CN117409250A - Small sample target detection method, device and medium - Google Patents
Small sample target detection method, device and medium Download PDFInfo
- Publication number
- CN117409250A CN117409250A CN202311412809.0A CN202311412809A CN117409250A CN 117409250 A CN117409250 A CN 117409250A CN 202311412809 A CN202311412809 A CN 202311412809A CN 117409250 A CN117409250 A CN 117409250A
- Authority
- CN
- China
- Prior art keywords
- small sample
- target
- target detection
- training
- model
- 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 198
- 238000000034 method Methods 0.000 claims abstract description 55
- 238000012549 training Methods 0.000 claims description 92
- 230000011218 segmentation Effects 0.000 claims description 39
- 238000009792 diffusion process Methods 0.000 claims description 29
- 238000012360 testing method Methods 0.000 claims description 25
- 238000013507 mapping Methods 0.000 claims description 18
- 238000009826 distribution Methods 0.000 claims description 11
- 238000004422 calculation algorithm Methods 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 238000002372 labelling Methods 0.000 claims description 6
- 230000001373 regressive effect Effects 0.000 claims description 6
- 230000002708 enhancing effect Effects 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 3
- 238000013434 data augmentation Methods 0.000 claims description 3
- 238000011084 recovery Methods 0.000 claims description 3
- 238000003860 storage Methods 0.000 claims description 3
- 238000013508 migration Methods 0.000 abstract description 2
- 230000005012 migration Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 3
- 206010026749 Mania Diseases 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008014 freezing Effects 0.000 description 1
- 238000007710 freezing Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000009966 trimming 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
- G06V10/70—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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised 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
- G06V10/20—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
- 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
- G06V10/766—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using regression, e.g. by projecting features on hyperplanes
-
- 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
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of target detection, and particularly discloses a method, a device and a medium for detecting a small sample target, wherein the method can be used for detecting small sample class data D with only a small quantity of marks novel Under the condition of (1), higher detection performance is realized, and other existing public data sets with a large number of labels can be fully utilized. Specifically, first, base class data D having a large number of labels is used base To pretrain the model and then to annotate the small sample class data D in small quantities novel And the method is optimized, realizes knowledge migration of the model, and has the capability of detecting the small sample targets.
Description
Technical Field
The invention belongs to the technical field of target detection, and particularly relates to a method, a device and a medium for detecting a small sample target.
Background
Object detection based on deep learning has made great progress in recent years, but often requires a large number of annotated image training neural networks. In the occasions of industrial product flaw detection, medical image recognition and the like, a large number of images are difficult to label, and only a small number of labeled images can be provided. The method for detecting the target by the small sample has wide application value by researching how to learn the target detection model from a small quantity of marked image data.
The closest, current state of the art in connection with this application includes:
1. the literature "XinWang, thomas e.huang, trevor Darrell, joseph E Gonzalez, and Fisher yu.frame Simple Few-Shot Object detection, pages1-12,2020," discloses a method for fusion-based pre-training and small sample-based fine-tuning of small sample targets: based on the Faster RCNN network architecture, the system is first pre-trained on a base class with a large amount of data, and then fine-tuned on a small sample class with a small amount of data.
Disadvantages: (1) The fast RCNN used in the pre-training is seriously dependent on the RPN module to generate a high-quality target candidate region, so that the detection accuracy is not high; (2) Trimming directly on the small sample class data after pre-training tends to result in overfitting due to the small sample class data.
2. Document "Shoufa Chen, peize Sun, yibing Song and Ping Luo. Diffuse: diffusion model for object detection. ArXiv preprint arXiv:2211.09788,2022" discloses a target detection method based on a diffusion model. Applying a diffusion model to the target detection; during training, noise is randomly added to the true value of the boundary frame, the diffusion process from the true value of the boundary frame to the random frame is realized, and the true value of the boundary frame is recovered through the inverse process of learning and noise diffusion; and generating candidate frames completely randomly during testing, and learning target boundary frames from the candidate frames through the inverse process of noise adding and diffusion.
Disadvantages: (1) The diffusion model is not limited during training and testing, and the generated random frames are particularly large; (2) During training, the random frames are particularly more, so that model learning is unstable and convergence is slow; (3) During testing, the reasoning speed is slow due to the fact that a plurality of random frames are generated completely randomly. (4) A significant amount of computing resources are required for both training and testing.
3. The literature Sangdoo Yun, dongyoon Han, seong Joon Oh, sanghyuk Chun, junsuk Chue, and Yongjoon Yoo. Cutmix: regularization strategy to train strong classifiers with localizable features, in Proceedings of the IEEE/CVF international conference on computer vision, pag-es6023-6032,2019, [ Alexey Bochkovskiy, chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: optimal speed and accuracy of object detection. ArXiv preprint arXv: 2004.10934,2020 ] discloses a data enhancement method based on block mapping: and selecting the whole or part of the bounding box from the bounding box according to the bounding box label of the target on the training image to obtain a box, and pasting the foreground and the background in the whole box to other images.
Disadvantages: and (3) based on the data enhancement of the block mapping, pasting the foreground in the block, and pasting the background in the block to other images, so that the pasted background area is not matched with the background area, and the enhanced image lacks reality.
4. The literature "Rother C, kolmogorov V, blake A." GrabCyt ": interactive foreground extraction using iter-ated graph cuts.ACM Transactions On Graphics (TOG), 2004,23 (3): 309-314." discloses a target segmentation method based on GrabCyt: assuming that the regions outside the bounding box are all "definite background", the pixels within the bounding box are "possible foreground", and segmentation is performed by the gaussian mixture model and the graph cut optimization.
Disadvantages: the segmented foreground tends to be insufficiently complete.
