CN114241285A - Ship rapid detection method based on knowledge distillation and semi-supervised learning - Google Patents
Ship rapid detection method based on knowledge distillation and semi-supervised learning Download PDFInfo
- Publication number
- CN114241285A CN114241285A CN202111410408.2A CN202111410408A CN114241285A CN 114241285 A CN114241285 A CN 114241285A CN 202111410408 A CN202111410408 A CN 202111410408A CN 114241285 A CN114241285 A CN 114241285A
- Authority
- CN
- China
- Prior art keywords
- ship
- sample
- network
- data
- representing
- 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 43
- 238000013140 knowledge distillation Methods 0.000 title claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 33
- 230000006870 function Effects 0.000 claims description 40
- 238000000034 method Methods 0.000 claims description 23
- 238000012544 monitoring process Methods 0.000 claims description 10
- 230000002776 aggregation Effects 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 3
- 230000002708 enhancing effect Effects 0.000 abstract description 2
- 238000010276 construction Methods 0.000 description 6
- 238000002372 labelling Methods 0.000 description 6
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a quick ship detection method based on knowledge distillation and semi-supervised learning, which comprises the following steps of: 1) constructing a ship data set by using the existing ship data and ship data acquired by a sea area to be deployed, marking a ship target on partial data in the data set, generating a training sample according to the prior frame sliding, and enhancing the data; 2) constructing a teacher network, pre-training by using a labeled sample, and then performing semi-supervised training by combining a label-free sample to realize ship detection; 3) constructing a lightweight student network, and guiding the student network to realize knowledge distillation on the marked samples and the unmarked samples by using the trained teacher network; 4) and converting the trained student network model into an ONNX format, and further optimizing and deploying the student network by adopting OpenVINO to realize rapid ship detection. The invention transfers the knowledge learned by the teacher network with high precision and large capacity to the lightweight student network, thereby realizing rapid ship detection while ensuring the detection precision.
Description
Technical Field
The invention relates to the technical field of ship detection, in particular to a ship rapid detection method based on knowledge distillation and semi-supervised learning.
Background
At present, the offshore economy is developed vigorously, and the real-time monitoring on the offshore environment is beneficial to timely finding out events such as blockage and smuggling, and the management level is improved. However, a large amount of monitoring image data is difficult to be processed effectively in a manual mode, so that the realization of rapid ship detection can early warn unknown targets in real time, improve the information classification of ships and construct an informationized ship management system, and the ship management system has important theoretical significance and application value.
The purpose of the vessel detection is to obtain the position and class of the vessel in the input image. At present, the ship detection method mainly comprises two main types: the method comprises a ship detection method based on background modeling and a ship detection method based on deep learning. The ship detection method based on background modeling mainly comprises the steps of modeling a background, and then quickly extracting a foreground target by utilizing background subtraction. However, such methods are prone to misidentifying a stationary vessel as background and can produce a large number of false positive targets in dynamically changing marine scenes. Furthermore, such methods lack further classification of the detected objects. The ship detection method based on deep learning mainly utilizes a neural network to extract target features, and performs fine-grained classification and position optimization on targets. However, such methods are prone to risk overfitting in cases where the data set is small in size due to the large number of parameters. And a large number of parameters enable the detection model to need to depend on special equipment such as a display card and the like to carry out accelerated operation, so that real-time ship detection is difficult to realize in general equipment.
Disclosure of Invention
The invention aims to overcome the defects and shortcomings of the prior art and provides a ship rapid detection method based on knowledge distillation and semi-supervised learning.
In order to achieve the purpose, the technical scheme provided by the invention is as follows: a ship rapid detection method based on knowledge distillation and semi-supervised learning comprises the following steps:
1) constructing a ship data set by using existing ship data and ship data acquired by a deployment sea area, marking a ship target on partial data in the data set, generating a training sample according to a priori frame sliding, and performing data enhancement, wherein the training sample comprises a marked sample and a non-marked sample;
2) constructing a teacher network, pre-training by using a labeled sample, and then performing semi-supervised training by combining a label-free sample to realize ship detection;
3) constructing a lightweight student network, and guiding the student network to realize knowledge distillation on the marked samples and the unmarked samples by using the trained teacher network;
4) and converting the trained student network into an ONNX format, further optimizing and deploying the student network by adopting OpenVINO, and finally realizing rapid ship detection through the optimized student network.
