CN113283278A - Anti-interference laser underwater target recognition instrument - Google Patents

Anti-interference laser underwater target recognition instrument Download PDF

Info

Publication number
CN113283278A
CN113283278A CN202110025036.5A CN202110025036A CN113283278A CN 113283278 A CN113283278 A CN 113283278A CN 202110025036 A CN202110025036 A CN 202110025036A CN 113283278 A CN113283278 A CN 113283278A
Authority
CN
China
Prior art keywords
feature
map
channel
neural network
feature map
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
Application number
CN202110025036.5A
Other languages
Chinese (zh)
Other versions
CN113283278B (en
Inventor
吕以豪
陈金水
高洁
王文海
卢建刚
刘兴高
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Original Assignee
Zhejiang University ZJU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University ZJU filed Critical Zhejiang University ZJU
Priority to CN202110025036.5A priority Critical patent/CN113283278B/en
Publication of CN113283278A publication Critical patent/CN113283278A/en
Application granted granted Critical
Publication of CN113283278B publication Critical patent/CN113283278B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • 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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/30—Assessment of water resources

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an anti-interference laser underwater target recognition instrument which is formed by sequentially connecting an underwater laser scanning device, a computing device and a display device, wherein the underwater laser scanning device scans an underwater area to be detected, transmits obtained laser point cloud to the computing device, converts the laser point cloud into two-dimensional image data and performs target recognition work. The computing device comprises a neural network identification module and a storage module. The neural network identification module utilizes the combination of the EfficientNet convolution neural network and an attention mechanism to guide the network to ignore noise in the image and improve the attention to the target effective characteristic position, thereby increasing the anti-interference capability of the network. The invention realizes the intelligent, anti-interference and high-precision laser underwater target identification.

