CN110929794B - Side-scan sonar image classification method based on multi-task learning - Google Patents
Side-scan sonar image classification method based on multi-task learning Download PDFInfo
- Publication number
- CN110929794B CN110929794B CN201911188526.6A CN201911188526A CN110929794B CN 110929794 B CN110929794 B CN 110929794B CN 201911188526 A CN201911188526 A CN 201911188526A CN 110929794 B CN110929794 B CN 110929794B
- Authority
- CN
- China
- Prior art keywords
- scan sonar
- task
- neural network
- convolutional neural
- sonar image
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of side-scan sonar image identification, and particularly relates to a side-scan sonar image classification method based on multi-task learning. The invention combines the multi-task learning idea and the convolutional neural network method, uses the convolutional neural network to automatically extract the features, can extract important features which cannot be sensed by human eyes compared with the traditional manually-arranged feature extractor, and can also reduce the influence of factors such as side scan sonar image noise, image edge loss, image feature deformation and the like on the feature extraction. According to the method, the idea of multi-task learning is utilized, and an optical image classification task is introduced, so that the feature space of the side scan sonar image can be enriched, and the problem of overfitting caused by incomplete feature extraction when the number of samples is too small is solved; the method can solve the problems of few side-scan sonar image samples and poor classification effect when the feature extraction is difficult, and has certain engineering and research values.
Description
Technical Field
The invention belongs to the technical field of side-scan sonar image identification, and particularly relates to a side-scan sonar image classification method based on multi-task learning.
Background
Compared with an optical imaging system, the side-scan sonar system can overcome the limitation of severe conditions in water and acquire images of the sea bottom in a long distance. The sonar image has the application which is difficult to be compared with other tools in the fields of submarine surveying and mapping, submarine geological investigation, deep water battle, submarine rescue and the like. However, due to the vast sea, it is very difficult to acquire enough side scan sonar images containing the target to be searched. On the other hand, because the quality of the side-scan sonar image may be affected by illumination, temperature, sand, stone and the like, and the target characteristics in the side-scan sonar image are far from the real target characteristics due to the factors, a good target classification effect is difficult to obtain only by using the limited side-scan sonar image.
At present, the classification of side-scan sonar images is mainly based on a traditional machine learning method, a feature extractor is designed firstly, and then a proper classifier is selected for classification. The most important of them is feature extraction, which directly determines the upper bound of classification accuracy. However, the imaging principle of the side scan sonar causes the problems of more noise points, obvious shadow and incomplete edge of the spliced side scan sonar images. This results in that the conventional point and line feature extraction methods such as Frstner, harris and Hough Transform cannot extract well the distinguishing features. The popular deep learning method based on the convolutional neural network achieves good effects in the aspects of image recognition and classification, even exceeds the level of human beings, however, the performances are based on strong computing power, a large number of samples and efficient algorithms. For side scan sonar images, due to the vast sea, scenes containing objects (such as airplanes, ships, etc.) are not many, and it is very difficult to scan the objects, which directly results in very few samples of the side scan sonar images. In this case, the effect achieved by the deep learning may not even be comparable to the conventional method that has been elaborated.
In summary, side scan sonars are widely used in ocean exploration, but due to the reasons of small sample size and poor quality of side scan sonar image targets, the current target identification method does not achieve good effect in side scan sonar image classification.
Disclosure of Invention
The invention aims to provide a side-scan sonar image classification method based on multi-task learning, which is used for finishing the training and classification of a side-scan sonar image classification network under the condition of small number of side-scan sonar image samples.
The purpose of the invention is realized by the following technical scheme: the method comprises the following steps:
step 1: constructing a classified side-scan sonar image data set according to the existing side-scan sonar data;
step 2: constructing an optical image auxiliary data set related to the target according to the target category in the side-scan sonar data set;
and 3, step 3: constructing a multi-task learning model based on a convolutional neural network; the multi-task learning model based on the convolutional neural network has two inputs which share a convolutional pooling layer; each input has a respective classification layer and calculates a loss function of the input independently; the total loss function of the multi-task learning model based on the convolutional neural network is the sum of the weighted loss functions of two inputs;
and 4, step 4: taking data in the side-scan sonar image data set and the optical image auxiliary data set as input of a multi-task learning model based on a convolutional neural network, setting network parameters and training the multi-task learning model based on the convolutional neural network to obtain an automatic classification network of the side-scan sonar image target;
and 5: verifying the automatic classification network performance of the side-scan sonar image target by using a verification set, and recording the highest accuracy; if the highest accuracy rate does not meet the requirement, adjusting the weight ratio of the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network, and returning to the step 4 for retraining;
step 6: and inputting the image detected by the side-scan sonar in real time into an automatic classification network of the side-scan sonar image target to obtain a classification result.