Further, the step 1) comprises the following steps:
1.1) acquiring existing ship data and acquiring data in a sea area to be deployed, constructing a ship data set, marking a ship target on partial data to form marked data and unmarked data, and calculating the ship target in the marked data through clustering to acquire a multi-scale prior frame;
1.2) generating a multi-scale ship sample as a training sample according to the sliding of the multi-scale prior frame, wherein the specific operation is as follows: generating an annotated sample in the annotated data in a sliding manner, dividing the annotated sample into a positive sample and a negative sample according to the intersection and comparison of the sample and the annotated frame, wherein the positive sample comprises a plurality of classes of ship targets and records a class label and a position regression label of the ship targets, the negative sample is a background sample, and generating a non-annotated sample in the non-annotated data in a sliding manner;
1.3) carrying out data enhancement on the training sample, including horizontal turning, color dithering, light variation and random clipping.
Further, in step 2), a teacher network is constructed and trained, and the method comprises the following steps:
2.1) building a teacher network with large model capacity as a ship discriminator, and pre-training the network by using a labeled sample;
In the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetatIn order to provide trainable parameters for the teacher's network,in order to supervise the loss,as a consistency regularization term, λ1A penalty factor is indicated.
Monitoring lossThe method comprises the steps of category loss and regression loss, wherein the category loss is calculated by adopting a cross entropy function, and the regression loss is calculated by adopting a smooth-L1 function; the consistency regularization item aims to enable different data enhancement samples of the same sample to have class consistency and improve the intra-class aggregation degree of target features, and the mathematical form of the consistency regularization item is as follows:
wherein N represents the number of data enhancements, fcRepresenting a cross entropy function, ftSample prediction function representing teacher network, d(0)Representing the original sample, d(i)The ith data representing the original sample enhances the sample.
Further, the step 3) comprises the following steps:
3.1) building a student network with a small number of parameters as a lightweight ship discriminator;
In the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetasIn order to train the trainable parameters of the student network,in order to supervise the loss,for the knowledge distillation regularization term, lambda2Representing a penalty coefficient;
monitoring lossThe method comprises the steps of category loss and regression loss, wherein the category loss is calculated by adopting a cross entropy function, and the regression loss is calculated by adopting a smooth-L1 function; the purpose of the knowledge distillation regular item is to enable the knowledge learned by the teacher network to be migrated into the student network, so that the student network can have consistent prediction with the teacher network, and the mathematical form is as follows:
wherein f iscRepresenting a cross entropy function, flRepresenting smooth-L1 function, fsSample prediction function representing a student network, ftSample prediction function representing teacher network, fTRepresents a sharpening function, T represents a sharpening coefficient, and the mathematical form of the sharpening function is shown as follows:
where p represents a one-dimensional input vector,representing the j-th component of pTo the power of one.
Further, the step 4) comprises the following steps:
4.1) converting the trained student network into an ONNX format, and further optimizing and deploying the student network by utilizing OpenVINO;
4.2) cutting the input image into 4 non-overlapping areas, and circularly monitoring the 4 areas through an optimized student network to further improve the speed of ship detection.
Compared with the prior art, the invention has the following advantages and beneficial effects:
1. the technical scheme combining knowledge distillation and semi-supervised learning is adopted, the knowledge learned by a teacher network is transferred to a lightweight student network through the consistency regular term and the knowledge distillation regular term, the technical problems of high labeling cost and large model parameter quantity in a fully supervised learning method are solved, and the technical effect of improving the detection speed while ensuring the detection precision is achieved.
