CN115082792A - Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning - Google Patents
Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning Download PDFInfo
- Publication number
- CN115082792A CN115082792A CN202210747277.5A CN202210747277A CN115082792A CN 115082792 A CN115082792 A CN 115082792A CN 202210747277 A CN202210747277 A CN 202210747277A CN 115082792 A CN115082792 A CN 115082792A
- Authority
- CN
- China
- Prior art keywords
- domain
- water surface
- surface target
- data
- region
- 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
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 135
- 238000001514 detection method Methods 0.000 title claims abstract description 118
- 238000013508 migration Methods 0.000 title claims abstract description 32
- 230000005012 migration Effects 0.000 title claims abstract description 23
- 230000003042 antagnostic effect Effects 0.000 title claims abstract description 8
- 238000012549 training Methods 0.000 claims abstract description 28
- 238000002372 labelling Methods 0.000 claims abstract description 15
- 238000013526 transfer learning Methods 0.000 claims abstract description 9
- 238000012546 transfer Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 74
- 238000000034 method Methods 0.000 claims description 24
- 150000001875 compounds Chemical class 0.000 claims description 9
- 238000011478 gradient descent method Methods 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000002689 soil Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning, which comprises the following steps of: 1) constructing a water surface target data set and a water surface target detection model by using the source domain data and the target domain data; 2) performing feature countermeasure transfer learning, training a water surface target detection model by using a water surface target data set, and realizing feature alignment of source domain data and target domain data in a countermeasure mode; 3) and performing semi-supervised learning, performing pseudo-labeling on partial target domain data by using the water surface target detection model, and performing semi-supervised training on the water surface target detection model after feature countermeasure transfer learning by using the water surface target data set. The invention trains the model by using the labeled sample data in other environments and transfers the knowledge learned by the model to a new environment, improves the detection precision of the cross-domain water surface target by combining a semi-supervised learning mode, reduces the data labeling cost of the model deployed in the new environment, and finally realizes the cross-domain water surface target detection.
Description
Technical Field
The invention relates to the technical field of water surface target detection, in particular to a cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning.
Background
Currently, the explosion of water economy brings about an increasingly lusterless water traffic. The water surface target detection method based on the optical image is beneficial to finding suspicious targets in time, and the management efficiency of the channel and the wharf is improved. However, as the appearance of the water surface target changes or the water surface environment changes, the detection performance of the detection model may be degraded. The calibration of a large amount of new data again can generate huge cost and is not favorable for continuous operation of a detection model, so that cross-domain water surface target detection is performed through feature countermeasure migration and semi-supervised learning, a continuous and reliable water surface monitoring system is facilitated to be constructed, and the method has important application value.
A large number of water surface target detection research works can be roughly divided into two categories: background modeling and deep learning based methods. The water surface target detection method based on background modeling models the water background, and obtains the foreground target through background subtraction. However, such methods are prone to generate a large number of false positives in cases where the background changes rapidly. Meanwhile, the detection precision is also reduced due to the change of light and dense water surface targets. Currently, with the continuous development of deep learning, a deep neural network model can learn richer features from data, and the performance of target detection is greatly improved. Therefore, the water surface target detection method based on deep learning is becoming the mainstream. However, such methods have many model parameters, and require a large amount of labeled data to prevent the risk of overfitting the model, and manual labeling of a large amount of data on water is time-consuming and costly. In addition, diversified environment and water surface targets are difficult to cover by using limited labeling data, so that extensive and robust training cannot be performed on a water surface target detection model.
Aiming at the problems, the invention provides a cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning. The core idea of the invention is as follows: the method comprises the steps of utilizing marked samples as source domain data, utilizing unmarked samples of an environment to be deployed as target domain data, migrating knowledge learned by a detection model from a source domain into a target domain in a characteristic countermeasure migration mode, marking the target domain in a small amount, improving the accuracy of cross-domain water surface target detection by combining a semi-supervised learning mode, and finally realizing cross-domain water surface target detection.
Disclosure of Invention
The invention aims to overcome the defects and shortcomings of the prior art, and provides a cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning.
In order to achieve the purpose, the technical scheme provided by the invention is as follows: the cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning comprises the following steps:
1) constructing a water surface target data set and a water surface target detection model by using the source domain data and the target domain data; the system comprises a water surface target detection model, a source domain data acquisition module, a target domain data acquisition module, a data acquisition module and a data acquisition module, wherein the source domain data is acquired from a labeled sample of a certain water surface environment, the target domain data is acquired from a non-labeled sample of a new environment, and the water surface target detection model comprises a basic detection network based on Resnet50, a domain confrontation discriminator based on a region and a domain confrontation category discriminator;
2) performing feature countermeasure transfer learning, training a water surface target detection model by using a water surface target data set, and realizing feature alignment of source domain data and target domain data in a countermeasure mode;
3) and performing semi-supervised learning, performing pseudo-labeling on partial target domain data by using the water surface target detection model after the characteristic countermeasure migration learning, performing semi-supervised training on the water surface target detection model after the characteristic countermeasure migration learning by using the water surface target data set, and performing the characteristic countermeasure migration learning, so that the performance of cross-domain water surface target detection is improved, and finally, accurate cross-domain water surface target detection is realized.
Further, in the step 1), collecting marked samples of the water surface environment as source domain data, collecting unmarked samples of a new environment as target domain data, and constructing a water surface target data set; and performing data enhancement on the samples in the water surface target data set, wherein the data enhancement comprises horizontal turning, random noise, rain generation and fog generation.
Further, in step 1), constructing a water surface target detection model, including:
a. constructing a structure of an Resnet 50-based basic detection network, wherein the structure comprises a Resnet 50-based feature extractor, a region proposal network, a region network, region pooling and an objective function;
after the structure of the basic detection network based on Resnet50 is built, pre-training is carried out, and the process comprises the following steps: firstly, inputting source domain data into a feature extractor to obtain a feature map, then generating a candidate region through a region proposing network, inputting the region pooled candidate region features into the region network to further optimize a detection result, and finally adjusting parameters of a water surface target detection model through optimizing a target function of a basic network; the objective function of the basic detection network based on Resnet50 is as follows:
in the formula, D S The source domain data is represented by a representation of,
a loss function representing the area proposed network,
representing the loss function, λ, of the area network 1 To represent
The penalty factor of (2) is determined,
and
both contain cross-entropy loss for classification and the Smooth-L1 function for position regression;
b. building a domain confrontation discriminator based on regions: the method comprises the steps that a gradient inversion layer and a convolution layer are connected in series to construct a region-based domain confrontation discriminator, wherein the output of a region proposal network is used as the input of the region-based domain confrontation discriminator, a region-based domain confrontation objective function is adopted to train the parameters of the region-based domain confrontation discriminator, and the parameters of a water surface target detection model are adjusted in a gradient back propagation mode; wherein the region-based domain confrontation objective function is:
in the formula (I), the compound is shown in the specification,
representing a region-based domain confrontation objective function,
the target domain data is represented by a representation of,
representing a set of candidate areas, p, generated by a network of area proposals c (r) represents the probability that the candidate region r belongs to the domain class c;
c. building a domain confrontation category discriminator: constructing a domain confrontation category discriminator by serially connecting a gradient inversion layer and a convolution layer, wherein the candidate region characteristics after region pooling are used as the input of the domain confrontation category discriminator, the domain confrontation category discriminator is weighted by utilizing the output of a region network, the parameters of the domain confrontation category discriminator are trained by adopting a domain confrontation category target function, and the parameters of a water surface target detection model are adjusted in a gradient back propagation mode; wherein, the objective function of the domain confrontation category is as follows:
in the formula (I), the compound is shown in the specification,
a domain confrontation category objective function is represented,
a multi-class domain counter-loss function is represented,
representing an entropy function;
the multi-class domain counter-loss function is in the mathematical form:
in the formula (I), the compound is shown in the specification,
representing the total number of target categories of source domain data and target domain data,
a class probability distribution indicating that the candidate region r belongs to class i, F a feature vector of the candidate region r, c a domain class,
representation calculation
And c;
the mathematical form of the entropy function is:
further, in step 2), the purpose of feature countermeasure transfer learning is to align the features of the source domain data and the target domain data in the water surface target data set by the water surface target detection model in a countermeasure manner, and therefore, a domain countermeasure discriminator based on a region is combined to train the feature countermeasure transfer learning, and a random gradient descent method is adopted to adjust the parameters of the water surface target detection model, and the objective function is as follows:
in the formula, D S The source domain data is represented by a representation of,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function, λ 2 To represent
The penalty factor of (2).