5. Document Prannay Kaul, weidi Xie, and Andrew Zisselman. Label, verify, correct: A simple few shot object detection method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14237-14247,2022, discloses a data augmentation method based on small sample class data annotation generation: firstly, an initial target detection model is trained on small sample class data, then reasoning is carried out on base class data, then a group of small sample class data labels are generated based on a KNN classifier, and the small sample class data are expanded.
Disadvantages: only the initial small sample target detection model is utilized, the basic class target detection pre-training model is not utilized, and wrong small sample class data labels are often generated.
Based on this, how to accurately learn the target detection model from a small amount of labeled image data is a technical problem that needs to be solved.
Disclosure of Invention
The present invention has been made to solve the above-mentioned problems occurring in the prior art. Therefore, a small sample target detection method, device and medium are needed to improve the target detection accuracy of small samples.
According to a first aspect of the present invention, there is provided a small sample target detection method, the method comprising:
target detection pre-training based on a diffusion model and target probability estimation;
merging element learning and small sample target detection of model fine adjustment;
pixel level data enhancement based on iterative GrabCut and random mapping;
expanding small sample class data based on contrast learning;
performing small sample target detection training, data enhancement and data expansion iteratively to obtain a final small sample target detection model;
and performing target detection by using a small sample target detection model, taking an image to be detected as input, and outputting a small sample class target in the image to be detected.
Further, the target detection pre-training based on the diffusion model and the target probability estimation specifically comprises the following steps:
base class data D based on a large number of labels base Pre-training a basic class target detection model M base To learn general knowledge of target detection.
Further, based on a large number of marked base class data D base A basic category target detection model M is pre-trained by the following method base ：
In the model training phase, for each input image:
extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
adding random noise on the marked boundary frame to realize the diffusion process from the true value of the boundary frame to the random distribution frame;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
and extracting the characteristics of the target candidate frame region from the characteristic map according to the position of each target candidate frame, learning the inverse process of the corresponding boundary frame true value and the noise diffusion through a position regressive device, realizing the recovery of the boundary frame true value from the random frame generated by the noise, and learning the marked target category based on the characteristics of the target candidate frame region through a target classifier.
Further, after the model training phase is completed, a model test phase is further included in which, for each input image:
Extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
generating a random frame in a completely random manner;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
extracting the characteristics of the target candidate frame areas from the characteristic map according to the position of each target candidate frame, and learning the marked target category based on the characteristics of the target candidate frame areas through a target classifier;
further, the small sample target detection of the fusion element learning and the model fine tuning specifically comprises:
construction of a Small sample target detection network M novel The small sample target detection network comprises an image encoder, a region feature extractor, a target probability estimator, a target classifier and a position regressive device, wherein the image encoder, the region feature extractor and the target probability estimator utilize a basic class target detection model M base Corresponding parameters in the step (a) are configured;
from base class data D base Target detection task T for randomly extracting a batch of small samples i { i=1, 2 … … }, for eachTarget detection task T for small samples i The number of target classes, the number of training and test images for each class are all equal to the small sample class data D novel Identical and divided into training data sets DTR i And test dataset DTE i The method comprises the steps of carrying out a first treatment on the surface of the In training data set DTR i Based on the small sample target detection network, the initial parameters of the target classifier and the position regressor are theta, and the training data set DTR of the network is calculated i The target detection loss is obtained by gradient descent to obtain updated parameters, and the updated parameters are calculated in a test data set DTE i Target detection loss L on i The method comprises the steps of carrying out a first treatment on the surface of the Target detection loss L obtained on target detection task of a batch of small samples extracted randomly i Averaging to obtain average detection loss L; gradient descending updating is carried out on the gradient of the current parameter theta according to the average detection loss L, and the updated parameter theta is obtained;
in case the average detection loss L is no longer reduced or is smaller than the set threshold value, the final parameter is obtained as the optimal initial parameter theta of the target classifier and the position regressor opt ；
With optimal initial parameters theta opt Configuring a target classifier and a position regressor in small sample class data D novel And (5) retraining to obtain a trained small sample target detection network, and realizing small sample target detection by using the trained small sample target detection network.
Further, the pixel level data enhancement based on the iterative GrabCut and the random mapping specifically includes:
A small sample class input image I and a bounding box label B are acquired,
clipping a region I 'from the input image I, wherein the size of the region I' is larger than that of the boundary box label B;
taking the area outside the boundary box as a determined background, and taking pixels in the boundary box as possible foreground to obtain an initial mask for the first segmentation;
according to the initial mask, executing GrabCut algorithm once to obtain a first segmentation result mask, wherein each pixel is endowed with three labels of a possible foreground, a possible background and a determined background;
assigning pixels marked as possible foreground as determined foreground according to the current segmentation result; for each row or column in the bounding box B, if all pixels are possible backgrounds, assigning the row or column pixel labels as possible foreground to obtain an initial mask for the next segmentation;
based on the initial mask of the next segmentation, executing the GrabCut algorithm again to obtain a next segmentation result;
in the final segmentation result, labeling the pixels as the determined foreground or the pixels of the possible foreground as the foreground region;
and extracting a pixel region corresponding to the foreground in the final segmentation result from the input image I, performing random scaling, and pasting the pixel region into an image containing the same type of targets as the input image I according to the random position.
Further, the small sample class data expansion based on contrast learning specifically includes:
detection of model M using pre-trained base class targets base Reasoning is carried out on other non-marked images to obtain a detection result R of the basic category target class base ；
Using a small sample target detection model M novel Reasoning is carried out on other non-marked images to obtain a detection result R of the small sample class target class novel ；
For R novel Each of which is scored as a target R of s, if the targets R and R base The overlapping proportion of a target r ' with a score of s ' is larger than a set threshold value, the score of s is n times larger than that of s ', wherein n is more than or equal to 0.5, and the result of the target r is used as the label of the label-free image and is added into a small sample class data set D novel Is a kind of medium.