2. According to the method, the candidate region is generated by using the multi-scale prior frame, so that the ship detection problem is converted into the classification and regression problem, the calculation complexity is reduced, and the actual scene is deployed by adopting a lightweight network model.
3. The invention converts the trained network model into ONNX format, can repeatedly call the network model in different frames, adopts OpenVINO to deploy and reason the network model, and is beneficial to further accelerating ship detection on a CPU.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a schematic diagram of a positive sample generated.
FIG. 3 is a graph of the classification accuracy of a teacher network and a student network.
FIG. 4 is a diagram of a result of a ship test performed by the method of the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.
The experimental platform of this example is python3.6, pytorch1.7.1, computer configuration: the CPU model is Intel (R) core (TM) i9-10900X, the memory is 32GB, and the graphics card model is NVIDIA Quadro M6000.
As shown in fig. 1, the method for quickly detecting a ship based on knowledge distillation and semi-supervised learning provided by this embodiment is roughly divided into four stages: the first stage is the construction and pretreatment of a data set, mainly the construction of a ship data set, the calculation of a ship prior frame, the generation of a training sample and the data enhancement; the second stage is the construction and training of a teacher network, which is mainly to construct the teacher network with large model capacity for pre-training and semi-supervised training; the third stage is the construction and training of the student network, mainly the construction of the lightweight student network, and the knowledge of the teacher network is transferred to the student network by knowledge distillation; the fourth stage is the optimization and deployment of the model, and mainly comprises the steps of converting the trained student network into an ONNX format and utilizing OpenVINO to deploy. The method comprises the following specific steps:
1) the method comprises the following steps of constructing a ship data set by utilizing existing ship data and ship data acquired by a deployment sea area, marking a ship target on partial data in the data set, generating training samples (including marked samples and unmarked samples) according to a priori frame sliding, and enhancing the data, wherein the training samples comprise the marked samples and unmarked samples:
1.1) construction and processing of data sets: and acquiring data in the Sea area A and utilizing a Sea landmark dataset to disclose a data set, and cutting all images into pictures with the size of 576 multiplied by 704 according to a camera data format to construct a ship data set. And marking the ship target on the partial data in the sea area A to form marked data and unmarked data. By clustering and calculating ship target frames in the marked data, 5 types of multi-scale prior frames for obtaining the ship target are mainly provided: (24, 32), (39, 69), (68, 141), (139, 297), (302, 689), where each prior box is represented by (height, width).
1.2) training sample generation: and sliding in the image of the ship data set according to the multi-scale prior frame to generate a multi-scale ship sample serving as a training sample. And (3) in the labeling data, according to the labeling sample generated by sliding the multi-scale prior frame, distinguishing a positive sample and a negative sample according to the intersection ratio of the sample frame and the labeling frame, wherein the intersection ratio is greater than 0.4 and is a positive sample, and the intersection ratio is less than 0.2 and is a negative sample. The positive sample comprises a plurality of classes of ship targets, and the class labels and the position regression labels of the ship targets are recorded, and the negative sample is a background sample. And generating unmarked samples in the unmarked data according to the sliding of the multi-scale prior frame, and storing the unmarked samples in the same folder. As shown in fig. 2, the positive sample labeling information includes a sample name, a category label, and a position regression vector. The position regression vector is calculated as follows:
in the formula, gx,gy,gwAnd ghRespectively labeling the central abscissa, central ordinate, width and length, a, of the target corresponding to the positive samplex,ay,awAnd ahThe center abscissa, center ordinate, width and length of the positive sample slide frame, respectively.
1.3) data enhancement: performing data enhancement on the generated training samples, including horizontal turning, color dithering, light variation and random cutting, so as to improve the diversity of the training samples;
2) the method comprises the following steps of constructing a teacher network, and realizing high-precision ship detection through pre-training and semi-supervised training, wherein the method comprises the following steps:
2.1) pre-training: a teacher network based on Resnet110 is set up as a ship discriminator, the teacher network is firstly classified and trained by using a large-scale classification task data set, and then classification and regression branches of the teacher network are further trained by using a labeled sample. Wherein all input training samples are scaled to a size of 32 x 32.