Description

Anti-interference laser underwater target recognition instrument
Technical Field
The invention relates to the field of laser underwater target identification, in particular to an anti-interference laser underwater target identification instrument.
Background
The laser underwater target recognition technology is an advanced detection technology which is being developed before a project, integrates a laser technology, a communication technology, a signal processing, target recognition and an electronic technology, and has wide application prospect. The research and development of the technology have important value in both theory and practical application.
To improve the performance and ability of existing deep learning networks, many scholars propose to further mimic the working patterns of the human brain, introducing a mechanism of attention into the neural network model. There have been many successful attempts to learn classical application scenarios in both natural language processing and image understanding analysis at two great depths. The former introduces an attention mechanism into a Recurrent Neural Network (RNN) and gives different importance to keywords, vocabularies, sentences, paragraphs and even chapter relations in input text sentences in the model training process, so that complex tasks such as content understanding and machine translation are completed more accurately. The latter combines an attention mechanism with a convolution neural network, and continuously learns a part which is more important for the completion of a final task in a feature map in a training process, so that higher weight is given to the part, which is equivalent to secondary enhancement of a self-training result, and the method is successfully applied to a plurality of visual application scenes such as classification, detection, segmentation and the like.
Disclosure of Invention
The invention aims to provide an anti-interference laser underwater target recognition instrument aiming at the defects of the prior art.
The technical scheme adopted by the invention for solving the technical problems is as follows: an anti-interference laser underwater target recognition instrument comprises an underwater laser scanning device, a computing device and a display device which are sequentially connected, wherein the underwater laser scanning device scans an underwater area to be detected, transmits obtained laser point clouds to the computing device, converts the laser point clouds into two-dimensional image data, and performs target recognition. The computing device comprises a neural network identification module which combines a traditional channel domain deep learning attention model with a class activation mapping map and utilizes an EfficientNet convolution neural network with a highly optimized structure to complete identification, and a storage module which stores data such as network parameters and an operating system.
Further, the neural network recognition module realizes the target recognition work and the process is as follows:
(1) and processing the laser point cloud data into three-channel picture data, and inputting the three-channel picture data into an EfficientNet convolution neural network model which is stored in a data storage module and contains an attention machine module.
(2) The convolutional neural network extracts features of an input picture, generates a feature map in multiple stages, gives a weight to each channel of the feature map, and multiplies the weight back to each element of the channel feature map, so that the activation amount of different channels of the layer of feature map presents new numerical value difference on the original basis, wherein the higher the weight of a certain channel is, the larger the numerical value of the newly generated feature map of the channel is, namely, the higher the importance is. The realization process is as follows:
and (2.1) performing global average pooling on the original feature map in the channel dimension, wherein the expression is as the expression (1). Wherein p isiIs the ith pixel point in the image, n is the total number of pixel points in the image,
Figure BDA0002890060560000021
is the pooled value of the output. Therefore, the feature map with the number of channels C can be converted into a feature vector F of 1 × 1 × C.
Figure BDA0002890060560000022
And (2.2) inputting the characteristic vector F into a plurality of layers of full-connection structures, and finally obtaining a weight vector W which is the same as the input dimension by reducing and increasing the characteristic dimension through mapping.
And (2.3) multiplying W back to the original feature map after passing through the sigmoid activation function layer, and giving different weights to different channels of the feature map.
(3) Generating a Class Activation Map (CAM), which is a visualization mode for the abstract feature extraction condition of the convolutional neural network, and can also be regarded as a visualization means for a network attention area, and the specific process is as follows:
and (3.1) performing global average pooling on the input feature map to finally obtain the feature vector with the dimension same as the number of the CONV channels.
And (3.2) calculating a Class Activation Map (CAM) corresponding to the current input sample by using the formula (2).
Figure BDA0002890060560000023
Wherein w1,w2...wnThe weight vector w corresponding to the category of the current input sample in the full connection layer is the same as the feature vector in dimension; miIs the ith channel image of the feature map; f is a function that zeroes out pixels in the image that are smaller than the mean. Expression (2) represents the situation that the weight vector w is multiplied by the corresponding feature map and then summed, and the finally obtained heat map is the area concerned by the network and corresponds to the input image. Because the element in w represents the weight of each element in the feature vector according to which the network makes the judgment that the input sample belongs to the category, and because the feature vector in the method is obtained by the last layer of feature graph through the GAP, the feature vector represents the activation condition of each channel of the feature graph, namely the distribution condition of the correspondingly represented feature in the input sample.
(4) For each block output signature graph of the EfficientNet convolutional neural network, an attention heat map is calculated, which is implemented as equation (3). Wherein h isiIs the weight of each channel of the feature map in the step (2), n is the number of channels of the feature map of the layer, MiIs the ith channel image of the feature map.
Figure BDA0002890060560000024
(5) And (3) making a square error Loss (MSE Loss) of a formula (4) by using a Heat map (Heat) generated by each block of the EfficientNet convolutional neural network and a Heat map (Heat) generated by a deeper block of the EfficientNet convolutional neural network, making a square error Loss (MSE Loss) by using a Heat map (Heat) of the last layer of feature map and a Class Activation Map (CAM) of the network, and adding all calculated losses into the total Loss of the network.
Figure BDA0002890060560000031
Wherein p isi，qiI-th pixel points of the two input images are respectively, and n is the total pixel point number of the image.
The process is a hierarchical structure, so that the feature graph output by each block from shallow to deep of the network can be guided by an attention system, thereby enhancing the key features and improving the identification precision of the network. Since each channel of the feature map is sensitive to features in a class of input data, the weight of the channel is the weight of the feature, which is a feature selection process.