Further, the iteration performs small sample target detection training, data enhancement and data expansion to obtain a final small sample target detection model, which specifically comprises the following steps:
1) Base class data set D marked with a large amount of data base On, pretraining a basic category target detection model M base ；
2) Small sample class data set D marked with small amount of data novel Training a small sample class target detection model M based on meta-learning and fine tuning novel ；
3) Small sample class data set D novel The pixel level data of the iteration GrabCut and the random mapping are enhanced to obtain an enhanced small sample class data set D novel ；
4) Finding unlabeled small sample class targets from other data sets based on small sample class data expansion of contrast learning, and adding the unlabeled small sample class targets to a small sample class data set D novel Obtaining the expanded small sample class data set D novel ；
5) Repeating the steps 2) to 4), iteratively training a new small sample target detection model, and enhancing and expanding small sample class data until the result converges.
According to a second aspect of the present invention, there is provided a small sample target detection apparatus, the apparatus comprising:
a first training module configured to pre-train target detection based on a diffusion model and a target probability estimate;
the second training module is configured to fuse the small sample target detection of the element learning and the model fine adjustment;
a data enhancement module configured to enhance pixel-level data based on the iterative GrabCut and the random map;
the data expansion module is configured to expand the small sample class data based on the contrast learning;
the iteration training module is configured to iterate small sample target detection training, data enhancement and data expansion to obtain a final small sample target detection model;
the target detection module is configured to detect targets by using a small sample target detection model, takes an image to be detected as input, and outputs small sample class targets in the image to be detected.
According to a third aspect of the present invention, there is provided a readable storage medium storing one or more programs executable by one or more processors to implement the method as described above.
The invention has at least the following beneficial effects:
1) And a small sample target detection method integrating element learning and model fine tuning. On the basis of target detection pre-training based on a diffusion model, for parameters needing fine adjustment, firstly, the optimal initialization parameters are found through meta learning and then fine adjustment is performed. The target detection pre-training model has higher detection performance, and the target detection of a small sample avoids overfitting as much as possible, so that the generalization performance of the target detection pre-training model is enhanced.
2) The method comprises the steps of a target detection pre-training method based on a diffusion model and target probability estimation. And sampling and generating a target candidate frame on the basis of the manized random frame according to the target probability of each region. The number of target candidate frames is greatly reduced, and the target candidate frames are mainly distributed in a region with high target probability; in the training stage, model learning is more stable and converges more quickly; in the test stage, the speed of model test reasoning is obviously improved. The need for computing resources is greatly reduced.
3) A data enhancement step based on pixel level mapping. Based on the bounding box labeling, a pixel-level segmentation is obtained, mapping only the foreground region to other images. The enhanced image is matched and coordinated in each region, and is closer to a real image.
4) A target segmentation step based on iterative GrabCut. According to the current segmentation result and the definition of the boundary box label, searching a foreground region which is possibly not segmented, and performing GrabCot segmentation iteratively to segment a complete foreground.
5) And (3) a small sample class data expansion step based on contrast learning. And finding a small sample class target on the unlabeled data set through the small sample target detection model and the basic class target detection pre-training model, and expanding small sample class data. On the unlabeled data set, the discovered small sample targets and the generated labels thereof are more reliable, and the detection performance of the small sample targets can be enhanced by expanding the targets to the small sample data set.
Drawings
FIG. 1 is a general flow chart of a small sample target detection method according to an embodiment of the invention;
FIG. 2 is a flow chart of a target detection pre-training based on a diffusion model and target probability estimation in accordance with an embodiment of the present invention;
FIG. 3 is a flow chart of a small sample target detection for fusion learning and model fine tuning according to an embodiment of the present invention;
FIG. 4 is a diagram of an example of pixel-level data enhancement based on iterative GrabCut and random mapping, with white, light gray, dark gray, and black in the mask representing "determined foreground", "possible background", "determined background", respectively, in accordance with an embodiment of the present invention;
FIG. 5 is a flow chart of small sample class annotation generation based on contrast learning, extending into a small sample dataset, according to an embodiment of the invention;
fig. 6 is a block diagram of a small sample object detection device according to an embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the drawings and detailed description to enable those skilled in the art to better understand the technical scheme of the present invention. Embodiments of the present invention will be described in further detail below with reference to the drawings and specific examples, but not by way of limitation. The order in which the steps are described herein by way of example should not be construed as limiting if there is no necessity for a relationship between each other, and it should be understood by those skilled in the art that the steps may be sequentially modified without disrupting the logic of each other so that the overall process is not realized.
Object detection based on deep learning has made great progress in recent years, but often requires a large number of annotated image training neural networks. In the situations of industrial product flaw detection, medical image recognition and the like, a large number of marked images are difficult to obtain, and only a small number of marked images can be provided. Small sample target detection (Few Shot Object Detection) aims to achieve target detection in an image by labeling a small number of samples.
Based on the above, the embodiment of the invention provides a small sample targetDetection method, giving small sample class data D with only a small quantity of labels novel To achieve higher detection performance, other existing public data sets (such as Pascal, COCO data sets, commonly referred to as base class data D) with a large number of labels can be fully utilized base ). Specifically, first, base class data D having a large number of labels is used base To pretrain the model and then to annotate the small sample class data D in small quantities novel And the method is optimized, realizes knowledge migration of the model, and has the capability of detecting the small sample targets. Small sample class data D novel And base class data D base The object classes contained are mutually exclusive, but the form of labeling is the same, i.e., each object in the image is labeled with its bounding box and class.