2.2) constructing a semi-supervised objective function of the teacher network and optimizing by adopting a random gradient descent method:
in the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetatIn order to provide trainable parameters for the teacher's network,in order to supervise the loss,as a consistency regularization term, λ1A penalty factor is indicated.
Monitoring lossThe method comprises the steps of class loss and regression loss, wherein the class loss is calculated by adopting a cross entropy function, and the regression loss is calculated by adopting a smooth-L1 function. The consistency regularization item aims to enable different data enhancement samples of the same sample to have class consistency and improve the intra-class aggregation degree of target features, and the mathematical form of the consistency regularization item is as follows:
wherein N represents the number of data enhancements, fcRepresenting a cross entropy function, ftSample prediction function representing teacher network, d(0)Representing the original sample, d(i)The ith data representing the original sample enhances the sample.
3) Constructing a lightweight student network, and guiding the student network to realize knowledge distillation on a labeled sample and a non-labeled sample by using the trained teacher network, wherein the method comprises the following steps:
3.1) building a student network based on Resnet8 as a lightweight ship discriminator.
3.2) constructing a knowledge distillation objective function of the student network and optimizing by adopting a random gradient descent method:
in the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetasIn order to train the trainable parameters of the student network,in order to supervise the loss,for the knowledge distillation regularization term, lambda2A penalty factor is indicated.
Monitoring lossThe method comprises the steps of class loss and regression loss, wherein the class loss is calculated by adopting a cross entropy function, and the regression loss is calculated by adopting a smooth-L1 function. The purpose of the knowledge distillation regular item is to enable the knowledge learned by the teacher network to be migrated into the student network, so that the student network can have consistent prediction with the teacher network, and the mathematical form is as follows:
wherein f iscRepresenting a cross entropy function, flRepresenting smooth-L1 function, fsSample prediction function representing a student network, ftSample prediction function representing teacher network, fTRepresents a sharpening function, T represents a sharpening coefficient, and the mathematical form of the sharpening function is shown as follows:
where p represents a one-dimensional input vector,representing the j-th component of pTo the power of one.
Fig. 3 is a graph showing the classification accuracy of a teacher network and a student network based on Resnet8 in different network configurations. The figure lists three teacher network-student network combinations of Resnet110-Resnet8, Resnet56-Resnet8 and Resnet8-Resnet8 for knowledge distillation in different training rounds. Where Resnet8-Resnet8 indicates that both the teacher network and the student network are network structures based on Resnet8, which is used as a reference in experiments. It can be seen that both Resnet110 and Resnet56, as teacher networks, can improve the performance of their corresponding student networks. As the performance of Resnet110 and Resnet56 are similar, the performance of their corresponding student networks is also similar.
4) Converting the trained student network into ONNX format, further optimizing and deploying the student network by adopting OpenVINO, and finally realizing rapid ship detection through the optimized student network, wherein the method comprises the following steps:
4.1) format conversion and model deployment: and converting the trained student network into an ONNX format, converting the precision of the model parameters in the ONNX format into floating point number precision 16 by utilizing an OpenVINO toolkit, deploying and reasoning, and further accelerating the speed of ship detection. The model size and the calculated speed of the ship detection are shown in table 1.
TABLE 1 comparison of model size and calculated speed for marine vessel inspection
4.2) region blocking: in remote monitoring, the position of the ship in the image changes little within 1 second. In order to reduce the calculation redundancy, the input image is divided into 4 non-overlapping regions, and the 4 regions are cyclically subjected to ship detection processing through the optimized student network, so as to further improve the speed of ship detection, and the detection result is shown in fig. 4. In the inference process, each 576 × 704 picture generates 2334 effective candidate frames through a sliding window, and the time required for completing all the regions is 2334/1000 × 74.7-174.3 ms, while the time required for completing one region is only 174.3/4-43.6 ms, so that the real-time ship detection performance is met.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.
