CN116645512A - Self-adaptive semantic segmentation method and device under severe conditions - Google Patents
Self-adaptive semantic segmentation method and device under severe conditions Download PDFInfo
- Publication number
- CN116645512A CN116645512A CN202310638992.XA CN202310638992A CN116645512A CN 116645512 A CN116645512 A CN 116645512A CN 202310638992 A CN202310638992 A CN 202310638992A CN 116645512 A CN116645512 A CN 116645512A
- Authority
- CN
- China
- Prior art keywords
- image
- domain data
- model
- target domain
- training
- 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
- 238000000034 method Methods 0.000 title claims abstract description 50
- 230000011218 segmentation Effects 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 46
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 230000000007 visual effect Effects 0.000 claims description 5
- 238000012937 correction Methods 0.000 claims description 4
- 238000010586 diagram Methods 0.000 claims description 4
- 230000005540 biological transmission Effects 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 5
- 230000006870 function Effects 0.000 description 8
- 238000002372 labelling Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- 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
-
- 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
- G06N3/096—Transfer learning
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a self-adaptive semantic segmentation method and device under severe conditions, which belong to the field of semantic segmentation, and are characterized in that pseudo tags are corrected; and then mixing the source domain data and the target domain data with the corrected pseudo tag according to a certain proportion, and training by mixing the source domain data and the target domain data, thereby improving the robustness and the robustness of the model under severe conditions. The method provided by the invention has the advantages that the trained model is used for semantic segmentation, so that the semantic segmentation effect under severe conditions can be improved, and the dependence on the data and the labels of the target domain is reduced.
Description
Technical Field
The invention belongs to the field of semantic segmentation, and particularly relates to a self-adaptive semantic segmentation method and device under severe conditions.
Background
Semantic segmentation plays an important role in the field of automatic driving and intelligent monitoring. In the prior art, a semantic segmentation model based on deep learning has good effect in a clear scene, but has great challenges in semantic segmentation under severe conditions (such as rain, fog, snow, night and the like). This is because in these cases, image collection is difficult, labeling effort is large, and the lack of high quality labeling data volume results in reduced model performance. In the existing method, the semantic segmentation model under a clear scene is generalized to severe conditions by using self-training as a main stream mode in a field self-adaption mode. However, during the self-training process, pseudo-tag noise caused by domain bias of the source domain dataset (the tagged dataset in a clear scene) and the target domain dataset (the untagged dataset in severe conditions) can negatively impact the model performance.
Disclosure of Invention
The invention aims to solve the technical problem of pseudo tag noise under severe conditions, and provides a self-adaptive semantic segmentation method and device under severe conditions.
In order to solve the technical problems, the invention adopts the following technical scheme:
a self-adaptive semantic segmentation method under severe conditions comprises the following steps:
step 1: acquiring a source domain data set and a target domain data set;
step 2: preprocessing image data in a source domain data set and a target domain data set;
step 3: predicting the preprocessed target domain data by using a teacher model, and generating a pseudo tag according to a prediction result;
step 4: correcting the pseudo tag;
step 5: mixing the source domain data and the target domain data with the corrected pseudo tag according to a preset proportion, respectively inputting the mixed domain data into a teacher model and a student model for mixed training, and obtaining a trained model after repeated iterative training;
step 6: and evaluating the model, and performing semantic segmentation on the input image by using the trained student model after the evaluation is passed.
Further, the method for generating the pseudo tag according to the prediction result in the step 3 is as follows:
predicting the target domain data by using a teacher model to obtain probability distribution of each pixel belonging to each category;
and taking the category with the highest probability in each pixel point as a pseudo tag of the pixel point.
Further, the method for correcting the pseudo tag comprises the following steps:
inputting the target domain data and the corresponding pseudo tag into a student model, and using a new pseudo tag predicted by the student model as a pseudo tag after the target domain data is corrected; in the cross entropy loss function of the student model:
the global weight parameter is used for setting the global weight of the source domain as 1, and taking the proportion of the part of the predicted value of the target domain larger than the threshold value as the global weight of the target domain;
the local weight parameter takes the difference of each pixel point in the target domain on the main and auxiliary classifiers as local weight, wherein the main and auxiliary classifiers refer to:
and setting the threshold parameter, wherein the initial value is 1/C, C is the number of categories, and then dynamically adjusting according to the predicted value of each pixel point category.