The general flow of the method is shown in figure 1, and specifically comprises the following steps 1-6.
And step 1, target detection pre-training based on a diffusion model and target probability estimation.
In the basic class data D with a large number of labels base On, pretraining a basic category target detection model M base Learning general knowledge of target detection, a flowchart is shown in fig. 2. In the training phase, for each image:
1) Taking VGG, resNet and other neural networks as image encoders to extract feature images;
2) Estimating the probability that each region is a target on the basis of the feature map;
3) Adding random noise to the marked boundary frame to realize the diffusion process from the true value of the boundary frame to the random distribution frame;
4) Sampling and generating target candidate frames according to the target probability of each region on the basis of a large number of random distribution frames, so that the total number of the target candidate frames can be reduced and the target candidate frames are mainly distributed in regions with large target probability;
5) And extracting the characteristics of the target candidate frame area from the characteristic diagram according to the position of each target candidate frame, learning the inverse process of the truth value and the manic diffusion of the boundary frame through a position regressive, namely recovering the truth value of the boundary frame from the random frame generated by manic addition, and learning the marked target category through a target classifier.
In the model test stage, a random frame is generated in a completely random mode, which is different from the mode that a random frame is generated by adding noise to a true value of a boundary frame in the training stage, and other processes are the same as those in the training stage.
The method combines the diffusion model and the target probability estimation, increases the constraint of the target probability estimation on the basis of generating a random frame through the diffusion model, and determines the number of sampling target candidate frames from the random frame according to the size of the target probability of each region, and has the following advantages: (1) The total number of target candidate frames can be reduced, and the target candidate frames are mainly distributed in the area with high target probability. (2) In the training stage, as the target candidate frames are mainly distributed in the region with high target probability, the inverse process of model learning diffusion is more stable and converges more quickly. (3) In the test stage, the number of target candidate frames is greatly reduced, so that the speed of model test reasoning is obviously improved. (4) substantially reducing the need for computing resources.
And 2, detecting a small sample target by fusion element learning and model fine adjustment.
The small sample target detection flow is shown in fig. 3, and the network structure is basically the same as the target detection pre-training on the base class data in fig. 2, but the training flow is different.
1) Since the image feature coding, the region feature extraction and the target probability estimation are all independent of the category, the model M is pre-trained base From a large number of base class data D base The learned knowledge in the process can be directly transferred to a small sample class, so that the parameters of an image encoder, a region feature extractor and target probability estimation in small sample target detection are directly used for a pre-training model M base And freezing the part of the parameters unchanged;
2) The target classifier and the position regressor need to be used for small sample class data D novel Training study was performed, but the data was less, and direct training easily resulted in overfitting. First pass through base class data D base And obtaining optimal initialization parameters by upper element learning. The initial parameters theta of the target classifier and the position regressor are random values;
3) From base class dataD base Target detection task T for randomly extracting a batch of small samples i { i=1, 2 … … }. Target detection task T for each small sample i The number of target classes, the number of training and test images per class, and the small sample class data D novel Identical and divided into training data sets DTR i And test dataset DTE i The method comprises the steps of carrying out a first treatment on the surface of the In training data set DTR i In the above, with the small sample target detection network in fig. 3, the initial parameters of the target classifier and the position regressor are θ, and the calculation network is in the training data set DTR i The target detection loss is obtained by gradient descent to obtain updated parameters, and the updated parameters are calculated in a test data set DTE i Target detection loss L on i The method comprises the steps of carrying out a first treatment on the surface of the Target detection loss L obtained on target detection task of a batch of small samples extracted randomly i Averaging to obtain average detection loss L; and gradient descending updating is carried out on the gradient of the current parameter theta according to the average detection loss L, so that the updated parameter theta is obtained.
4) The above operation step 3) is repeated until the average detection loss L is no longer reduced or less than a given threshold. The final parameter theta is the optimal initial parameter theta of the target classifier and the position regressor opt 。
5) In small sample class data D novel With the small sample target detection network of fig. 3, the target classifier and location regressor initial parameters use the optimal initial parameters θ opt Training a small sample target detection network M novel 。
And 3, enhancing pixel level data based on the iteration GrabCut and the random mapping.
In small sample target detection, data enhancement helps to enhance the diversity of small sample class data and generalization of the model. One common method of data enhancement is block level mapping, which scales and then pastes all regions within the bounding box to the background region, as shown in fig. 4 (j). In this example, the sky background area within the bounding box is also scaled and pasted onto the horse's tree background, significantly not conforming to the real scene of the image. Aiming at the defects of the existing method, a data enhancement method based on pixel-level foreground segmentation and mapping is provided. The pixel level foreground segmentation can be realized through the GrabCut algorithm, but a complete target is not obtained, and an iterative GrabCut algorithm is provided for the defect. The specific flow is as follows:
Given a small sample class input image I and bounding box label B, as shown in FIG. 4 (a);
1) Firstly, cutting a region I' with simpler background from an input image I, wherein the size of the region I is 1.2 times of that of a boundary box label B, as shown in (B) of fig. 4;
2) According to the definition of the bounding box, the areas outside the bounding box are all "definite backgrounds", the pixels in the bounding box are "possible prospects", and the initial Mask (Mask) for the first segmentation is obtained as shown in (c) of fig. 4;
3) Performing GrabCot algorithm once according to the initial mask to obtain a first segmentation result mask, wherein each pixel is endowed with three labels of possible foreground, possible background and determined background, as shown in (d) of fig. 4;
4) According to the current segmentation result, some pixels are marked as 'possible foreground', and in order to fix the pixels as foreground pixels which are not changed any more, the labels of the pixels are assigned as 'determined foreground'; for each row or column within bounding box B, if all pixels are "possible background", a definition that does not match the bounding box indicates that the foreground exists in that row or column without segmentation, the row or column pixel label is assigned to "possible foreground". Such an initial mask for the next division is obtained as shown in fig. 4 (e);
5) Executing the GrabCut algorithm again to obtain the next segmentation result, as shown in (f) of FIG. 4;
6) Repeating steps 4) and 5) until convergence, wherein the pixel labels are either "definite foreground" or "possible foreground" in the final segmentation result, and are regarded as foreground regions, as shown by the white regions in fig. 4 (g);
7) And (3) extracting a pixel area corresponding to the foreground in the final segmentation result from the input image I, randomly scaling, and pasting the pixel area into an image (as shown in (h) in fig. 4, a pasted image can also be the input image I) containing the same type of target with the input image I according to a random position, wherein the result is shown in (I) in fig. 4. The image generated by the pixel-level data enhancement is more realistic than the square frame-level data enhancement in fig. 4 (j).