The present invention may further comprise:
in the step 3, a multitask learning model based on a convolutional neural network is constructed, the convolutional neural network is used for feature extraction, the convolutional neural network is constructed by imitating a biological visual mechanism, and the convolutional kernel sharing of the middle layer and the sparsity of interlayer connection enable the convolutional neural network to automatically and stably learn the image features with smaller calculated amount and have no additional feature engineering;
the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network in the step 3 are specifically as follows:
the penalty function for the optical image auxiliary dataset classification task is:
wherein M is the number of auxiliary task samples in the optical image auxiliary data set participating in training; f (x) is a fitting function of the convolution and pooling layers; g is a radical of formula 1 (x) Fitting function of full connection layer for auxiliary task; x is the number of s (i) An ith sample participating in training in the optical image auxiliary data set; y is s (i) A label corresponding to the ith sample in the optical image auxiliary data set;
the loss function for the side scan sonar image dataset classification task is:
wherein N is the number of target task samples in the side scan sonar image data set participating in training; f (x) is a fitting function of the convolution and pooling layers; g 2 (x) A fitting function of a fully connected layer for the target task; x is a radical of a fluorine atom t (i) The ith sample participating in training is collected for the side-scan sonar image data; y is t (i) A label corresponding to the ith sample in the side scan sonar image data set;
the final set total loss function of the convolutional neural network-based multitask learning model is:
J=α*J s +β*J t
wherein, α and β represent the weights of the auxiliary task and the target task respectively, and are initialized to 1:1.
The invention has the beneficial effects that:
the invention combines the multi-task learning idea and the convolutional neural network method, uses the convolutional neural network to automatically extract the features, can extract important features which cannot be sensed by human eyes compared with the traditional manually-arranged feature extractor, and can also reduce the influence of factors such as side scan sonar image noise, image edge loss, image feature deformation and the like on the feature extraction. According to the method, the idea of multi-task learning is utilized, and an optical image classification task is introduced, so that the feature space of the side scan sonar image can be enriched, and the problem of overfitting caused by incomplete feature extraction when the number of samples is too small is solved; the method can solve the problems of few side-scan sonar image samples and poor classification effect when the feature extraction is difficult, and has certain engineering and research values.
Drawings
FIG. 1 is a general flow diagram of the present invention.
Fig. 2 is an image of a side-scan sonar in the side-scan sonar image data set.
Fig. 3 is an optical image in an optical image auxiliary data set.
Fig. 4 is a diagram of a network architecture of the present invention.
FIG. 5 (a) is a graph of the loss function of the network training process at 1:1 weight ratio.
FIG. 5 (b) is a graph of the accuracy of the network training process under the 1:1 weight ratio.
FIG. 6 is a comparison table of prediction accuracy under different weighting ratios according to the present invention.
FIG. 7 is a table comparing the side-scan sonar image prediction results of the present invention with other methods.
Detailed Description
The invention is further described below with reference to the accompanying drawings.
The invention belongs to the field of side scan sonar image identification, and particularly relates to a side scan sonar image classification method based on multi-task learning (MTL-CNN); aiming at the limitation of the current side-scan sonar image classification method, the invention aims to provide a side-scan sonar image classification method based on multi-task learning, and the invention aims to finish the training and classification of a side-scan sonar image classification network end to end by using a deep learning method under the condition that the number of side-scan sonar image samples is small.