The technical conception of the invention is as follows: the method is characterized in that the underwater laser scanning device scans an underwater area to be detected, transmits obtained laser point clouds to the computing device, converts the laser point clouds into two-dimensional image data, and performs target identification. The computing device comprises a neural network identification module which combines a traditional channel domain deep learning attention model with a class activation mapping map and utilizes an EfficientNet convolution neural network with a highly optimized structure to complete identification, and a storage module which stores data such as network parameters and an operating system. Due to the flow of the water body and the change of the refractive index, more noise points exist in the collected laser images of various underwater targets, and the recognition effect of the deep learning network is influenced. The neural network identification module in the patent combines the EfficientNet convolutional neural network with an attention mechanism, guides the network to ignore noise in an image, improves attention to a target effective characteristic position, and accordingly increases the anti-interference capability of the network.
The invention has the following beneficial effects: the anti-interference laser underwater target recognition instrument has high reasoning speed and anti-interference capability, and can quickly, stably and accurately complete the recognition of the laser underwater target.
Drawings
FIG. 1 is a flow chart of the identification implementation process of an anti-interference laser underwater target identification instrument.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings.
The anti-interference laser underwater target recognition instrument comprises an underwater laser scanning device, a computing device and a display device which are sequentially connected, wherein the underwater laser scanning device scans an underwater area to be detected, the obtained laser point cloud is transmitted to the computing device and converted into two-dimensional image data for target recognition, and the recognition result is displayed by the display device. The computing device comprises a neural network identification module which combines a traditional channel domain deep learning attention model with a class activation mapping map and utilizes an EfficientNet convolution neural network with a highly optimized structure to complete identification, and a storage module which stores data such as network parameters and an operating system.
The working process of the invention is shown in figure 1, and the specific steps are as follows:
1) the underwater laser scanning device carries out laser scanning in the operation area to form point cloud data and then returns the point cloud data to the computing device through the cable.
2) The calculation device is integrated with an EfficientNet convolutional neural network model containing an attention machine module and a data storage module, and the identification process is as follows:
and 2.1) processing the laser point cloud data into three-channel picture data, and inputting the three-channel picture data into an EfficientNet convolution neural network model which is stored in a data storage module and contains an attention machine module.
2.2) the convolutional neural network extracts the features of the input picture to generate feature maps in multiple stages, gives a weight to each channel of the feature maps, and multiplies the weight back to each element of the channel feature maps, so that the activation amount of different channels of the layer of feature maps presents new numerical value differences on the original basis, wherein the higher the weight of a certain channel is, the larger the numerical value of the newly generated feature map of the channel is, namely, the higher the importance is. The realization process is as follows:
2.2.1) performing global average pooling on the original feature map in the channel dimension, wherein the expression is as the expression (1). Wherein p isiIs the ith pixel point in the image, n is the total number of pixel points in the image,
Figure BDA0002890060560000041
is the pooled value of the output. Therefore, the feature map with the number of channels C can be converted into a feature vector F of 1 × 1 × C.
Figure BDA0002890060560000042
2.2.2) inputting the characteristic vector F into a plurality of layers of full-connection structures, and finally obtaining the weight vector W which is the same as the input dimension by reducing and increasing the characteristic dimension through mapping.
2.2.3) multiplying W back to the original feature map after passing through the sigmoid activation function layer, and giving different weights to different channels of the feature map.
2.3) generating a Class Activation Map (CAM), which is a visualization mode of the abstract feature extraction condition of the convolutional neural network and can also be regarded as a visualization means for a network attention area, and the specific process is as follows:
2.3.1) performing global average pooling on the input feature map to finally obtain the feature vector with the dimension same as the number of the CONV channels.
2.3.2) calculate the Class Activation Map (CAM) for the current input sample using equation (2):
Figure BDA0002890060560000043
wherein, w1,w2...wnThe weight vector w corresponding to the category of the current input sample in the full connection layer has the same dimension as the feature vector; miIs a characteristic diagramThe ith channel image of (1); f is a function that zeroes out pixels in the image that are smaller than the mean. Expression (2) represents the situation that the weight vector w is multiplied by the corresponding feature map and then summed, and the finally obtained heat map is the area concerned by the network and corresponds to the input image. Because the element in w represents the weight of each element in the feature vector according to which the network makes the judgment that the input sample belongs to the category, and because the feature vector in the method is obtained by the last layer of feature graph through the GAP, the feature vector represents the activation condition of each channel of the feature graph, namely the distribution condition of the correspondingly represented feature in the input sample.
2.4) calculating an attention heat map for each block output characteristic map of the EfficientNet convolutional neural network, wherein the attention heat map is realized as the formula (3). Wherein h isiIs 2.2) the weight of each channel of the feature map, n is the number of channels of the feature map of the layer, MiIs the ith channel image of the feature map.
Figure BDA0002890060560000044
2.5) the square error Loss (MSE Loss) of the formula (4) of the Heat map (Heat) generated by each block of the EfficientNet convolutional neural network and the Heat map (Heat) generated by the block of the deeper layer. The Heat map (Heat) of the last layer of feature map and the Class Activation Map (CAM) of the network do square error Loss (MSE Loss). All losses calculated are added to the total loss of the network.
Figure BDA0002890060560000051
Wherein p isi，qiI-th pixel points of the two input images are respectively, and n is the total pixel point number of the image.
The process is a hierarchical structure, so that the feature graph output by each block from shallow to deep of the network can be guided by an attention system, thereby enhancing the key features and improving the identification precision of the network. Since each channel of the feature map is sensitive to features in a class of input data, the weight of the channel is the weight of the feature, which is a feature selection process.
The above-described embodiments are intended to illustrate rather than to limit the invention, and any modifications and variations of the present invention are within the spirit and scope of the claims.