And 4, expanding small sample class data based on contrast learning.
Considering that a large number of unlabeled images exist on the Internet, unlabeled small sample targets can exist in the images, and based on a contrast learning method, unlabeled small sample targets are found and added into a small sample data set, so that the small sample data is further expanded.
1) Detection of model M using pre-trained base class targets base Reasoning is carried out on other non-marked images I to obtain a detection result R of the basic category target category base ；
2) Using a small sample target detection model M novel Reasoning is carried out on other non-marked images I to obtain a detection result R of the small sample class target class novel ；
3) Assuming that an unlabeled small sample class target exists on the unlabeled image I, if the model M is detected by the base class target base False detection is a basic class target, and classification scores are lower due to class judgment errors, if the model M is detected by a small sample class target at the same time novel The target score for correctly detecting a small sample would be high. According to this property, for R novel Is a target R of s if it is equal to R base Some of the targets r ' scoring s ' have a higher overlap ratio (e.g., greater than 80%) and the score s is much greater than s ' (e.g., 1 times greater), meaning that the target r is likely to be a small sample class target in the unlabeled image I, the result of the target r can be added to the small sample class dataset D as a label for the unlabeled image I novel Thereby realizing the expansion of the small sample class data set.
An example is shown in FIG. 5, where (b) and (c) in FIG. 5 are each the base class target detector M base And a small sample class object detector M novel The results of the detection are shown in FIG. 5 (c)The car with a score of 0.75 in (a) overlaps with the gluck with a score of 0.2 in (b) and the score is much higher, and the car detected in (c) in fig. 5 is regarded as a small sample class label, which can be added to the small sample class data D novel And (3) realizing the expansion of small sample data.
And 5, carrying out small sample target detection training, data enhancement and data expansion in an iteration mode to obtain a final small sample target detection model.
The iterative optimization flow comprises the following steps:
1) Base class data set D marked with a large amount of data base On, pretraining a basic category target detection model M base ；
2) Small sample class data set D marked with small amount of data novel Training a small sample class target detection model M based on meta-learning and fine tuning novel ；
3) Small sample class data set D novel The pixel level data of the iteration GrabCut and the random mapping are enhanced to obtain an enhanced small sample class data set D novel ；
4) Finding unlabeled small sample class targets from other data sets based on small sample class data expansion of contrast learning, and adding the unlabeled small sample class targets to a small sample class data set D novel Obtaining the expanded small sample class data set D novel ；
5) Repeating the steps 2) to 4), iteratively training a new small sample target detection model, and enhancing and expanding small sample class data until the result converges.
And 6, performing target detection by using a small sample target detection model, inputting an image to be detected, and detecting a small sample class target in the image to be detected.
In summary, the advantages of the present invention are as follows:
1. the small sample target detection method provided by the invention is more stable and faster in convergence in the model training stage, the model reasoning speed in the test stage is obviously improved, and the demand on computing resources is greatly reduced.
2. The provided small sample target detection method can avoid overfitting as much as possible and enhance generalization performance.
3. The pixel data enhancement method based on the iteration GrabCut and the mapping is provided, the enhanced image is coordinated with each region, the image is closer to a real image, the diversity of training data can be increased, and the detection performance and generalization capability of the model are improved.
4. According to the small sample data expansion method based on contrast learning, targets are found from a large amount of unlabeled data on the Internet and are automatically labeled, training data can be increased, and the detection performance and generalization capability of the model are improved.
An embodiment of the present invention provides a small sample target detection apparatus, as shown in fig. 6, the apparatus 600 includes:
a first training module 601 configured to target detection pre-training based on a diffusion model and target probability estimates;
A second training module 602 configured to fuse small sample target detection for meta-learning and model fine tuning;
a data enhancement module 603 configured to enhance pixel-level data based on the iterative GrabCut and random mapping;
a data expansion module 604 configured to expand the small sample class data based on contrast learning;
the iterative training module 605 is configured to iteratively perform small sample target detection training, data enhancement and data expansion to obtain a final small sample target detection model;
the target detection module 606 is configured to perform target detection by using a small sample target detection model, take an image to be detected as an input, and output a small sample class target in the image to be detected.
In some embodiments, the first training module is further configured to:
base class data D based on a large number of labels base Pre-training a basic class target detection model M base To learn general knowledge of target detection.
In some embodiments, the first training module is further configured to:
in the model training phase, for each input image:
extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
Adding random noise on the marked boundary frame to realize the diffusion process from the true value of the boundary frame to the random distribution frame;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
and extracting the characteristics of the target candidate frame region from the characteristic map according to the position of each target candidate frame, learning the inverse process of the corresponding boundary frame true value and the noise diffusion through a position regressive device, realizing the recovery of the boundary frame true value from the random frame generated by the noise, and learning the marked target category based on the characteristics of the target candidate frame region through a target classifier.