The invention finds out a side-scan sonar image containing a target from the existing side-scan sonar images; finding out an optical image which is labeled and has the same target as that in the side scan sonar image from the existing ImageNet data set; and then a multi-task learning model based on a convolutional neural network is built, the network is trained by combining an optical data set and a side-scan sonar data set, and the classification capability of the deep learning network on a side-scan sonar target can be improved by training the network by combining the optical data set. Considering that the features of the side-scan sonar images are difficult to extract by using a traditional method, the conventional deep learning target identification and classification method can extract features with self-adaptability, but needs a large number of samples, and the real samples of the side-scan sonar images are high in acquisition cost and difficulty and cannot meet the requirements of the deep learning method. Therefore, the side-scan sonar image classification method based on multi-task learning can solve the problem that side-scan sonar image samples are few, and therefore side-scan sonar image targets can be well classified.
A side scan sonar image classification method based on multitask learning comprises the following steps:
step 1: constructing a classified side-scan sonar image data set according to the existing side-scan sonar data;
step 2: constructing an optical image auxiliary data set related to the target according to the target category in the side scan sonar data set;
and step 3: constructing a multi-task learning model based on a convolutional neural network; the multi-task learning model based on the convolutional neural network has two inputs which share a convolutional pooling layer; each input has a respective classification layer and calculates its own loss function independently; the total loss function of the multi-task learning model based on the convolutional neural network is the sum of the weighted loss functions of two inputs;
and 4, step 4: taking data in the side-scan sonar image data set and the optical image auxiliary data set as input of a multi-task learning model based on a convolutional neural network, setting network parameters and training the multi-task learning model based on the convolutional neural network to obtain an automatic classification network of the side-scan sonar image target;
and 5: verifying the automatic classification network performance of the side-scan sonar image target by using a verification set, and recording the highest accuracy; if the highest accuracy rate does not meet the requirement, adjusting the weight ratio of the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network, and returning to the step 4 for retraining;
step 6: and inputting the image detected by the side-scan sonar in real time into an automatic classification network of the side-scan sonar image target to obtain a classification result.
Constructing a multitask learning model based on a convolutional neural network, extracting features by using the convolutional neural network, constructing a visual mechanism of a simulated organism by using the convolutional neural network, wherein the convolutional kernel of the middle layer shares sparsity connected with the layers, so that the convolutional neural network can automatically and stably learn the image features with small calculated amount and has no additional feature engineering;
the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network in the step 3 are specifically as follows:
the penalty function for the optical image auxiliary dataset classification task is:
wherein M is the number of auxiliary task samples in the optical image auxiliary data set participating in training; f (x) is a fitting function of the convolved, pooled layers; g 1 (x) A fitting function of a fully connected layer for the auxiliary task; x is the number of s (i) The ith sample participating in training in the optical image auxiliary data set; y is s (i) A label corresponding to the ith sample in the optical image auxiliary data set;
the loss function for the side scan sonar image dataset classification task is:
wherein N is the number of target task samples in the side scan sonar image data set participating in training; f (x) being a convolution, pooling layerFitting a function; g 2 (x) A fitting function of a fully connected layer for the target task; x is a radical of a fluorine atom t (i) The ith sample participating in training is collected for the side scan sonar image data; y is t (i) A label corresponding to the ith sample in the side scan sonar image data set;
the final set total loss function of the convolutional neural network-based multi-task learning model is:
J=α*J s +β*J t
wherein, α and β represent the weights of the auxiliary task and the target task respectively, and are initialized to 1:1.
Example 1:
a side scan sonar image classification method based on multitask learning specifically comprises the following steps:
step 1: constructing a side scan sonar image data set, and classifying the existing side scan sonar data set, wherein three types of airplanes, ships and other objects except the airplanes and the ships are taken as examples for explanation;
step 2: acquiring images of the airplane and the ship under an optical scene similar to the images of the airplane and the ship in the side scan sonar image data set from the existing data set;
and step 3: and constructing a multi-task learning model based on the convolutional neural network, wherein the model is modified on the basis of the VGG11 convolutional neural network model. The model has two inputs and shares the convolution pooling layer. Each input has its own classification level (fully connected level) and its own loss function is computed separately. The total loss function of the multitask network model is the sum of the weighted loss functions of the two inputs, and the weight ratio is initialized (generally 1:1);
and 4, step 4: using the two data sets in the steps 1 and 2 as the input of the network model in the step 3, then training the network model, and recording the highest accuracy of the model on the side scan sonar image verification set;
and 5: if the accuracy is at the maximum (or close to convergence), it is available, and then the process is finished; otherwise, the weight ratio is adjusted, and the step 4 is returned to.