Further, the method for performing the hybrid training in the step 5 is as follows:
step 5.1: initializing student model and teacher model parameters;
step 5.2, randomly sampling an image and a label of a batch from the source domain data set, and randomly sampling an image of a batch from the target domain data set;
step 5.3: training a cross entropy loss function of a student model by using a mixed image and mixed labels, wherein the mixed image is obtained by taking object pixel points corresponding to one half of classes of class labels from one image of a source domain data set, the rest pixel points are filled with other pixel points in a target domain, and the mixed labels are formed by mixing labels of various classes taken from the source domain image in the mixed image and pseudo labels of the pixel points in the taken target domain image;
step 5.4: training a cross entropy loss function of a student model by using an enhanced mixed image and enhanced mixed labels, wherein the enhanced mixed image is formed by taking half of object pixel points corresponding to classes in class labels from a source domain image, filling the rest pixel points with pixel points in an image after the target domain image is subjected to visual enhancement, and the enhanced mixed labels are formed by mixing labels of various classes taken from the source domain image in the enhanced mixed image and pseudo labels of the pixel points taken from the enhanced target domain image;
step 5.5: updating the student model, and obtaining a new teacher model according to the weighted summation of the student model weight at the current moment and the previous teacher model weight;
step 5.6: stopping training after the training reaches the preset times, and outputting the trained student model.
Further, the method for performing visual enhancement in step 5.4 is:
the image is enhanced using an atmospheric scattering model,
the following formula is adopted for rain, fog and snow:
I(x)＝J(x)t(x)+A(1-t(x))
the following formula was used for the night:
1-I(x)＝(1-J(x))t(x)+A(1-t(x))，
wherein I (x) represents an input original image, J (x) represents an enhanced clear image, a is atmospheric light, and t (x) is a transmission diagram of the original image.
Further, the method for updating the teacher model in step 5.5 is as follows:
ψ t+1 ←αψ t +(1-α)θ t
θ t representing student model after t training iteration, ψ t 、ψ t+1 Respectively representing the teacher model after the t time and the t+1st training iteration, wherein alpha is a momentum coefficient.
Further, the method for dynamically adjusting the threshold parameter is as follows:
and the threshold value of each category is weighted and summed with the average value of the predicted value of each pixel point category at the current moment through the threshold value of the category at the previous moment.
The invention also provides a self-adaptive semantic segmentation device under severe conditions, which comprises the following modules:
an input module: for acquiring a source domain data set and a target domain data set;
and a pretreatment module: the method comprises the steps of preprocessing image data in a source domain data set and a target domain data set, and respectively extracting features of the preprocessed images to obtain a feature map of each training sample;
a pseudo tag generation module: the method comprises the steps of predicting target domain data by using a teacher model, and generating a pseudo tag according to a prediction result;
and a pseudo tag correction module: for correcting the pseudo tag;
and (3) a hybrid training module: the method comprises the steps of mixing source domain data and target domain data with corrected pseudo labels according to a preset proportion, respectively inputting the mixed domain data into a teacher model and a student model for mixed training, and obtaining a trained model after repeated iterative training;
and a segmentation module: the method is used for evaluating the model, and semantic segmentation is carried out on the input image by using the trained model after the evaluation is passed.
By adopting the technical scheme, the invention has the following beneficial effects:
the invention provides a self-adaptive semantic segmentation method and a device under severe conditions, which are characterized in that a pseudo tag is corrected; and then mixing the source domain data and the target domain data with the corrected pseudo tag according to a certain proportion, and training by mixing the source domain data and the target domain data, thereby improving the robustness and the robustness of the model under severe conditions. The semantic segmentation is carried out by using the model trained by the method, so that the semantic segmentation effect under severe conditions can be improved.
Drawings
FIG. 1 is a flow chart of a system of the present invention;
figure 2 shows a schematic diagram of a model framework.