Claims (2)

1. An anti-interference laser underwater target recognition instrument is characterized by comprising an underwater laser scanning device, a computing device and a display device which are sequentially connected, wherein the underwater laser scanning device scans an underwater area to be detected, the obtained laser point cloud is transmitted to the computing device and converted into two-dimensional image data to perform target recognition work, and a recognition result is displayed by the display device. The computing device comprises a neural network identification module which combines a traditional channel domain deep learning attention model with a class activation mapping map and utilizes an EfficientNet convolution neural network with a highly optimized structure to complete identification, and a storage module which stores data such as network parameters and an operating system.
2. The anti-jamming laser underwater target recognition instrument according to claim 1, wherein the neural network recognition module realizes target recognition work by the following process:
(1) and processing the laser point cloud data into three-channel picture data, and inputting the three-channel picture data into an EfficientNet convolution neural network model which is stored in a data storage module and contains an attention machine module.
(2) The convolutional neural network extracts features of an input picture, generates a feature map in multiple stages, gives a weight to each channel of the feature map, and multiplies the weight back to each element of the channel feature map, so that the activation amount of different channels of the layer of feature map presents new numerical value difference on the original basis, wherein the higher the weight of a certain channel is, the larger the numerical value of the newly generated feature map of the channel is, namely, the higher the importance is. The realization process is as follows:
(2.1) performing global average on the original characteristic diagram in the channel dimensionPooling, which is expressed as formula (1). Wherein p isiIs the ith pixel point in the image, n is the total number of pixel points in the image,
Figure FDA0002890060550000011
is the pooled value of the output. Therefore, the feature map with the number of channels C can be converted into a feature vector F of 1 × 1 × C.
Figure FDA0002890060550000012
And (2.2) inputting the characteristic vector F into a plurality of layers of full-connection structures, and finally obtaining the weight vector W which is the same as the input dimension by reducing and increasing the characteristic dimension through mapping.
And (2.3) multiplying W back to the original feature map after passing through the sigmoid activation function layer, and giving different weights to different channels of the feature map.
(3) Generating a Class Activation Map (CAM), which is a visualization mode for the abstract feature extraction condition of the convolutional neural network, and can also be regarded as a visualization means for a network attention area, and the specific process is as follows:
and (3.1) performing global average pooling on the input feature map to finally obtain the feature vector with the dimension same as the number of the CONV channels.
And (3.2) calculating a Class Activation Map (CAM) corresponding to the current input sample by using the formula (2).
Figure FDA0002890060550000013
Wherein w1,w2...wnThe weight vector w corresponding to the category of the current input sample in the full connection layer has the same dimension as the feature vector; miIs the ith channel image of the feature map; f is a function that zeroes out pixels in the image that are smaller than the mean. Expression (2) represents the situation that the weight vector w is multiplied by the corresponding feature map and then summed, and the finally obtained heat map is the area concerned by the network and corresponds to the input image. Since the elements in w representThe network makes the weight of each element in the feature vector according to which the input sample belongs to the category judgment, and because the feature vector in the method is obtained by the last layer of feature graph through the GAP, the feature vector represents the activation condition of each channel of the feature graph, namely the distribution condition of the correspondingly represented feature in the input sample.
(4) For each block output signature graph of the EfficientNet convolutional neural network, an attention heat map is calculated, which is implemented as equation (3). Wherein h isiIs the weight of each channel of the feature map in the step (2), n is the number of channels of the feature map of the layer, MiIs the ith channel image of the feature map.
Figure FDA0002890060550000021
(5) And (3) making a Heat map (Heat) generated by each block of the EfficientNet convolutional neural network and a Heat map (Heat) generated by a block at a deeper layer thereof as a square error Loss (MSE Loss) of a formula (4), making a Heat map (Heat) of the last layer of feature map and a Class Activation Map (CAM) of the network as a square error Loss (MSE Loss), and adding all calculated losses into the total Loss of the network.
Figure FDA0002890060550000022
Wherein p isi，qiI-th pixel points of the two input images are respectively, and n is the total pixel point number of the image.
CN202110025036.5A 2021-01-08 2021-01-08 Anti-interference laser underwater target recognition instrument Active CN113283278B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110025036.5A CN113283278B (en) 2021-01-08 2021-01-08 Anti-interference laser underwater target recognition instrument

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110025036.5A CN113283278B (en) 2021-01-08 2021-01-08 Anti-interference laser underwater target recognition instrument

Publications (2)

Publication Number Publication Date
CN113283278A true CN113283278A (en) 2021-08-20
CN113283278B CN113283278B (en) 2023-03-24