Claims (5)
1. A ship rapid detection method based on knowledge distillation and semi-supervised learning is characterized by comprising the following steps:
1) constructing a ship data set by using existing ship data and ship data acquired by a deployment sea area, marking a ship target on partial data in the data set, generating a training sample according to a priori frame sliding, and performing data enhancement, wherein the training sample comprises a marked sample and a non-marked sample;
2) constructing a teacher network, pre-training by using a labeled sample, and then performing semi-supervised training by combining a label-free sample to realize ship detection;
3) constructing a lightweight student network, and guiding the student network to realize knowledge distillation on the marked samples and the unmarked samples by using the trained teacher network;
4) and converting the trained student network into an ONNX format, further optimizing and deploying the student network by adopting OpenVINO, and finally realizing rapid ship detection through the optimized student network.
2. The ship rapid detection method based on knowledge distillation and semi-supervised learning as recited in claim 1, wherein the step 1) comprises the following steps:
1.1) acquiring existing ship data and acquiring data in a sea area to be deployed, constructing a ship data set, marking a ship target on partial data to form marked data and unmarked data, and calculating the ship target in the marked data through clustering to acquire a multi-scale prior frame;
1.2) generating a multi-scale ship sample as a training sample according to the sliding of the multi-scale prior frame, wherein the specific operation is as follows: generating an annotated sample in the annotated data in a sliding manner, dividing the annotated sample into a positive sample and a negative sample according to the intersection and comparison of the sample and the annotated frame, wherein the positive sample comprises a plurality of classes of ship targets and records a class label and a position regression label of the ship targets, the negative sample is a background sample, and generating a non-annotated sample in the non-annotated data in a sliding manner;
1.3) carrying out data enhancement on the training sample, including horizontal turning, color dithering, light variation and random clipping.
3. The ship rapid detection method based on knowledge distillation and semi-supervised learning as claimed in claim 1, wherein in step 2), a teacher network is constructed and trained, and the method comprises the following steps:
2.1) building a teacher network with large model capacity as a ship discriminator, and pre-training the network by using a labeled sample;
In the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetatIn order to provide trainable parameters for the teacher's network,in order to supervise the loss,as a consistency regularization term, λ1Representing a penalty coefficient;
the consistency regularization item aims to enable different data enhancement samples of the same sample to have class consistency and improve the intra-class aggregation degree of target features, and the mathematical form of the consistency regularization item is as follows:
wherein N represents the number of data enhancements, fcRepresenting a cross entropy function, ftSample prediction function representing teacher network, d(0)Representing the original sample, d(i)The ith data representing the original sample enhances the sample.
4. The ship rapid detection method based on knowledge distillation and semi-supervised learning as recited in claim 1, wherein the step 3) comprises the following steps:
3.1) building a student network with a small number of parameters as a lightweight ship discriminator;
In the formula, DLRepresenting a set of annotated samples, DURepresents a set of unlabeled samples, | DLI represents the number of marked samples, | DUI represents the number of unlabeled samples, d represents one sample in the set, thetasIn order to train the trainable parameters of the student network,in order to supervise the loss,for the knowledge distillation regularization term, lambda2Representing a penalty coefficient;