Detailed Description
The following description of the embodiments of the present invention will be made apparent and fully in view of the accompanying drawings, in which some, but not all embodiments of the invention are shown. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
FIG. 1 shows a specific embodiment of an adaptive semantic segmentation method under severe conditions, comprising the following steps:
step 1: a source domain data set and a target domain data set are acquired.
Step 2: image data in the source domain data set and the target domain data set are preprocessed.
In this embodiment, after the data set is acquired, preprocessing is further required for each image, which includes operations such as image size adjustment, data enhancement, etc., where the purpose of image size adjustment is to make the images have the same specification, and the purpose of data enhancement is to enhance the quality of the images under severe conditions.
Step 3: and predicting the preprocessed target domain data by using a teacher model, and generating a pseudo tag according to a prediction result. Because the teacher model is a relatively mature model and does not have a gradient back propagation function, the teacher model is used for carrying out preliminary prediction on the target domain data, and the accuracy of prediction can be improved.
In this embodiment, as shown in fig. 2, the method for generating the pseudo tag according to the prediction result is:
predicting the target domain data by using a teacher model to obtain probability distribution of each pixel belonging to each category;
and taking the category with the highest probability in each pixel point as a pseudo tag of the pixel point.
Step 4: and correcting the pseudo tag.
In this embodiment, the method for correcting the pseudo tag is:
inputting the target domain data and the corresponding pseudo tag into a student model, and using a new pseudo tag predicted by the student model as a pseudo tag after the target domain data is corrected; in the cross entropy loss function of the student model:
the global weight parameter is used for setting the global weight of the source domain as 1, and taking the proportion of the part of the predicted value of the target domain larger than the threshold value as the global weight of the target domain;
and taking the difference of each pixel point in the target domain on the main classifier and the auxiliary classifier as a local weight, wherein the local weight is determined by using the KL divergences of the main classifier F_main and the auxiliary classifier F_auxliary.
And (3) setting a threshold parameter, wherein an initial value is 1/C, C is the number of categories, the initial value is set to be very small, and then dynamically adjusting according to the predicted value of each pixel point category. The method for dynamically adjusting the threshold parameter in the embodiment is as follows: and the threshold value of each category is weighted and summed with the average value of the predicted value of each pixel point category at the current moment through the threshold value of the category at the previous moment. As the number of network training increases, the predicted value for each category is continually adjusted and increases, so too does the threshold. The dynamic change of the threshold value enables the diversity of the pseudo tag to be considered more in the early stage of training, and the credibility of the pseudo tag to be considered more in the later stage, so that the quality of the pseudo tag is further improved.
The correction of the pseudo tag by using the student model is focused on how to take values of all parameters in the cross entropy loss function of the student model, and the global weight and the local weight parameters used in the embodiment are all uncertainty weights and continuously change along with the change of the target domain predicted value, and the threshold parameters are continuously adjusted along with the change of the class predicted value of each pixel point.
Step 5: mixing the source domain data and the target domain data with the corrected pseudo tag according to a preset proportion, respectively inputting the mixed training into a teacher model and a student model, and obtaining a trained model after repeated iterative training.
As shown in the model framework of fig. 2, the method for performing the hybrid training in this embodiment is:
step 5.1: initializing student model and teacher model parameters;
step 5.2, randomly sampling an image and a label of a batch from the source domain data set, and randomly sampling an image of a batch from the target domain data set;
step 5.3: training a cross entropy loss function of a student model by using a mixed image and mixed labels, wherein the mixed image is obtained by taking object pixel points corresponding to one half of classes of class labels from one image of a source domain data set, the rest pixel points are filled with other pixel points in a target domain, and the mixed labels are formed by mixing labels of various classes taken from the source domain image in the mixed image and pseudo labels of the pixel points in the taken target domain image; mixing the source domain data and the target domain data according to a certain proportion, and adding the pseudo tag into the training data at the same time so as to further improve the performance of the model.