And 2, acquiring the images of the airplane and the ship under the optical scene similar to the images of the airplane and the ship in the side scan sonar image data set from the existing data set, wherein the data set can assist the side scan sonar data set to train and solve the problem of insufficient data of the side scan sonar data set.
And 3, building a multi-task learning model based on the convolutional neural network, and extracting features by using the convolutional neural network. The convolutional neural network is constructed by imitating the visual mechanism of a living being, and the convolution kernel sharing of the middle layer and the sparsity of connection between the layers enable the convolutional neural network to automatically and stably learn the image characteristics with small calculation amount and no additional characteristic engineering.
And 3, building a multi-task learning model based on the convolutional neural network. The model combines the multi-task learning idea and the convolutional neural network in the claim 3, is modified on the basis of the VGG11, and is added with an input and an output. The two inputs to the model are a side-scan sonar image and an optical image, respectively, which share a convolutional pooling layer and have separate classification layers. By sharing the convolution pooling layer, the bottom-layer characteristics of the optical image with sufficient data volume can be fully utilized by the side-scan sonar image classification task, and the problem of insufficient side-scan sonar image characteristic extraction is solved; the total loss function of the multi-task network is the sum of two input weighting loss functions, and by simultaneously optimizing the loss functions of optical image and side-scan sonar image classification, the optical image classification task can extract more features, and the features can be better applied to the classification task of the side-scan sonar image.
In the step 5, the optimal proportion is found by adjusting the weight ratio of the loss functions of the auxiliary task (optical image target classification network training) and the target task (side scan sonar image target classification network training), so that the classification effect of the whole network is optimal, the classification of the side scan sonar images is facilitated, and the network generalization performance is stronger.
Compared with the prior art, the invention has the advantages that: combines the multi-task learning idea and the convolution neural network method. The convolutional neural network is used for automatic feature extraction, and compared with a traditional manually-arranged feature extractor, the convolutional neural network can extract important features which cannot be sensed by human eyes, and can also reduce the influence of factors such as side-scan sonar image noise, image edge loss, image feature deformation and the like on feature extraction; by utilizing the idea of multi-task learning and introducing an optical image classification task, the feature space of the side scan sonar image can be enriched, and the problem of overfitting caused by incomplete feature extraction when the number of samples is too small is avoided; the method can solve the problems of few side-scan sonar image samples and poor classification effect when the feature extraction is difficult, and has certain engineering and research values.
The invention will be further described by way of practical examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.
As shown in fig. 1, is a flow chart of the method of the present invention, which specifically comprises the following steps:
step 1: constructing a side scan sonar image data set, and dividing the side scan sonar image data set into three types, namely an airplane, a ship and other targets, as shown in an attached figure 2;
step 2: constructing an optical image data set which is divided into two types, namely an airplane and a ship, as shown in figure 3;
and step 3: a convolutional neural network based multi-task learning (MTL-CNN) model was constructed as shown in fig. 4. The model has two inputs, a side-scan sonar image and an optical image. The two inputs share the convolutional pooling layer of the convolutional neural network and have independent fully-connected layers. The penalty function for the ImageNet dataset (helper dataset) classification task is:
where M represents the number of helper samples involved in the training, f (x) represents the fitting function of the convolved, pooled layers, g 1 (x) Fitting function, x, of the fully connected layer representing the auxiliary task s (i) Representing the i-th sample involved in the training, y s (i) Representing the label corresponding to the ith sample.
The penalty function for the side scan sonar dataset (target dataset) classification task is:
where N represents the number of target task samples involved in the training, f (x) represents the fitting function of the convolution, pooling layers, g 2 (x) Fitting function, x, of the fully connected layer representing the auxiliary task t (i) Representing the i-th sample involved in the training, y t (i) Representing the label corresponding to the ith sample.