In some embodiments, the first training module is further configured to, after completion of the model training phase, further comprise a model testing phase in which, for each input image:
extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
generating a random frame in a completely random manner;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
extracting the characteristics of the target candidate frame areas from the characteristic map according to the position of each target candidate frame, and learning the marked target category based on the characteristics of the target candidate frame areas through a target classifier;
In some embodiments, the second training module is further configured to:
construction of a Small sample target detection network M novel The small sample target detection network comprises an image encoder, a region feature extractor, a target probability estimator, a target classifier and a position regressive deviceThe image encoder, region feature extractor and target probability estimator utilize a base class target detection model M base Corresponding parameters in the step (a) are configured;
from base class data D base Target detection task T for randomly extracting a batch of small samples i { i=1, 2 … … }, target detection task T for each small sample i The number of target classes, the number of training and test images for each class are all equal to the small sample class data D novel Identical and divided into training data sets DTR i And test dataset DTE i The method comprises the steps of carrying out a first treatment on the surface of the In training data set DTR i Based on the small sample target detection network, the initial parameters of the target classifier and the position regressor are theta, and the training data set DTR of the network is calculated i The target detection loss is obtained by gradient descent to obtain updated parameters, and the updated parameters are calculated in a test data set DTE i Target detection loss L on i The method comprises the steps of carrying out a first treatment on the surface of the Target detection loss L obtained on target detection task of a batch of small samples extracted randomly i Averaging to obtain average detection loss L; gradient descending updating is carried out on the gradient of the current parameter theta according to the average detection loss L, and the updated parameter theta is obtained;
in case the average detection loss L is no longer reduced or is smaller than the set threshold value, the final parameter is obtained as the optimal initial parameter theta of the target classifier and the position regressor opt ；
With optimal initial parameters theta opt Configuring a target classifier and a position regressor in small sample class data D novel And (5) retraining to obtain a trained small sample target detection network, and realizing small sample target detection by using the trained small sample target detection network.
In some embodiments, the data enhancement module is further configured to:
a small sample class input image I and a bounding box label B are acquired,
clipping a region I 'from the input image I, wherein the size of the region I' is larger than that of the boundary box label B;
taking the area outside the boundary box as a determined background, and taking pixels in the boundary box as possible foreground to obtain an initial mask for the first segmentation;
according to the initial mask, executing GrabCut algorithm once to obtain a first segmentation result mask, wherein each pixel is endowed with three labels of a possible foreground, a possible background and a determined background;
Assigning pixels marked as possible foreground as determined foreground according to the current segmentation result; for each row or column in the bounding box B, if all pixels are 'possible backgrounds', assigning the row or column pixel labels as possible foreground to obtain an initial mask for the next segmentation;
based on the initial mask of the next segmentation, executing the GrabCut algorithm again to obtain a next segmentation result;
in the final segmentation result, labeling the pixels as the determined foreground or the pixels of the possible foreground as the foreground region;
and extracting a pixel region corresponding to the foreground in the final segmentation result from the input image I, performing random scaling, and pasting the pixel region into an image containing the same type of targets as the input image I according to the random position.
In some embodiments, the data augmentation module is further configured to:
detection of model M using pre-trained base class targets base Reasoning is carried out on other non-marked images to obtain a detection result R of the basic category target class base ；
Using a small sample target detection model M novel Reasoning is carried out on other non-marked images to obtain a detection result R of the small sample class target class novel ；
For R novel Each of which is scored as a target R of s, if the targets R and R base The overlapping proportion of a target r ' with a score of s ' is larger than a set threshold value, the score of s is n times larger than that of s ', wherein n is more than or equal to 0.5, and the result of the target r is used as the label of the label-free image and is added into a small sample class data set D novel Is a kind of medium.
In some embodiments, the iterative training module is further configured to:
1) Base class data set D marked with a large amount of data base On, pretraining a basic category target detection model M base ；
2) Small sample class data set D marked with small amount of data novel Training a small sample class target detection model M based on meta-learning and fine tuning novel ；
3) Small sample class data set D novel The pixel level data of the iteration GrabCut and the random mapping are enhanced to obtain an enhanced small sample class data set D novel ；
4) Finding unlabeled small sample class targets from other data sets based on small sample class data expansion of contrast learning, and adding the unlabeled small sample class targets to a small sample class data set D novel Obtaining the expanded small sample class data set D novel ；
5) Repeating the steps 2) to 4), iteratively training a new small sample target detection model, and enhancing and expanding small sample class data until the result converges.
It should be noted that, the device in this embodiment and the method described in the foregoing belong to the same technical idea, and the same technical effects can be achieved, which are not repeated here.
Embodiments of the present invention provide a readable storage medium storing one or more programs executable by one or more processors to implement the methods described in the above embodiments.
The above description is intended to be illustrative and not restrictive. For example, the above-described examples (or one or more aspects thereof) may be used in combination with each other. For example, other embodiments may be used by those of ordinary skill in the art upon reading the above description. In addition, in the above detailed description, various features may be grouped together to streamline the invention. This is not to be interpreted as an intention that the features of the claimed invention are essential to any of the claims. Rather, inventive subject matter may lie in less than all features of a particular inventive embodiment. Thus, the following claims are hereby incorporated into the detailed description as examples or embodiments, with each claim standing on its own as a separate embodiment, and it is contemplated that these embodiments may be combined with one another in various combinations or permutations. The scope of the invention should be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.
Claims (10)
1. A method of small sample target detection, the method comprising:
target detection pre-training based on a diffusion model and target probability estimation;
merging element learning and small sample target detection of model fine adjustment;
pixel level data enhancement based on iterative GrabCut and random mapping;
expanding small sample class data based on contrast learning;
performing small sample target detection training, data enhancement and data expansion iteratively to obtain a final small sample target detection model;
and performing target detection by using a small sample target detection model, taking an image to be detected as input, and outputting a small sample class target in the image to be detected.