Step 5.4: training a cross entropy loss function of a student model by using an enhanced mixed image and enhanced mixed labels, wherein the enhanced mixed image is obtained by taking half of object pixels corresponding to the classes in the class labels from a source domain image, the rest pixels are filled with pixels in the image after the target domain image is visually enhanced, and the enhanced mixed labels are formed by mixing labels of various classes taken in the source domain image in the enhanced mixed image and pseudo labels of the pixels in the taken enhanced target domain image. In this embodiment, the difference between the enhanced hybrid tag and the hybrid tag in step 5.3 is that the pseudo tag in the target domain is used, one is that the pseudo tag predicted by the teacher model for the original target domain in step 5.3, and the pseudo tag predicted by the teacher model for the visually enhanced target data in step 5.4, so that the quality of the pseudo tag is improved.
In this embodiment, the method for performing visual enhancement is:
the atmospheric scattering model is used for enhancing the image, and the following formula is adopted for rain, fog and snow:
I(x)＝J(x)t(x)+A(1-t(x))
the following formula was used for the night:
1-I(x)＝(1-J(x))t(x)+A(1-t(x))，
wherein I (x) represents an input original image, J (x) represents an enhanced clear image, a is atmospheric light, and t (x) is a transmission diagram of the original image. The atmospheric scattering model comes from the atmospheric visibility formula proposed in 1997 by Koschmieder in its paper Theorie der horizontalen Sichtweite. The formula can be used for describing the propagation and scattering of light in the atmosphere, so that the transparency of the atmosphere is calculated, and the haze effect in pictures or videos is further achieved.
Step 5.5: and updating the student model, and updating the teacher model by using EMA momentum according to the weight of the student model at the current moment and the weight of the previous teacher model.
The method for updating the teacher model in this embodiment is:
ψ t+1 ←αψ t +(1-α)θ t
θ t representing student model after t training iteration, ψ t 、ψ t+1 Respectively representing the teacher model after the t time and the t+1st training iteration, wherein alpha is a momentum coefficient, and the alpha value in the embodiment is 0.99.
Step 5.6: stopping training after the training reaches the preset times, and outputting the trained student model.
Step 6: and evaluating the model, and performing semantic segmentation on the input image by using the trained student model after the evaluation is passed.
In this embodiment, a cross-validation method is used to evaluate the model, including calculating indexes such as pixel accuracy, average cross-over ratio (mIoU), etc. of each class.
The embodiment further reduces pseudo tag noise and domain deviation by adopting a dynamic threshold value, uncertainty weight and the like, and trains by mixing source domain and target domain data so as to improve the robustness and robustness of the model under severe conditions. Experimental results show that the technical scheme of the invention can obviously improve the semantic segmentation effect of the model under severe conditions, and has better robustness and robustness. Thereby being more robust in the automatic driving and intelligent monitoring directions by using semantic segmentation models
The invention also provides a self-adaptive semantic segmentation device under severe conditions, which comprises the following modules:
an input module: for acquiring a source domain data set and a target domain data set;
and a pretreatment module: the method comprises the steps of preprocessing image data in a source domain data set and a target domain data set, and respectively extracting features of the preprocessed images to obtain a feature map of each training sample;
a pseudo tag generation module: the method comprises the steps of predicting target domain data by using a teacher model, and generating a pseudo tag according to a prediction result;
and a pseudo tag correction module: for correcting the pseudo tag;
and (3) a hybrid training module: the method comprises the steps of mixing source domain data and target domain data with corrected pseudo labels according to a preset proportion, respectively inputting the mixed domain data into a teacher model and a student model for mixed training, and obtaining a trained model after repeated iterative training;
and a segmentation module: the method is used for evaluating the model, and semantic segmentation is carried out on the input image by using the trained model after the evaluation is passed.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.
Claims (8)
1. The self-adaptive semantic segmentation method under the severe condition is characterized by comprising the following steps of:
step 1: acquiring a source domain data set and a target domain data set;
step 2: preprocessing image data in a source domain data set and a target domain data set;
step 3: predicting the preprocessed target domain data by using a teacher model, and generating a pseudo tag according to a prediction result;
step 4: correcting the pseudo tag;
step 5: mixing the source domain data and the target domain data with the corrected pseudo tag according to a preset proportion, inputting the mixed data into a student model for mixed training, and obtaining a trained student model after repeated iterative training;
step 6: and evaluating the model, and performing semantic segmentation on the input image by using the trained student model after the evaluation is passed.
2. The adaptive semantic segmentation method under severe conditions according to claim 1, wherein the method for generating pseudo tags according to the prediction result in step 3 is as follows:
predicting the target domain data by using a teacher model to obtain probability distribution of each pixel belonging to each category;
and taking the category with the highest probability in each pixel point as a pseudo tag of the pixel point.
3. The method for adaptively partitioning semantics under severe conditions as in claim 2, wherein the method for correcting the pseudo tag is:
inputting the target domain data and the corresponding pseudo tag into a student model, and using a new pseudo tag predicted by the student model as a pseudo tag after the target domain data is corrected; in the cross entropy loss function of the student model:
the global weight parameter is used for setting the global weight of the source domain as 1, and taking the proportion of the part of the predicted value of the target domain larger than the threshold value as the global weight of the target domain;
the local weight parameter takes the difference of each pixel point in the target domain on the main and auxiliary classifiers as local weight, wherein the main and auxiliary classifiers refer to:
and setting the threshold parameter, wherein the initial value is 1/C, C is the number of categories, and then dynamically adjusting according to the predicted value of each pixel point category.
4. A method for adaptive semantic segmentation under severe conditions according to claim 3, wherein the method for performing hybrid training in step 5 is:
step 5.1: initializing student model and teacher model parameters;
step 5.2, randomly sampling an image and a label of a batch from the source domain data set, and randomly sampling an image of a batch from the target domain data set;
step 5.3: training a cross entropy loss function of a student model by using a mixed image and mixed labels, wherein the mixed image is obtained by taking object pixel points corresponding to one half of classes of class labels from one image of a source domain data set, the rest pixel points are filled with other pixel points in a target domain, and the mixed labels are formed by mixing labels of various classes taken from the source domain image in the mixed image and pseudo labels of the pixel points in the taken target domain image;
step 5.4: training a cross entropy loss function of a student model by using an enhanced mixed image and enhanced mixed labels, wherein the enhanced mixed image is formed by taking half of object pixel points corresponding to classes in class labels from a source domain image, filling the rest pixel points with pixel points in an image after the target domain image is subjected to visual enhancement, and the enhanced mixed labels are formed by mixing labels of various classes taken from the source domain image in the enhanced mixed image and pseudo labels of the pixel points taken from the enhanced target domain image;
step 5.5: updating the student model, and obtaining a new teacher model according to the weighted summation of the student model weight at the current moment and the previous teacher model weight;
step 5.6: stopping training after the training reaches the preset times, and outputting the trained student model.
5. The method for adaptive semantic segmentation under severe conditions according to claim 4, wherein the method for visual enhancement in step 5.4 is as follows:
the image is enhanced using an atmospheric scattering model,
the following formula is adopted for rain, fog and snow:
I(x)＝J(x)t(x)+A(1-t(x))
the following formula was used for the night:
1-I(x)＝(1-J(x))t(x)+A(1-t(x))，
wherein I (x) represents an input original image, J (x) represents an enhanced clear image, a is atmospheric light, and t (x) is a transmission diagram of the original image.
6. The adaptive semantic segmentation method under severe conditions according to claim 4, wherein the method for updating the teacher model in step 5.5 is as follows:
ψ t+1 ←αψ t +(1-α)θ t
θ t representing student model after t training iteration, ψ t 、ψ t+1 Respectively representing the teacher model after the t time and the t+1st training iteration, wherein alpha is a momentum coefficient.
7. A method for adaptive semantic segmentation under severe conditions according to claim 3, wherein the method for dynamically adjusting the threshold parameter is:
and the threshold value of each category is weighted and summed with the average value of the predicted value of each pixel point category at the current moment through the threshold value of the category at the previous moment.