Further, the step 3) comprises the following steps:
3.1) performing pseudo-labeling on part of unmarked target domain data by using the water surface target detection model after the transfer-resistant learning by using the characteristics, and adjusting the accuracy of the pseudo-labeling in a manual mode;
3.2) carrying out semi-supervised training by using a domain confrontation discriminator and a domain confrontation category discriminator based on a region, and adjusting parameters of a water surface target detection model by adopting a random gradient descent method, wherein an objective function is as follows:
in the formula, D S The source domain data is represented by a representation of,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function,
representing a domain confrontation class objective function, λ 3 And λ 4 Respectively represent
And
the penalty factor of (2);
and finally, deploying a water surface target detection model with convergent training in a new environment to realize accurate cross-domain water surface target detection.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1. the invention adopts the technical scheme of combining the characteristic countermeasure migration and the semi-supervised learning, and utilizes the characteristic countermeasure migration to carry out characteristic alignment and pseudo labeling on the source domain data and the target domain data, so that the cross-domain target detection is simplified into a task similar to the semi-supervised detection.
2. The method and the system transfer the knowledge learned by the source domain data to the unmarked target domain through the characteristic countermeasure transfer, realize the cross-domain water surface target detection through a semi-supervised learning mode, effectively reduce the marking cost of the data and improve the intelligent level of the water management.
3. The invention provides a domain confrontation discriminator based on a region, which reduces the distribution difference of cross-domain features by aligning the features through maximizing the confrontation domain loss, thereby reducing the influence of the distribution difference on semi-supervised learning and improving the recall rate of cross-domain water surface targets.
4. The invention provides a domain confrontation category discriminator which performs weighted migration by using the similarity of target features, reduces the uncertainty of target categories through entropy regularization, and improves the feature discrimination of different categories of water surface targets, thereby improving the detection precision of a water surface target detection model.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a schematic diagram of source domain data and target domain data.
Fig. 3 is a block diagram of a water surface target detection model.
FIG. 4 is a diagram of a water surface target detection result achieved by the method of the present invention; wherein, "passanger ship" represents a passenger ship, "bulk cargo carrier" represents a bulk cargo ship, "speed boat" represents an express ship, "soil boat" represents a sailing ship, "boat" represents a boat, and "fishing boat" represents a fishing ship.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.
The embodiment discloses a cross-domain water surface target detection method based on feature anti-migration and semi-supervised learning, wherein an experiment platform is Python3.6, Pytrich1.7.1, and a computer is configured: CPU model is Intel (R) core (TM) i9-10900X, memory 64GB, and GPU model is NVIDIA Quadro M6000. As shown in fig. 1, the flow of the cross-domain water surface target detection method can be roughly divided into four stages: the first stage is the construction of a water surface target data set, and mainly comprises the steps of collecting source domain data and target domain data and performing data enhancement; the second stage is the construction of a water surface target detection model, which comprises a basic detection network based on Resnet50, a domain confrontation discriminator based on a region and a domain confrontation category discriminator, and is pre-trained by utilizing source domain data; the third stage is feature countermeasure transfer learning, and mainly uses a countermeasure mode to enable the water surface target detection model to realize feature alignment of source domain data and target domain data; the fourth stage is semi-supervised learning, which mainly comprises the steps of carrying out pseudo-labeling and manual fine adjustment on a small amount of target domain data, combining with feature pair migration to improve the performance of cross-domain water surface target detection, and finally realizing cross-domain water surface target detection. The embodiment specifically comprises the following steps:
1) constructing a water surface target data set and a water surface target detection model by using the source domain data and the target domain data; the structure of the water surface target detection model is shown in fig. 3, and comprises three modules, namely a basic detection network based on Resnet50, a domain confrontation discriminator based on a region and a domain confrontation category discriminator, specifically, the following:
a. and (3) construction of a water surface target data set: using public labeled data sets Singapore landmark dataset and SeaShips as source domain data, 14 water surface target categories are included: the method comprises the steps of collecting unlabeled samples of an A water environment as target domain data, and constructing a water surface target data set by using source domain data and target domain data, wherein the sizes of all samples in the water surface target data set are adjusted to be 1066 and 600 in length and width so as to reduce model calculation amount. Referring to fig. 2, in the drawings, (a) and (b) are schematic diagrams of source domain data and target domain data in a water surface target data set, it can be seen that a certain difference exists between the target appearance and the background style of the source domain data and the target domain data, and the difference easily causes that a detection model obtained by training the source domain data is difficult to generalize into a target domain;
b. data enhancement: according to the characteristics of the water surface environment, an album library is adopted to perform data enhancement on the training samples in the water surface target data set, wherein the data enhancement comprises horizontal turning, random noise, rain generation and fog generation.
c. Constructing a structure of a basic detection network based on Resnet50, which mainly comprises a feature extractor based on Resnet50, an area proposal network, an area network, area pooling and an objective function, wherein the function of the area proposal network is to extract candidate areas according to the shapes of a priori anchor boxes by using input data, and the widths and the heights of the shapes of the priori anchor boxes are set to be 15 shapes (46, 24), (32, 32), (24, 46), (184, 96), (128 ), (96, 184), (368, 192), (256 ), (192, 368), (736, 384), (512 ), (384, 736); the function of the area network is to further calculate the shape and the class of the detection target by using the candidate area output by the area proposal network, and optimize the detection result.
After the structure of the basic detection network based on Resnet50 is built, pre-training is carried out, and the process comprises the following steps: firstly, inputting source domain data into a feature extractor to obtain a feature map, then generating a candidate region through a region proposing network, inputting the region pooled candidate region features into the region network to further optimize a detection result, and finally adjusting parameters of a water surface target detection model through optimizing a target function of a basic network; the objective function of the basic detection network based on Resnet50 is as follows:
in the formula, D S The source domain data is represented by a representation of,
a loss function representing the area proposed network,
representing the loss function, λ, of the area network 1 To represent
The penalty factor of (2) is determined,
and
both contain cross-entropy loss for classification and the Smooth-L1 function for position regression.
d. Building a domain confrontation discriminator based on regions: the method comprises the following steps of constructing a region-based domain confrontation discriminator by serially connecting a gradient inversion layer and a convolution layer, wherein the calculation rule of the gradient inversion layer is to realize identity transformation in forward operation, and the gradient direction is automatically negated in reverse operation; wherein the region-based domain confrontation objective function is:
in the formula (I), the compound is shown in the specification,
representing a region-based domain confrontation objective function,
which represents the data of the target domain, and,
representing a set of candidate areas, p, generated by a network of area proposals c (r) represents the probability that the candidate region r belongs to the domain class c;
e. building a domain confrontation category discriminator: establishing a domain confrontation type discriminator by serially connecting a gradient inversion layer and a convolution layer, wherein the domain confrontation type discriminator has the functions of taking the candidate region characteristics after region pooling as the input of the domain confrontation type discriminator, weighting the domain confrontation type discriminator by utilizing the output of a region network, training the parameters of the domain confrontation type discriminator by adopting a domain confrontation type target function, and adjusting the parameters of a water surface target detection model in a gradient back propagation mode; wherein, the domain confrontation category objective function is:
in the formula (I), the compound is shown in the specification,
a domain confrontation category objective function is represented,
a multi-class domain counter-loss function is represented,
representing an entropy function;
the multi-class domain counter-loss function is in the mathematical form:
in the formula (I), the compound is shown in the specification,
representing the total number of target categories of source domain data and target domain data,
a class probability distribution indicating that the candidate region r belongs to class i, F a feature vector of the candidate region r, c a domain class,
representation calculation
And c;
the mathematical form of the entropy function is:
2) performing feature countermeasure migration learning, training a water surface target detection model by using a water surface target data set, and realizing feature alignment of source domain data and target domain data in a countermeasure mode, wherein the specific conditions are as follows:
the purpose of feature countermeasure transfer learning is to align the features of source domain data and target domain data in a water surface target data set by a water surface target detection model in a countermeasure mode, therefore, a domain countermeasure discriminator based on a region is combined to train the feature countermeasure transfer learning, parameters of the water surface target detection model are adjusted by adopting a random gradient descent method, and the target function is as follows:
in the formula, D S The source domain data is represented by a representation of,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function, λ 2 To represent
The penalty factor of (2);
loading pre-training parameters of a basic detection network based on Resnet50 before training, inputting a source domain and a target domain image batch to be 1 respectively by using a water surface target detection model in the training process, and optimizing an objective function by using a random gradient descent method, wherein the learning rate of the objective function is 1 multiplied by 10 -3 Momentum of 0.9, weight attenuation ratio of 1 × 10 -4 The training steps are 20000, lambda 1 Is set to 1, λ 2 Set to 0.1.
3) Performing semi-supervised learning, performing pseudo-labeling on partial target domain data by using a water surface target detection model after feature countermeasure migration learning, performing semi-supervised training on the water surface target detection model after feature countermeasure migration learning by using a water surface target data set, performing anti-migration learning by combining features, improving the performance of cross-domain water surface target detection, and finally realizing cross-domain water surface target detection by using a water surface target detection model with convergent training, wherein the method comprises the following specific steps:
3.1) performing pseudo-labeling on part of unmarked target domain data by using the water surface target detection model after the transfer-resistant learning by using the characteristics, and adjusting the accuracy of the pseudo-labeling in a manual mode;
3.2) carrying out semi-supervised training by using a domain confrontation discriminator and a domain confrontation category discriminator based on a region, and adjusting parameters of a water surface target detection model by adopting a random gradient descent method, wherein an objective function is as follows:
in the formula, D S The source domain data is represented by a representation of,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function,
representing a domain confrontation class objective function, λ 3 And λ 4 Respectively represent
And
the penalty factor of (2);
loading parameters of a water surface target detection model after characteristic anti-migration learning before training, wherein in the training process, the input source domain and target domain images of the water surface target detection model are respectively 1 in batch, optimizing an objective function by using a random gradient descent method, and the learning rate is 1 multiplied by 10 -3 Momentum is 0.9, and weight decay rate is 1 × 10 -4 The training step number is 20000, lambda 1 Is set to 1, λ 3 And λ 4 Are all set to 0.1. The detection result is shown in fig. 4, and includes a conventional scene, a scale change scene and an illumination change scene, and as can be seen from the figure, the method has good detection accuracy in various scenes, realizes accurate cross-domain water surface target detection, and is worthy of popularization.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.
Claims (5)
1. The cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning is characterized by comprising the following steps of:
1) constructing a water surface target data set and a water surface target detection model by using the source domain data and the target domain data; the system comprises a water surface target detection model, a source domain data acquisition module, a target domain data acquisition module, a data acquisition module and a data acquisition module, wherein the source domain data is acquired from a labeled sample of a certain water surface environment, the target domain data is acquired from a non-labeled sample of a new environment, and the water surface target detection model comprises a basic detection network based on Resnet50, a domain confrontation discriminator based on a region and a domain confrontation category discriminator;
2) performing feature countermeasure transfer learning, training a water surface target detection model by using a water surface target data set, and realizing feature alignment of source domain data and target domain data in a countermeasure mode;
3) and performing semi-supervised learning, performing pseudo-labeling on partial target domain data by using the water surface target detection model after the characteristic countermeasure migration learning, performing semi-supervised training on the water surface target detection model after the characteristic countermeasure migration learning by using the water surface target data set, and performing the characteristic countermeasure migration learning, so that the performance of cross-domain water surface target detection is improved, and finally, accurate cross-domain water surface target detection is realized.
2. The method for detecting the cross-domain water surface target based on the feature-based anti-migration and semi-supervised learning of claim 1, wherein in the step 1), a labeled sample of a water surface environment is collected as source domain data, and a label-free sample of a new environment is collected as target domain data, so as to construct a water surface target data set; and performing data enhancement on the samples in the water surface target data set, wherein the data enhancement comprises horizontal turning, random noise, rain generation and fog generation.
3. The method for detecting the cross-domain water surface target based on the feature-based anti-migration and semi-supervised learning of claim 1, wherein in the step 1), a water surface target detection model is constructed, and the method comprises the following steps:
a. constructing a structure of an Resnet 50-based basic detection network, wherein the structure comprises a Resnet 50-based feature extractor, a region proposal network, a region network, region pooling and an objective function;
after the structure of the basic detection network based on Resnet50 is built, pre-training is carried out, and the process comprises the following steps: firstly, inputting source domain data into a feature extractor to obtain a feature map, then generating a candidate region through a region proposing network, inputting the region pooled candidate region features into the region network to further optimize a detection result, and finally adjusting parameters of a water surface target detection model through optimizing a target function of a basic network; the objective function of the basic detection network based on Resnet50 is as follows:
in the formula, D S The source domain data is represented by a representation of,
a loss function representing the area proposed network,
representing the loss function, λ, of the area network 1 To represent
The penalty factor of (2) is determined,
and
both contain cross-entropy loss for classification and the Smooth-L1 function for position regression;
b. building a domain confrontation discriminator based on regions: the method comprises the steps that a gradient inversion layer and a convolution layer are connected in series to construct a region-based domain confrontation discriminator, wherein the output of a region proposal network is used as the input of the region-based domain confrontation discriminator, a region-based domain confrontation objective function is adopted to train the parameters of the region-based domain confrontation discriminator, and the parameters of a water surface target detection model are adjusted in a gradient back propagation mode; wherein the region-based domain confrontation objective function is:
in the formula (I), the compound is shown in the specification,
representing a region-based domain confrontation objective function,
the target domain data is represented by a representation of,
representing a set of candidate areas, p, generated by a network of area proposals c (r) represents the probability that the candidate region r belongs to the domain class c;
c. building a domain confrontation category discriminator: constructing a domain confrontation category discriminator by serially connecting a gradient inversion layer and a convolution layer, wherein the candidate region characteristics after region pooling are used as the input of the domain confrontation category discriminator, the domain confrontation category discriminator is weighted by utilizing the output of a region network, the parameters of the domain confrontation category discriminator are trained by adopting a domain confrontation category target function, and the parameters of a water surface target detection model are adjusted in a gradient back propagation mode; wherein, the domain confrontation category objective function is:
in the formula (I), the compound is shown in the specification,
a domain confrontation category objective function is represented,
a multi-class domain counter-loss function is represented,
representing an entropy function;
the multi-class domain counter-loss function is in the mathematical form:
in the formula (I), the compound is shown in the specification,
representing the total number of target categories of source domain data and target domain data,
a class probability distribution indicating that the candidate region r belongs to class i, F a feature vector of the candidate region r, c a domain class,
representation calculation
And c;
the mathematical form of the entropy function is:
4. the method for detecting the cross-domain water surface target based on the feature-based anti-migration and semi-supervised learning of claim 1, wherein in the step 2), the purpose of the feature-based anti-migration learning is to align the features of the source domain data and the target domain data in the water surface target data set by the water surface target detection model in an anti-manner, for this purpose, a region-based domain anti-collision discriminator is combined to train the feature-based anti-migration learning, and a stochastic gradient descent method is adopted to adjust the parameters of the water surface target detection model, and the objective function is as follows:
in the formula, D S Which represents the data of the source domain,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function, λ 2 To represent
The penalty factor of (2).
5. The method for detecting the cross-domain water surface target based on the feature-based anti-migration and semi-supervised learning according to claim 1, wherein the step 3) comprises the following steps:
3.1) performing pseudo-labeling on part of unmarked target domain data by using the water surface target detection model after the transfer-resistant learning by using the characteristics, and adjusting the accuracy of the pseudo-labeling in a manual mode;
3.2) carrying out semi-supervised training by using a domain confrontation discriminator and a domain confrontation category discriminator based on a region, and adjusting parameters of a water surface target detection model by adopting a random gradient descent method, wherein an objective function is as follows:
in the formula, D S The source domain data is represented by a representation of,
the target domain data is represented by a representation of,
indicating the objective function of detecting the network based on the basis of Resnet50,
representing a region-based domain confrontation objective function,
representing a domain confrontation class objective function, λ 3 And λ 4 Respectively represent
And
the penalty factor of (2);
and finally, deploying a water surface target detection model with convergent training in a new environment to realize accurate cross-domain water surface target detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210747277.5A CN115082792A (en) | 2022-06-29 | 2022-06-29 | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210747277.5A CN115082792A (en) | 2022-06-29 | 2022-06-29 | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115082792A true CN115082792A (en) | 2022-09-20 |
Family
ID=83255397
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210747277.5A Pending CN115082792A (en) | 2022-06-29 | 2022-06-29 | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115082792A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110234085A (en) * | 2019-05-23 | 2019-09-13 | 深圳大学 | Based on the indoor location fingerprint to anti-migration network drawing generating method and system |
| KR20190138238A (en) * | 2018-06-04 | 2019-12-12 | 삼성전자주식회사 | Deep Blind Transfer Learning |
| CN113313166A (en) * | 2021-05-28 | 2021-08-27 | 华南理工大学 | Ship target automatic labeling method based on feature consistency learning |
| CN114241285A (en) * | 2021-11-25 | 2022-03-25 | 华南理工大学 | Ship rapid detection method based on knowledge distillation and semi-supervised learning |
| US20220193915A1 (en) * | 2020-12-18 | 2022-06-23 | Strong Force Vcn Portfolio 2019, Llc | Robotic Vision System with Variable Lens for Value Chain Networks |
| US20220198339A1 (en) * | 2020-12-23 | 2022-06-23 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for training machine learning model based on cross-domain data |
-
2022
- 2022-06-29 CN CN202210747277.5A patent/CN115082792A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20190138238A (en) * | 2018-06-04 | 2019-12-12 | 삼성전자주식회사 | Deep Blind Transfer Learning |
| CN110234085A (en) * | 2019-05-23 | 2019-09-13 | 深圳大学 | Based on the indoor location fingerprint to anti-migration network drawing generating method and system |
| US20220193915A1 (en) * | 2020-12-18 | 2022-06-23 | Strong Force Vcn Portfolio 2019, Llc | Robotic Vision System with Variable Lens for Value Chain Networks |
| US20220198339A1 (en) * | 2020-12-23 | 2022-06-23 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for training machine learning model based on cross-domain data |
| CN113313166A (en) * | 2021-05-28 | 2021-08-27 | 华南理工大学 | Ship target automatic labeling method based on feature consistency learning |
| CN114241285A (en) * | 2021-11-25 | 2022-03-25 | 华南理工大学 | Ship rapid detection method based on knowledge distillation and semi-supervised learning |
Non-Patent Citations (1)
| Title |
|---|
| 周立君;刘宇;白璐;茹志兵;于帅;: "一种基于GAN和自适应迁移学习的样本生成方法", 应用光学, no. 01, 15 January 2020 (2020-01-15) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109919108B (en) | Remote sensing image rapid target detection method based on deep hash auxiliary network | |
| CN109902806B (en) | Method for determining target bounding box of noise image based on convolutional neural network | |
| CN107909082B (en) | Sonar image target identification method based on deep learning technology | |
| CN112308019B (en) | SAR ship target detection method based on network pruning and knowledge distillation | |
| US20230184927A1 (en) | Contextual visual-based sar target detection method and apparatus, and storage medium | |
| CN108710875B (en) | A kind of take photo by plane road vehicle method of counting and device based on deep learning | |
| CN109447008B (en) | Crowd analysis method based on attention mechanism and deformable convolutional neural network | |
| CN112766087A (en) | Optical remote sensing image ship detection method based on knowledge distillation | |
| CN110334705A (en) | A kind of Language Identification of the scene text image of the global and local information of combination | |
| CN113239830B (en) | Remote sensing image cloud detection method based on full-scale feature fusion | |
| CN110633708A (en) | Deep network significance detection method based on global model and local optimization | |
| CN110598586A (en) | Target detection method and system | |
| CN109635726B (en) | Landslide identification method based on combination of symmetric deep network and multi-scale pooling | |
| CN106815323A (en) | A kind of cross-domain vision search method based on conspicuousness detection | |
| CN111753677A (en) | Multi-angle remote sensing ship image target detection method based on characteristic pyramid structure | |
| Li et al. | Robust deep neural networks for road extraction from remote sensing images | |
| CN110334656A (en) | Multi-source Remote Sensing Images Clean water withdraw method and device based on information source probability weight | |
| CN111192240B (en) | Remote sensing image target detection method based on random access memory | |
| CN116110022A (en) | Lightweight traffic sign detection method and system based on response knowledge distillation | |
| CN114549909A (en) | Pseudo label remote sensing image scene classification method based on self-adaptive threshold | |
| CN112556682B (en) | Automatic detection algorithm for underwater composite sensor target | |
| CN113989612A (en) | Remote sensing image target detection method based on attention and generation countermeasure network | |
| CN116503602A (en) | Unstructured environment three-dimensional point cloud semantic segmentation method based on multi-level edge enhancement | |
| CN115082792A (en) | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning | |
| CN116664826A (en) | Small sample point cloud semantic segmentation method |
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 |