2. The method according to claim 1, wherein the target detection pre-training based on the diffusion model and the target probability estimation, in particular comprises:
base class data D based on a large number of labels base Pre-training a basic class target detection model M base To learn general knowledge of target detection.
3. The method according to claim 2, characterized in that the base class data D based on a multitude of labels base A basic category target detection model M is pre-trained by the following method base ：
In the model training phase, for each input image:
Extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
adding random noise on the marked boundary frame to realize the diffusion process from the true value of the boundary frame to the random distribution frame;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
and extracting the characteristics of the target candidate frame region from the characteristic map according to the position of each target candidate frame, learning the inverse process of the corresponding boundary frame true value and the noise diffusion through a position regressive device, realizing the recovery of the boundary frame true value from the random frame generated by the noise, and learning the marked target category based on the characteristics of the target candidate frame region through a target classifier.
4. A method according to claim 3, further comprising a model test phase after completion of the model training phase, in which for each input image:
extracting a feature map from an input image by using the neural network as an image encoder;
estimating the probability that each region is a target according to the feature map;
generating a random frame in a completely random manner;
sampling and generating a target candidate frame according to the size of the target probability of each region based on the random distribution frame;
And extracting the characteristics of the target candidate frame region from the characteristic map according to the position of each target candidate frame, and learning the marked target category based on the characteristics of the target candidate frame region through a target classifier.
5. The method according to claim 1, wherein the fusion learning and model fine-tuning small sample target detection specifically comprises:
construction of a Small sample target detection network M novel The small sample target detection network comprises an image encoder, a region feature extractor, a target probability estimator, a target classifier and a bitA regressor, wherein the image encoder, the regional characteristic extractor and the target probability estimator utilize a basic category target detection model M base Corresponding parameters in the step (a) are configured;
from base class data D base Target detection task T for randomly extracting a batch of small samples i { i=1, 2 … … }, target detection task T for each small sample i The number of target classes, the number of training and test images for each class are all equal to the small sample class data D novel Identical and divided into training data sets DTR i And test dataset DTE i The method comprises the steps of carrying out a first treatment on the surface of the In training data set DTR i Based on the small sample target detection network, the initial parameters of the target classifier and the position regressor are theta, and the training data set DTR of the network is calculated i The target detection loss is obtained by gradient descent to obtain updated parameters, and the updated parameters are calculated in a test data set DTE i Target detection loss L on i The method comprises the steps of carrying out a first treatment on the surface of the Target detection loss L obtained on target detection task of a batch of small samples extracted randomly i Averaging to obtain average detection loss L; gradient descending updating is carried out on the gradient of the current parameter theta according to the average detection loss L, and the updated parameter theta is obtained;
in case the average detection loss L is no longer reduced or is smaller than the set threshold value, the final parameter is obtained as the optimal initial parameter theta of the target classifier and the position regressor opt ；
With optimal initial parameters theta opt Configuring a target classifier and a position regressor in small sample class data D novel And (5) retraining to obtain a trained small sample target detection network, and realizing small sample target detection by using the trained small sample target detection network.
6. The method according to claim 1, wherein the pixel-level data enhancement based on iterative GrabCut and random mapping, in particular comprises:
a small sample class input image I and a bounding box label B are acquired,
clipping a region I 'from the input image I, wherein the size of the region I' is larger than that of the boundary box label B;
Taking the area outside the boundary box as a determined background, and taking pixels in the boundary box as possible foreground to obtain an initial mask for the first segmentation;
according to the initial mask, executing GrabCut algorithm once to obtain a first segmentation result mask, wherein each pixel is endowed with three labels of a possible foreground, a possible background and a determined background;
assigning pixels marked as possible foreground as determined foreground according to the current segmentation result; for each row or column in the bounding box B, if all pixels are possible backgrounds, assigning the row or column pixel labels as possible foreground to obtain an initial mask for the next segmentation;
based on the initial mask of the next segmentation, executing the GrabCut algorithm again to obtain a next segmentation result;
in the final segmentation result, labeling the pixels as the determined foreground or the pixels of the possible foreground as the foreground region;
and extracting a pixel region corresponding to the foreground in the final segmentation result from the input image I, performing random scaling, and pasting the pixel region into an image containing the same type of targets as the input image I according to the random position.
7. The method according to claim 1, wherein the small sample class data augmentation based on contrast learning specifically comprises:
Detection of model M using pre-trained base class targets base Reasoning is carried out on other non-marked images to obtain a detection result R of the basic category target class base ；
Using a small sample target detection model M novel Reasoning is carried out on other non-marked images to obtain a detection result R of the small sample class target class novel ；
For R novel Each of which is scored as a target R of s, if the targets R and R base The overlapping proportion of a target r 'with a score of s' is larger than a set threshold, and the score s is larger than sn times, wherein n is more than or equal to 0.5, the result of the target r is used as the label of the non-labeled image and is added to the small sample class data set D novel Is a kind of medium.
8. The method according to claim 1, wherein the iterating performs small sample target detection training, data enhancement, and data expansion to obtain a final small sample target detection model, and specifically includes:
1) Base class data set D marked with a large amount of data base On, pretraining a basic category target detection model M base ；
2) Small sample class data set D marked with small amount of data novel Training a small sample class target detection model M based on meta-learning and fine tuning novel ；
3) Small sample class data set D novel The pixel level data of the iteration GrabCut and the random mapping are enhanced to obtain an enhanced small sample class data set D novel ；
4) Finding unlabeled small sample class targets from other data sets based on small sample class data expansion of contrast learning, and adding the unlabeled small sample class targets to a small sample class data set D novel Obtaining the expanded small sample class data set D novel ；
5) Repeating the steps 2) to 4), iteratively training a new small sample target detection model, and enhancing and expanding small sample class data until the result converges.