monitoring lossThe method comprises the steps of category loss and regression loss, wherein the category loss is calculated by adopting a cross entropy function, and the regression loss is calculated by adopting a smooth-L1 function; the purpose of the knowledge distillation regular item is to enable the knowledge learned by the teacher network to be migrated into the student network, so that the student network can have consistent prediction with the teacher network, and the mathematical form is as follows:
wherein f iscRepresenting a cross entropy function, flRepresenting smooth-L1 function, fsSample prediction function representing a student network, ftSample prediction function representing teacher network, fTRepresents a sharpening function, T represents a sharpening coefficient, and the mathematical form of the sharpening function is shown as follows:
5. The ship rapid detection method based on knowledge distillation and semi-supervised learning as recited in claim 1, wherein the step 4) comprises the following steps:
4.1) converting the trained student network into an ONNX format, and further optimizing and deploying the student network by utilizing OpenVINO;
4.2) cutting the input image into 4 non-overlapping areas, and circularly monitoring the 4 areas through an optimized student network to further improve the speed of ship detection.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111410408.2A CN114241285B (en) | 2021-11-25 | 2021-11-25 | Ship rapid detection method based on knowledge distillation and semi-supervised learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111410408.2A CN114241285B (en) | 2021-11-25 | 2021-11-25 | Ship rapid detection method based on knowledge distillation and semi-supervised learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114241285A true CN114241285A (en) | 2022-03-25 |
CN114241285B CN114241285B (en) | 2024-05-28 |
Family
ID=80751110
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111410408.2A Active CN114241285B (en) | 2021-11-25 | 2021-11-25 | Ship rapid detection method based on knowledge distillation and semi-supervised learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114241285B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115018852A (en) * | 2022-08-10 | 2022-09-06 | 四川大学 | Abdominal lymph node detection method and device based on semi-supervised learning |
CN115019183A (en) * | 2022-07-28 | 2022-09-06 | 北京卫星信息工程研究所 | Remote sensing image model migration method based on knowledge distillation and image reconstruction |
CN115082800A (en) * | 2022-07-21 | 2022-09-20 | 阿里巴巴达摩院(杭州)科技有限公司 | Image segmentation method |
CN115082792A (en) * | 2022-06-29 | 2022-09-20 | 华南理工大学 | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning |
CN115131747A (en) * | 2022-08-25 | 2022-09-30 | 合肥中科类脑智能技术有限公司 | Knowledge distillation-based power transmission channel engineering vehicle target detection method and system |
CN115239638A (en) * | 2022-06-28 | 2022-10-25 | 厦门微图软件科技有限公司 | Industrial defect detection method, device and equipment and readable storage medium |
CN116168256A (en) * | 2023-04-19 | 2023-05-26 | 浙江华是科技股份有限公司 | Ship detection method, system and computer storage medium |
CN118196145A (en) * | 2024-05-16 | 2024-06-14 | 云南师范大学 | Teacher classroom vision tracking method based on single-point labeling and weak semi-supervision |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112464879A (en) * | 2020-12-10 | 2021-03-09 | 山东易视智能科技有限公司 | Ocean target detection method and system based on self-supervision characterization learning |
CN112529178A (en) * | 2020-12-09 | 2021-03-19 | 中国科学院国家空间科学中心 | Knowledge distillation method and system suitable for detection model without preselection frame |
CN112766087A (en) * | 2021-01-04 | 2021-05-07 | 武汉大学 | Optical remote sensing image ship detection method based on knowledge distillation |
CN113313166A (en) * | 2021-05-28 | 2021-08-27 | 华南理工大学 | Ship target automatic labeling method based on feature consistency learning |
-
2021
- 2021-11-25 CN CN202111410408.2A patent/CN114241285B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112529178A (en) * | 2020-12-09 | 2021-03-19 | 中国科学院国家空间科学中心 | Knowledge distillation method and system suitable for detection model without preselection frame |
CN112464879A (en) * | 2020-12-10 | 2021-03-09 | 山东易视智能科技有限公司 | Ocean target detection method and system based on self-supervision characterization learning |
CN112766087A (en) * | 2021-01-04 | 2021-05-07 | 武汉大学 | Optical remote sensing image ship detection method based on knowledge distillation |
CN113313166A (en) * | 2021-05-28 | 2021-08-27 | 华南理工大学 | Ship target automatic labeling method based on feature consistency learning |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115239638A (en) * | 2022-06-28 | 2022-10-25 | 厦门微图软件科技有限公司 | Industrial defect detection method, device and equipment and readable storage medium |
CN115082792A (en) * | 2022-06-29 | 2022-09-20 | 华南理工大学 | Cross-domain water surface target detection method based on feature antagonistic migration and semi-supervised learning |
CN115082792B (en) * | 2022-06-29 | 2024-05-28 | 华南理工大学 | Cross-domain water surface target detection method based on feature countermeasure migration and semi-supervised learning |
CN115082800A (en) * | 2022-07-21 | 2022-09-20 | 阿里巴巴达摩院(杭州)科技有限公司 | Image segmentation method |
CN115019183A (en) * | 2022-07-28 | 2022-09-06 | 北京卫星信息工程研究所 | Remote sensing image model migration method based on knowledge distillation and image reconstruction |
CN115019183B (en) * | 2022-07-28 | 2023-01-20 | 北京卫星信息工程研究所 | Remote sensing image model migration method based on knowledge distillation and image reconstruction |
CN115018852A (en) * | 2022-08-10 | 2022-09-06 | 四川大学 | Abdominal lymph node detection method and device based on semi-supervised learning |
CN115018852B (en) * | 2022-08-10 | 2022-12-06 | 四川大学 | Abdominal lymph node detection method and device based on semi-supervised learning |
CN115131747A (en) * | 2022-08-25 | 2022-09-30 | 合肥中科类脑智能技术有限公司 | Knowledge distillation-based power transmission channel engineering vehicle target detection method and system |
CN116168256A (en) * | 2023-04-19 | 2023-05-26 | 浙江华是科技股份有限公司 | Ship detection method, system and computer storage medium |
CN118196145A (en) * | 2024-05-16 | 2024-06-14 | 云南师范大学 | Teacher classroom vision tracking method based on single-point labeling and weak semi-supervision |
Also Published As
Publication number | Publication date |
---|---|
CN114241285B (en) | 2024-05-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN114241285B (en) | Ship rapid detection method based on knowledge distillation and semi-supervised learning | |
Moniruzzaman et al. | Deep learning on underwater marine object detection: A survey | |
CN112308019B (en) | SAR ship target detection method based on network pruning and knowledge distillation | |
CN107133569B (en) | Monitoring video multi-granularity labeling method based on generalized multi-label learning | |
Wang et al. | Tree leaves detection based on deep learning | |
Yun et al. | Selfreformer: Self-refined network with transformer for salient object detection | |
Hechun et al. | Survey of deep learning based object detection | |
Wu et al. | Research on insulator defect detection algorithm of transmission line based on CenterNet | |
Castellano et al. | Multi-view convolutional network for crowd counting in drone-captured images | |
Zhao et al. | BiTNet: a lightweight object detection network for real-time classroom behavior recognition with transformer and bi-directional pyramid network | |
Zhang et al. | Underwater target detection algorithm based on improved YOLOv4 with SemiDSConv and FIoU loss function | |
Ansari et al. | A survey of recent trends in two-stage object detection methods | |
Li et al. | Object detection for uav images based on improved yolov6 | |
Luo et al. | RBD-Net: robust breakage detection algorithm for industrial leather | |
CN117611901A (en) | Small sample image classification method based on global and local contrast learning | |
Diwate et al. | Optimization in object detection model using yolo. v3 | |
Alsaadi et al. | An automated mammals detection based on SSD-mobile net | |
Ma | Fixed-point tracking of English reading text based on mean shift and multi-feature fusion | |
CN116071544A (en) | Image description prediction method oriented to weak supervision directional visual understanding | |
Zhu et al. | Multi-size object detection assisting fault diagnosis of power systems based on improved cascaded faster R-CNNs | |
Li | Investigation of you only look once networks for vision-based small object detection | |
CN114898290A (en) | Real-time detection method and system for marine ship | |
Huang et al. | Efficient Small-Object Detection in Airport Surface Based on Maintain Feature High Resolution | |
Ali et al. | Review on Image Segmentation Methods Using Deep Learning | |
Matta et al. | Single Use Plastic Bottle Recognition and Classification Using Yolo V5 and V8 Architectures |
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 |