Family

ID=77275485

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110025036.5A Active CN113283278B (en) 2021-01-08 2021-01-08 Anti-interference laser underwater target recognition instrument

Country Status (1)

Country Link
CN (1) CN113283278B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114140792A (en) * 2022-02-08 2022-03-04 山东力聚机器人科技股份有限公司 Micro target detection method and device based on dynamic sliding window

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019153908A1 (en) * 2018-02-11 2019-08-15 北京达佳互联信息技术有限公司 Image recognition method and system based on attention model
US20190324439A1 (en) * 2017-08-02 2019-10-24 Strong Force Iot Portfolio 2016, Llc Data monitoring systems and methods to update input channel routing in response to an alarm state
CN110738698A (en) * 2019-11-08 2020-01-31 青岛合启立智能科技有限公司 Floating type seabed data measuring method and device and electronic equipment
CN111126134A (en) * 2019-11-11 2020-05-08 浙江大学 Radar radiation source deep learning identification method based on non-fingerprint signal eliminator
CN111243042A (en) * 2020-02-28 2020-06-05 浙江德尚韵兴医疗科技有限公司 Ultrasonic thyroid nodule benign and malignant characteristic visualization method based on deep learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190324439A1 (en) * 2017-08-02 2019-10-24 Strong Force Iot Portfolio 2016, Llc Data monitoring systems and methods to update input channel routing in response to an alarm state
WO2019153908A1 (en) * 2018-02-11 2019-08-15 北京达佳互联信息技术有限公司 Image recognition method and system based on attention model
CN110738698A (en) * 2019-11-08 2020-01-31 青岛合启立智能科技有限公司 Floating type seabed data measuring method and device and electronic equipment
CN111126134A (en) * 2019-11-11 2020-05-08 浙江大学 Radar radiation source deep learning identification method based on non-fingerprint signal eliminator
CN111243042A (en) * 2020-02-28 2020-06-05 浙江德尚韵兴医疗科技有限公司 Ultrasonic thyroid nodule benign and malignant characteristic visualization method based on deep learning

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114140792A (en) * 2022-02-08 2022-03-04 山东力聚机器人科技股份有限公司 Micro target detection method and device based on dynamic sliding window

Also Published As

Publication number Publication date
CN113283278B (en) 2023-03-24

Similar Documents

Publication Publication Date Title
CN110188598B (en) Real-time hand posture estimation method based on MobileNet-v2
CN110263912B (en) Image question-answering method based on multi-target association depth reasoning
CN107844743A (en) A kind of image multi-subtitle automatic generation method based on multiple dimensioned layering residual error network
CN111444968A (en) Image description generation method based on attention fusion
CN111259940B (en) Target detection method based on space attention map
US20230260255A1 (en) Three-dimensional object detection framework based on multi-source data knowledge transfer
CN110020681A (en) Point cloud feature extracting method based on spatial attention mechanism
CN110084240A (en) A kind of Word Input system, method, medium and calculate equipment
CN113536925A (en) Crowd counting method based on attention guide mechanism
CN113283278B (en) Anti-interference laser underwater target recognition instrument
CN114676852A (en) General countermeasure disturbance generation method based on correlation class activation mapping
CN116258931B (en) Visual finger representation understanding method and system based on ViT and sliding window attention fusion
Li et al. CLS-DETR: A DETR-series object detection network using classification information to accelerate convergence
CN116958740A (en) Zero sample target detection method based on semantic perception and self-adaptive contrast learning
CN114067359B (en) Pedestrian detection method integrating human body key points and visible part attention characteristics
CN112598065B (en) Memory-based gating convolutional neural network semantic processing system and method
Si et al. Image semantic segmentation based on improved DeepLab V3 model
CN111914751B (en) Image crowd density identification detection method and system
CN114529949A (en) Lightweight gesture recognition method based on deep learning
CN114821337A (en) Semi-supervised SAR image building area extraction method based on time phase consistency pseudo-label
CN114723049A (en) Class activation mapping method and device based on gradient optimization
CN114049567A (en) Self-adaptive soft label generation method and application in hyperspectral image classification
CN116563313B (en) Remote sensing image soybean planting region segmentation method based on gating and attention fusion
Yang et al. Underwater bubble plume image generative model based on noise prior and multi conditional labels
Yang et al. Research on real-time detection of road vehicle targets based on YOLOV4 improved algorithm

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