The optimized loss function for the final setting is:
J=α*J s +β*J t
wherein, α and β represent the weights of the auxiliary task and the target task respectively, and are generally initialized to 1:1;
and 4, step 4: both data sets in steps 1, 2 are used as input to the network model in step 3. The number of the aerial images in the side scan sonar image data set is 120, the number of the steamship is 160, and the number of the steamship is 250; the optical image data set comprises 1200 images of an airplane and a ship. Initializing alpha: beta is 1:1, other parameters use default values, then the network model is trained, and the highest accuracy of the model on the side scan sonar image test set is recorded. 5 (a) and 5 (b) are loss function and accuracy chart of the network training process under the weight ratio of 1:1, respectively;
and 5: if the accuracy is at the maximum (or close to convergence), ending; otherwise, the weight ratio is adjusted, and the step 4 is returned to. Through several experiments, the optimal weight ratio of the auxiliary task to the target task is determined to be 1. Fig. 6 shows the accuracy (accuracy), precision (Precision), and Recall (Recall) of the network trained under different specific gravities on the side scan sonar image. In addition, compared with other methods such as a common Convolutional Neural Network (CNN) method and a Fine-tuning (Fine-tuning) method, the effectiveness of the convolutional neural network-based multitask learning method is proved, as shown in fig. 7.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (2)
1. A side scan sonar image classification method based on multitask learning is characterized by comprising the following steps:
step 1: constructing a classified side-scan sonar image data set according to the existing side-scan sonar data;
and 2, step: constructing an optical image auxiliary data set related to the target according to the target category in the side-scan sonar data set;
and step 3: constructing a multi-task learning model based on a convolutional neural network; the multi-task learning model based on the convolutional neural network has two inputs which share a convolutional pooling layer; each input has a respective classification layer and calculates its own loss function independently; the total loss function of the multi-task learning model based on the convolutional neural network is the sum of the weighted loss functions of two inputs;
the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network are specifically as follows:
the penalty function for the optical image auxiliary dataset classification task is:
wherein M is the number of auxiliary task samples in the optical image auxiliary data set participating in training; f (x) is a fitting function of the convolution and pooling layers; g is a radical of formula 1 (x) A fitting function of a fully connected layer for the auxiliary task; x is the number of s (i) An ith sample participating in training in the optical image auxiliary data set; y is s (i) A label corresponding to the ith sample in the optical image auxiliary data set;
the loss function for the side scan sonar image dataset classification task is:
wherein N is the number of target task samples in the side scan sonar image data set participating in training; f (x) is a fitting function of the convolution and pooling layers; g 2 (x) A fitting function of a fully connected layer for the target task; x is the number of t (i) The ith sample participating in training is collected for the side scan sonar image data; y is t (i) A label corresponding to the ith sample in the side scan sonar image data set;
the final set total loss function of the convolutional neural network-based multi-task learning model is:
J=α*J s +β*J t
wherein, alpha and beta respectively represent the weights of the auxiliary task and the target task and are initialized to 1:1;
and 4, step 4: taking data in the side-scan sonar image data set and the optical image auxiliary data set as input of a multi-task learning model based on a convolutional neural network, setting network parameters and training the multi-task learning model based on the convolutional neural network to obtain an automatic classification network of the side-scan sonar image target;
and 5: verifying the automatic classification network performance of the side-scan sonar image target by using a verification set, and recording the highest accuracy; if the highest accuracy rate does not meet the requirement, adjusting the weight ratio of the loss functions of the two inputs of the multi-task learning model based on the convolutional neural network, and returning to the step 4 for retraining;
step 6: and inputting the image detected by the side-scan sonar in real time into an automatic classification network of the side-scan sonar image target to obtain a classification result.
2. The side-scan sonar image classification method based on multitask learning according to claim 1, characterized in that:
and 3, constructing a multitask learning model based on a convolutional neural network, extracting features by using the convolutional neural network, constructing the visual mechanism of a bionic by using the convolutional neural network, and enabling the convolutional neural network to automatically and stably learn the image features with a small calculated amount without additional feature engineering due to the sharing of convolutional kernels of the middle layer and the sparsity of connection between the layers.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911188526.6A CN110929794B (en) | 2019-11-28 | 2019-11-28 | Side-scan sonar image classification method based on multi-task learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911188526.6A CN110929794B (en) | 2019-11-28 | 2019-11-28 | Side-scan sonar image classification method based on multi-task learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110929794A CN110929794A (en) | 2020-03-27 |
CN110929794B true CN110929794B (en) | 2022-12-13 |
Family
ID=69847513
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911188526.6A Active CN110929794B (en) | 2019-11-28 | 2019-11-28 | Side-scan sonar image classification method based on multi-task learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110929794B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106897662A (en) * | 2017-01-06 | 2017-06-27 | 北京交通大学 | The localization method of the face key feature points based on multi-task learning |
CN107679250A (en) * | 2017-11-01 | 2018-02-09 | 浙江工业大学 | A kind of multitask layered image search method based on depth own coding convolutional neural networks |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10360494B2 (en) * | 2016-11-30 | 2019-07-23 | Altumview Systems Inc. | Convolutional neural network (CNN) system based on resolution-limited small-scale CNN modules |
CN107239736A (en) * | 2017-04-28 | 2017-10-10 | 北京智慧眼科技股份有限公司 | Method for detecting human face and detection means based on multitask concatenated convolutional neutral net |
CN111373419A (en) * | 2017-10-26 | 2020-07-03 | 奇跃公司 | Gradient normalization system and method for adaptive loss balancing in deep multitask networks |
IL274424B2 (en) * | 2017-11-14 | 2024-07-01 | Magic Leap Inc | Meta-learning for multi-task learning for neural networks |
CN108510000B (en) * | 2018-03-30 | 2021-06-15 | 北京工商大学 | Method for detecting and identifying fine-grained attribute of pedestrian in complex scene |
CN109325527A (en) * | 2018-09-05 | 2019-02-12 | 哈尔滨工程大学 | A kind of sidescan-sonar image target automatic classification method based on transfer learning and deep learning |
US10426442B1 (en) * | 2019-06-14 | 2019-10-01 | Cycle Clarity, LLC | Adaptive image processing in assisted reproductive imaging modalities |
-
2019
- 2019-11-28 CN CN201911188526.6A patent/CN110929794B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106897662A (en) * | 2017-01-06 | 2017-06-27 | 北京交通大学 | The localization method of the face key feature points based on multi-task learning |
CN107679250A (en) * | 2017-11-01 | 2018-02-09 | 浙江工业大学 | A kind of multitask layered image search method based on depth own coding convolutional neural networks |
Also Published As
Publication number | Publication date |
---|---|
CN110929794A (en) | 2020-03-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021000906A1 (en) | Sar image-oriented small-sample semantic feature enhancement method and apparatus | |
CN108182441B (en) | Parallel multichannel convolutional neural network, construction method and image feature extraction method | |
CN109308483B (en) | Dual-source image feature extraction and fusion identification method based on convolutional neural network | |
CN107909082B (en) | Sonar image target identification method based on deep learning technology | |
CN110414554B (en) | Stacking ensemble learning fish identification method based on multi-model improvement | |
CN109344855B (en) | Depth model face beauty evaluation method based on sequencing guided regression | |
CN113408492B (en) | Pedestrian re-identification method based on global-local feature dynamic alignment | |
CN111046967A (en) | Underwater image classification method based on convolutional neural network and attention mechanism | |
CN110059741A (en) | Image-recognizing method based on semantic capsule converged network | |
CN108921019A (en) | A kind of gait recognition method based on GEI and TripletLoss-DenseNet | |
CN106529570B (en) | Image classification method based on depth ridge ripple neural network | |
CN110287770B (en) | Water individual target matching identification method based on convolutional neural network | |
CN109948467A (en) | Method, apparatus, computer equipment and the storage medium of recognition of face | |
Rodrigues et al. | Automatic fish species classification based on robust feature extraction techniques and artificial immune systems | |
CN110991516A (en) | Side-scan sonar image target classification method based on style migration | |
CN110880010A (en) | Visual SLAM closed loop detection algorithm based on convolutional neural network | |
CN113111857A (en) | Human body posture estimation method based on multi-mode information fusion | |
CN110633727A (en) | Deep neural network ship target fine-grained identification method based on selective search | |
CN111723823B (en) | Underwater target detection method based on third party transfer learning | |
CN111694977A (en) | Vehicle image retrieval method based on data enhancement | |
CN110096976A (en) | Human behavior micro-Doppler classification method based on sparse migration network | |
CN110991554B (en) | Improved PCA (principal component analysis) -based deep network image classification method | |
CN109948662B (en) | Face image depth clustering method based on K-means and MMD | |
CN114495170A (en) | Pedestrian re-identification method and system based on local self-attention inhibition | |
CN110598737A (en) | Online learning method, device, equipment and medium of deep learning model |
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 |