9. A small sample target detection method apparatus, the apparatus comprising:
a first training module configured to pre-train target detection based on a diffusion model and a target probability estimate;
the second training module is configured to fuse the small sample target detection of the element learning and the model fine adjustment;
a data enhancement module configured to enhance pixel-level data based on the iterative GrabCut and the random map;
the data expansion module is configured to expand the small sample class data based on the contrast learning;
the iteration training module is configured to iterate small sample target detection training, data enhancement and data expansion to obtain a final small sample target detection model;
the target detection module is configured to detect targets by using a small sample target detection model, takes an image to be detected as input, and outputs small sample class targets in the image to be detected.
10. A readable storage medium storing one or more programs executable by one or more processors to implement the method of any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311412809.0A CN117409250B (en) | 2023-10-27 | 2023-10-27 | Small sample target detection method, device and medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311412809.0A CN117409250B (en) | 2023-10-27 | 2023-10-27 | Small sample target detection method, device and medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117409250A true CN117409250A (en) | 2024-01-16 |
| CN117409250B CN117409250B (en) | 2024-04-30 |
Family
ID=89486760
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311412809.0A Active CN117409250B (en) | 2023-10-27 | 2023-10-27 | Small sample target detection method, device and medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117409250B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112861720A (en) * | 2021-02-08 | 2021-05-28 | 西北工业大学 | Remote sensing image small sample target detection method based on prototype convolutional neural network |
| US20210319266A1 (en) * | 2020-04-13 | 2021-10-14 | Google Llc | Systems and methods for contrastive learning of visual representations |
| CN114078197A (en) * | 2021-11-05 | 2022-02-22 | 清华大学深圳国际研究生院 | Small sample target detection method and device based on support sample characteristic enhancement |
| CN114841257A (en) * | 2022-04-21 | 2022-08-02 | 北京交通大学 | Small sample target detection method based on self-supervision contrast constraint |
| CN115661542A (en) * | 2022-11-08 | 2023-01-31 | 中国科学院计算技术研究所 | Small sample target detection method based on feature relation migration |
| CN116071660A (en) * | 2023-03-10 | 2023-05-05 | 广西师范大学 | Target detection method based on small sample |
-
2023
- 2023-10-27 CN CN202311412809.0A patent/CN117409250B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210319266A1 (en) * | 2020-04-13 | 2021-10-14 | Google Llc | Systems and methods for contrastive learning of visual representations |
| CN112861720A (en) * | 2021-02-08 | 2021-05-28 | 西北工业大学 | Remote sensing image small sample target detection method based on prototype convolutional neural network |
| CN114078197A (en) * | 2021-11-05 | 2022-02-22 | 清华大学深圳国际研究生院 | Small sample target detection method and device based on support sample characteristic enhancement |
| CN114841257A (en) * | 2022-04-21 | 2022-08-02 | 北京交通大学 | Small sample target detection method based on self-supervision contrast constraint |
| CN115661542A (en) * | 2022-11-08 | 2023-01-31 | 中国科学院计算技术研究所 | Small sample target detection method based on feature relation migration |
| CN116071660A (en) * | 2023-03-10 | 2023-05-05 | 广西师范大学 | Target detection method based on small sample |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117409250B (en) | 2024-04-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107368845B (en) | Optimized candidate region-based Faster R-CNN target detection method | |
| CN109299274B (en) | Natural scene text detection method based on full convolution neural network | |
| CN108681752B (en) | Image scene labeling method based on deep learning | |
| CN111640125B (en) | Aerial photography graph building detection and segmentation method and device based on Mask R-CNN | |
| CN110766008A (en) | Text detection method facing any direction and shape | |
| CN113673338A (en) | Natural scene text image character pixel weak supervision automatic labeling method, system and medium | |
| CN112541491B (en) | End-to-end text detection and recognition method based on image character region perception | |
| CN111310609B (en) | Video target detection method based on time sequence information and local feature similarity | |
| EP3973507B1 (en) | Segmentation for holographic images | |
| CN114820655B (en) | Weak supervision building segmentation method taking reliable area as attention mechanism supervision | |
| CN112819840B (en) | High-precision image instance segmentation method integrating deep learning and traditional processing | |
| CN112733858B (en) | Image character rapid identification method and device based on character region detection | |
| CN113793357A (en) | Bronchopulmonary segment image segmentation method and system based on deep learning | |
| CN113392930A (en) | Traffic sign target detection method based on multi-level divide-and-conquer network | |
| CN117409250B (en) | Small sample target detection method, device and medium | |
| CN116912144A (en) | Data enhancement method based on discipline algorithm and channel attention mechanism | |
| CN111612803A (en) | Vehicle image semantic segmentation method based on image definition | |
| CN115578721A (en) | Streetscape text real-time detection method based on attention feature fusion | |
| KR102026280B1 (en) | Method and system for scene text detection using deep learning | |
| US11908133B2 (en) | Method for cell image segmentation using scribble labels, recording medium and device for performing the method | |
| CN116310293A (en) | Method for detecting target of generating high-quality candidate frame based on weak supervised learning | |
| CN114708591A (en) | Document image Chinese character detection method based on single character connection | |
| CN114998132A (en) | Weak supervision shadow detection method for mining complementary features through double networks | |
| CN108428240B (en) | Salient object segmentation method adaptive to input information | |
| CN113420760A (en) | Handwritten Mongolian detection and identification method based on segmentation and deformation LSTM |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |