in the above method embodiment, and the beneficial effects achieved by the present system embodiment are also the same as those achieved by the above method embodiment.
While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (10)
1. The ship detection method based on adaptive data enhancement is characterized by comprising the following steps of:
acquiring a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture;
based on the neural network model, carrying out picture feature extraction processing on the enhanced picture to obtain a feature picture;
performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
performing loss calculation on the characteristic picture with multi-scale dimensionality through a loss function to obtain a final loss value;
updating the neural network model according to the final loss value, and constructing a ship detection model;
and detecting the picture to be detected based on the ship detection model to obtain a detection result.
2. The ship detection method based on adaptive data enhancement according to claim 1, wherein the step of obtaining a picture data set through a camera and performing data enhancement processing to obtain an enhanced picture specifically comprises:
acquiring a picture data set through a camera;
extracting pictures according to the picture data set to obtain a first picture and a second picture;
marking and cutting the ship image in the first picture to obtain an enhanced material;
and pasting the enhanced material to the second picture according to a preset rule to obtain an enhanced picture.
3. The adaptive data enhancement-based ship detection method according to claim 2, wherein the step of labeling and cropping the ship image in the first picture to obtain enhanced material specifically comprises:
labeling the ship image in the first picture according to a ship detection frame to obtain a ship image with a label, wherein the ship detection frame is provided with a detection label;
cutting the marked ship image to obtain a cut image;
judging whether the cut image is cut off or not according to the detection label of the ship detection frame to obtain the cut image which is cut off and the cut image which is not cut off;
and integrating the cut image and the non-cut image to obtain the enhanced material.
4. The ship detection method based on adaptive data enhancement according to claim 3, wherein the step of pasting the enhancement material to the second picture according to a preset rule to obtain the enhanced picture specifically comprises:
the preset rule comprises selecting a pasting area in the second picture, zooming the enhancement material and judging the pasting position of the enhancement material;
judging that the enhancement material is a cut image which is cut off, and selecting the edge which is pasted on the second picture to obtain the pasting position of the edge target;
judging that the enhancement material is an uncut cut image, and randomly selecting a position on the second picture to obtain a target pasted central point;
judging that the size of the enhancement material is larger than a preset size, and performing reduction processing on the enhancement material;
and when the cross ratio value of the pasting position of the enhancement material is judged to be larger than or equal to the preset value, reselecting the pasting position of the enhancement material, and pasting the enhancement material to the second picture to obtain the enhanced picture until the cross ratio value of the pasting position of the enhancement material is judged to be smaller than the preset value.
5. The ship detection method based on adaptive data enhancement according to claim 4, wherein the step of performing picture feature extraction processing on the enhanced picture based on the neural network model to obtain a feature picture specifically comprises:
the neural network model comprises four stage layers, each stage layer comprises a plurality of residual blocks, and each residual block comprises a first convolution layer, a second convolution layer, a first batch of normalization layers, a second batch of normalization layers, a linear rectification layer and a down-sampling layer;
and performing feature extraction processing on the enhanced picture through convolution calculation to obtain a feature picture.
6. The ship detection method based on adaptive data enhancement according to claim 5, wherein the step of performing multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions specifically comprises:
performing channel dimension lifting processing on the feature picture based on an extended dimension method to obtain a multi-dimensional feature picture;
calculating attention values of the multi-dimensional feature pictures based on a channel weighting method to obtain multi-dimensional feature pictures;
and performing fusion processing on the multi-scale feature picture based on a dimension fusion method to obtain the feature picture with multi-scale dimensions.
7. The ship detection method based on adaptive data enhancement according to claim 6, wherein the step of performing loss calculation on the feature picture with multi-scale dimensions through a loss function to obtain a final loss value specifically comprises:
the loss functions include a thermodynamic loss function, a center point offset value loss function, a length-width prediction loss function, and an aspect ratio loss function;
respectively performing loss calculation on the feature picture with multi-scale dimensionality through a thermodynamic diagram loss function, a central point offset value loss function, a length-width prediction loss function and an aspect ratio loss function to obtain a thermodynamic diagram loss value, a central point offset loss value, an aspect ratio prediction loss value and an aspect ratio loss value;
and integrating the thermodynamic diagram loss value, the central point offset loss value, the aspect ratio prediction loss value and the aspect ratio loss value to obtain a final loss value.
8. The adaptive data enhancement based ship detection method according to claim 7, wherein the length-width prediction loss function is expressed as follows:
in the above formula, LsizeThe length and width predicted loss value is represented,
indicates the predicted length and width, skRepresenting a genuine label.