8. The self-adaptive semantic segmentation device under severe conditions is characterized by comprising the following modules:
an input module: for acquiring a source domain data set and a target domain data set;
and a pretreatment module: the method comprises the steps of preprocessing image data in a source domain data set and a target domain data set, and respectively extracting features of the preprocessed images to obtain a feature map of each training sample;
a pseudo tag generation module: the method comprises the steps of predicting target domain data by using a teacher model, and generating a pseudo tag according to a prediction result;
and a pseudo tag correction module: for correcting the pseudo tag;
and (3) a hybrid training module: the method comprises the steps of mixing source domain data and target domain data with corrected pseudo labels according to a preset proportion, respectively inputting the mixed domain data into a teacher model and a student model for mixed training, and obtaining a trained model after repeated iterative training;
and a segmentation module: the method is used for evaluating the model, and semantic segmentation is carried out on the input image by using the trained model after the evaluation is passed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310638992.XA CN116645512A (en) | 2023-06-01 | 2023-06-01 | Self-adaptive semantic segmentation method and device under severe conditions |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310638992.XA CN116645512A (en) | 2023-06-01 | 2023-06-01 | Self-adaptive semantic segmentation method and device under severe conditions |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116645512A true CN116645512A (en) | 2023-08-25 |
Family
ID=87624360
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310638992.XA Pending CN116645512A (en) | 2023-06-01 | 2023-06-01 | Self-adaptive semantic segmentation method and device under severe conditions |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116645512A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117636086A (en) * | 2023-10-13 | 2024-03-01 | 中国科学院自动化研究所 | Passive domain adaptive target detection method and device |
-
2023
- 2023-06-01 CN CN202310638992.XA patent/CN116645512A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117636086A (en) * | 2023-10-13 | 2024-03-01 | 中国科学院自动化研究所 | Passive domain adaptive target detection method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108986050B (en) | Image and video enhancement method based on multi-branch convolutional neural network | |
| CN113688723B (en) | Infrared image pedestrian target detection method based on improved YOLOv5 | |
| WO2022111219A1 (en) | Domain adaptation device operation and maintenance system and method | |
| CN114022432B (en) | Insulator defect detection method based on improved yolov5 | |
| CN108229550B (en) | Cloud picture classification method based on multi-granularity cascade forest network | |
| CN111160553B (en) | Novel field self-adaptive learning method | |
| CN116645512A (en) | Self-adaptive semantic segmentation method and device under severe conditions | |
| CN110211052A (en) | A kind of single image to the fog method based on feature learning | |
| CN114067173A (en) | Small sample low-quality image target detection method based on multi-definition integrated self-training | |
| CN111680702A (en) | Method for realizing weak supervision image significance detection by using detection frame | |
| CN116563687A (en) | Teacher-student network method for semi-supervised directivity target detection | |
| CN112016594B (en) | Collaborative training method based on field self-adaption | |
| CN116468938A (en) | Robust image classification method on label noisy data | |
| CN112417752A (en) | Cloud layer track prediction method and system based on convolution LSTM neural network | |
| CN104318528A (en) | Foggy weather image restoration method based on multi-scale WLS filtering | |
| CN114596477A (en) | Foggy day train fault detection method based on field self-adaption and attention mechanism | |
| CN114663665A (en) | Gradient-based confrontation sample generation method and system | |
| CN113280820A (en) | Orchard visual navigation path extraction method and system based on neural network | |
| US20230342938A1 (en) | Adaptive Semi-Supervised Image Segmentation Method Based on Uncertainty Knowledge Domain and System thereof | |
| CN113077438A (en) | Cell nucleus region extraction method and imaging method for multi-cell nucleus color image | |
| CN116993975A (en) | Panoramic camera semantic segmentation method based on deep learning unsupervised field adaptation | |
| CN116433909A (en) | Similarity weighted multi-teacher network model-based semi-supervised image semantic segmentation method | |
| CN115797904A (en) | Active learning method for multiple scenes and multiple tasks in intelligent driving visual perception | |
| CN113256517A (en) | Video rain removing method based on semi-supervised probability map model | |
| CN114862724A (en) | Contrast type image defogging method based on exponential moving average knowledge distillation |
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 |