9. The adaptive data enhancement based ship detection method according to claim 8, wherein the aspect ratio loss function is expressed as follows:
in the above formula, LRatio、
Represents the aspect ratio loss value, x, y represent coordinate points on the thermodynamic diagram, c represents the target class,
indicates the predicted length of the detected object,
indicates the predicted width of the detected object, HxycRepresenting the true length of the detected object, WxycRepresenting the true width of the detected object.
10. The ship detection system based on adaptive data enhancement is characterized by comprising the following modules:
the enhancement module is used for acquiring a picture data set through a camera and carrying out data enhancement processing to obtain an enhanced picture;
the characteristic extraction module is used for extracting the image characteristics of the enhanced image based on the neural network model to obtain a characteristic image;
the multi-scale dimension calculation module is used for carrying out multi-scale dimension attention calculation on the feature picture to obtain the feature picture with multi-scale dimensions;
the loss calculation module is used for performing loss calculation on the characteristic picture with the multi-scale dimensionality through a loss function to obtain a final loss value;
the updating module is used for updating the neural network model according to the final loss value and constructing a ship detection model;
and the detection module is used for detecting the picture to be detected based on the ship detection model to obtain a detection result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210249886.8A CN114627415A (en) | 2022-03-14 | 2022-03-14 | Ship detection method and system based on adaptive data enhancement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210249886.8A CN114627415A (en) | 2022-03-14 | 2022-03-14 | Ship detection method and system based on adaptive data enhancement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114627415A true CN114627415A (en) | 2022-06-14 |
Family
ID=81901972
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210249886.8A Pending CN114627415A (en) | 2022-03-14 | 2022-03-14 | Ship detection method and system based on adaptive data enhancement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114627415A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115331113A (en) * | 2022-10-12 | 2022-11-11 | 浙江华是科技股份有限公司 | Ship target detection model training method and system and computer storage medium |
| CN116052094A (en) * | 2023-03-07 | 2023-05-02 | 浙江华是科技股份有限公司 | Ship detection method, system and computer storage medium |
-
2022
- 2022-03-14 CN CN202210249886.8A patent/CN114627415A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115331113A (en) * | 2022-10-12 | 2022-11-11 | 浙江华是科技股份有限公司 | Ship target detection model training method and system and computer storage medium |
| CN116052094A (en) * | 2023-03-07 | 2023-05-02 | 浙江华是科技股份有限公司 | Ship detection method, system and computer storage medium |
| CN116052094B (en) * | 2023-03-07 | 2023-06-09 | 浙江华是科技股份有限公司 | Ship detection method, system and computer storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109299274B (en) | Natural scene text detection method based on full convolution neural network | |
| CN111080645B (en) | Remote sensing image semi-supervised semantic segmentation method based on generation type countermeasure network | |
| CN108009543B (en) | License plate recognition method and device | |
| CN111368600B (en) | Remote sensing image target detection and identification method and device, readable storage medium and equipment | |
| CN114627415A (en) | Ship detection method and system based on adaptive data enhancement | |
| CN111126359A (en) | High-definition image small target detection method based on self-encoder and YOLO algorithm | |
| CN111797829A (en) | License plate detection method and device, electronic equipment and storage medium | |
| CN109977997A (en) | Image object detection and dividing method based on convolutional neural networks fast robust | |
| CN114627052A (en) | Infrared image air leakage and liquid leakage detection method and system based on deep learning | |
| CN113869138A (en) | Multi-scale target detection method and device and computer readable storage medium | |
| CN113657409A (en) | Vehicle loss detection method, device, electronic device and storage medium | |
| CN113903022B (en) | Text detection method and system based on feature pyramid and attention fusion | |
| CN110852327A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| CN111553351A (en) | Semantic segmentation based text detection method for arbitrary scene shape | |
| CN113159215A (en) | Small target detection and identification method based on fast Rcnn | |
| CN114897738A (en) | Image blind restoration method based on semantic inconsistency detection | |
| CN114723945A (en) | Vehicle damage detection method and device, electronic equipment and storage medium | |
| CN112364709A (en) | Cabinet intelligent asset checking method based on code identification | |
| CN115272826A (en) | Image identification method, device and system based on convolutional neural network | |
| CN114743201A (en) | Multimeter reading identification method and system based on rotating target detection | |
| CN114219988A (en) | ViT-framework-based rapid multi-category rock mine classification method, device and storage medium | |
| CN113269717A (en) | Building detection method and device based on remote sensing image | |
| CN110874170A (en) | Image area correction method, image segmentation method and device | |
| CN117612136A (en) | Automatic driving target detection method based on increment small sample learning | |
| CN110188682B (en) | Optical remote sensing image target detection method based on geometric structure double-path convolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |