CN110287849B - Lightweight depth network image target detection method suitable for raspberry pi - Google Patents

Lightweight depth network image target detection method suitable for raspberry pi Download PDF

Info

Publication number
CN110287849B
CN110287849B CN201910534572.0A CN201910534572A CN110287849B CN 110287849 B CN110287849 B CN 110287849B CN 201910534572 A CN201910534572 A CN 201910534572A CN 110287849 B CN110287849 B CN 110287849B
Authority
CN
China
Prior art keywords
network
feature
target detection
image
characteristic
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
Application number
CN201910534572.0A
Other languages
Chinese (zh)
Other versions
CN110287849A (en
Inventor
任坤
黄泷
范春奇
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.)
China Industrial Internet (Beijing) Technology Group Co.,Ltd.
Original Assignee
Beijing University of Technology
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 Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN201910534572.0A priority Critical patent/CN110287849B/en
Publication of CN110287849A publication Critical patent/CN110287849A/en
Application granted granted Critical
Publication of CN110287849B publication Critical patent/CN110287849B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/25Fusion techniques
    • G06F18/253Fusion techniques of extracted features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/29Graphical models, e.g. Bayesian networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/20Image preprocessing
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/40Scenes; Scene-specific elements in video content
    • G06V20/46Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Landscapes

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

Abstract

A lightweight deep network image target detection method suitable for a raspberry pie belongs to the field of deep learning and target detection, and comprises the steps of firstly collecting images containing targets to be detected, preprocessing the collected images, and using the preprocessed images for network training; secondly, inputting the preprocessed image into a depth separable expansion convolution neural network for feature extraction to obtain feature maps with different resolutions; inputting the feature maps with different resolutions into a feature pyramid network for feature fusion to generate a fusion feature map carrying more abundant information; and then, classifying and positioning the target to be detected by adopting a detection network to the fusion characteristic graph, and finally performing non-maximum value inhibition to obtain an optimal target detection result. The invention overcomes the difficulties that the image target detection method based on the deep neural network is difficult to realize on the raspberry platform and the image target detection method based on the lightweight network is low in detection accuracy on the raspberry platform.

Description

Lightweight depth network image target detection method suitable for raspberry pi
Technical Field
The invention belongs to the field of deep learning and target detection, and particularly relates to a lightweight deep network image target detection method suitable for a raspberry group.
Background
Object detection is a fundamental task in computer vision. The main purpose of object detection is to locate objects of interest from an input image or video, accurately classify the class of each object, and provide a bounding box for each object. Early target detection technologies adopted a manual feature extraction method, and the manually extracted features were combined with a classifier to implement a target detection task. The method for manually extracting features is not only complex, but also the extracted features have no good expression capability and robustness, so researchers propose a target detection method based on a convolutional neural network. The convolution neural network can autonomously learn useful characteristics of the image, so that the limitation of manually designing the characteristics is saved, and the accuracy of target detection is improved. These advantages make the convolutional neural network-based method rapidly replacing the traditional method a mainstream research direction in the field of target detection.
At present, an image target detection model based on a convolutional neural network optimizes a network model by deepening a network hierarchy so as to improve detection accuracy. With the deepening of network level, hardware resources required by the training model are changed from a common hardware platform to a large-scale high-performance server, and large-scale intensive computation makes it difficult to realize the depth detection model in a micro computing platform (such as raspberry pie) with limited resources. In order to solve the problems, the conventional technical scheme mainly compresses and accelerates the deep convolutional neural network, reduces network parameters and calculated amount, enables memory occupation and calculation power of a deep neural network model to meet the requirement of low configuration, and has the cost of greatly reducing detection accuracy.
Disclosure of Invention
In order to solve the technical problems, the invention aims to provide a lightweight depth network image target detection method suitable for a raspberry group, and overcomes the difficulties that an image target detection method based on a depth neural network is difficult to realize on a raspberry group platform and the detection accuracy of the image target detection method based on the lightweight network on the raspberry group platform is low.
In order to achieve the technical purpose, the technical scheme of the invention is as follows:
a lightweight depth network image target detection method suitable for a raspberry pi comprises the following steps:
(1) collecting images containing targets to be detected, and preprocessing the collected images for network training;
(2) inputting the image obtained after the preprocessing in the step (1) into a depth separable expansion convolution neural network for feature extraction to obtain feature maps with different resolutions;
(3) selecting the different resolution characteristic graphs obtained in the step (2) to input into a characteristic pyramid network for characteristic fusion, and generating a fusion characteristic graph carrying more abundant information;
(4) and (4) inputting the fusion characteristic diagram generated in the step (3) into a detection network to classify and position the target to be detected, and finally performing non-maximum value inhibition to obtain an optimal target detection result.
Further, the specific process of step (1) is as follows:
(a) selecting the types of targets to be detected, collecting images containing the targets of the types, and marking the targets, namely marking a boundary frame and type information of each target to be detected appearing in each image;
(b) when the number of the collected images is small, the existing images are utilized to carry out data enhancement operation. The method of turning, translating, rotating or adding noise and the like is adopted to create more images, so that the trained neural network has better effect;
(c) uniformly converting the image resolution into 224 × 224 to adapt to the input size;
(d) and optimizing the image based on the number of positive and negative samples, and dividing to obtain a training image set and a testing image set.
Further, the specific process of step (2) is as follows:
(A) firstly, carrying out primary feature extraction on an input image through a standard convolution block of 7 × 7 to obtain a 112 × 64 feature map, wherein 64 represents the number of channels of the feature map;
(B) sequentially extracting the depth features of the 3 depth separable volume blocks from the 112 × 64 feature maps obtained in the step (a) to obtain feature maps of 56 × 256, 28 × 512 and 14 × 1024 respectively;
(C) and (C) performing final feature extraction on the 14 × 1024 feature map obtained in the step (B) through a depth separable extended volume block to obtain a feature map with 14 × 1024 resolution.
Wherein, the depth separable volume block in step (B) can greatly compress the network parameters, which is specifically explained as follows:
3 x 3 Standard convolution with Hi*WiInput tensor L of MiAnd using a convolution kernel K of 3X 3M NsTo obtain Hi*WiOutput tensor L of NjIn which H isi,WiRespectively representing the length and width of an input image, M representing the number of channels of an input feature map, N representing the number of channels of an output feature map, and 3 x 3 representing the spatial dimension of a convolution kernel. The 3 x 3 standard convolution requires the computational cost to be:
Hi*Wi*M*N*3*3。
the depth separable convolution decomposes the standard convolution into two steps: 3 x 3 depth convolution and 1 x 1 point-by-point convolution. 3 x 3 depth convolution the input feature maps are each convolved using only a single convolution kernel. Point-by-point convolution then linearly combines the output of the depth convolution layer with a 1 x 1 convolution kernel.
Depth separable convolution employing Hi*WiInput tensor L of MiAnd using a depth convolution kernel K of 3X 1X MdTo obtain Hi*WiOutput tensor L of 1jThen, a point-by-point convolution kernel K of 1X 1M N is adoptedpTo obtain Hi*WiOutput tensor L of Nk. The depth separable convolution requires a computational cost of:
Hi*Wi*M*3*3+M*N*3*3
depth separable convolution by representing the convolution as a filtering and combining process, the computation cost of depth separable convolution is only that of conventional convolution
Figure BDA0002100789170000041
After point-by-point convolution (1 x 1), a ReLU layer is adopted for non-linearization, gradient disappearance is avoided, and network sparsity is increased to avoid overfitting. And no ReLU layer is added after the deep convolution (3 × 3) to ensure information circulation between feature maps and reduce calculation.
In addition, the depth separable expanded convolution block in the step (C) can effectively expand the receptive field of the convolution kernel and improve the regression rate and the positioning accuracy of the target under the condition of not increasing the network parameter number.
Further, the specific process of step (3) is as follows:
(I) respectively carrying out 1 × 1 convolution operation on the 28 × 512 and 14 × 1024 feature maps obtained by feature extraction in the step (2), and unifying the number of channels into 256 to obtain 28 × 256 and 14 × 256 feature maps;
and (II) adjusting the plurality of feature maps with different spatial resolutions obtained in the step (I) to the same resolution through upsampling, and then performing splicing processing to generate fused feature maps 56 x 256, 28 x 256 and 14 x 256 which carry more abundant information.
Further, the specific process of step (4) is as follows:
(i) and (3) taking the fused feature map obtained in the step (II) as an input, generating a plurality of default frames for each pixel of the input feature map, and then respectively detecting by the positioning sub-network and the classification sub-network. The detection value contains two parts: bounding box position and category confidence;
(ii) the positioning subnetwork predicts a bounding box for each default box; the classification sub-network predicts for each default box the confidence of all its classes;
(iii) and inhibiting the confidence degrees of the object types in the plurality of prediction frames and the position offset of the prediction frames relative to the default frame by using non-maximum value inhibition, and selecting the prediction frame with the minimum target loss function as the optimal prediction frame to obtain the object type and the position of the prediction frame in the optimal prediction frame.
Wherein the target loss function L (x, L, c, g) of the detection network in step (iii) is defined by the classification loss function Lconf(x, c) and a localization loss function Lloc(x, l, g) composition:
Figure BDA0002100789170000051
wherein x is a default frame on the feature map, L is a prediction frame, c is a confidence prediction value of the default frame on the feature map on each category, g is a real frame, and L is a confidence value of the default frame on the feature map on each categoryconf(x, c) denotes the softmax classification loss function of the default box on the feature map over the set of category scores c, Lloc(x, l, g) represents a position loss function, N represents the number of default boxes matched with the real boxes, and the weight coefficient α is set to 1 by cross validation.
The detection network realizes more accurate target positioning and classification by optimizing the loss function.
Adopt the beneficial effect that above-mentioned technical scheme brought:
the invention provides a method adopting deep separable convolution, which reduces redundant information in a characteristic diagram, realizes the great compression of network parameters under the condition of extremely small precision loss, and reduces the requirements on hardware memory and computing power; the depth separable expansion convolution is introduced to increase the receptive field of the characteristic diagram, and the small target detection effect and the target positioning precision are enhanced on the premise of not increasing network parameters; and the characteristic pyramid is used for carrying out multi-scale characteristic fusion, so that the characteristics under all scales have abundant image information, and the detection and target positioning precision of the small target is further improved. The method has the advantages of low memory occupation and low calculation power requirement, and can realize the target detection task on the raspberry dispatching platform.
Drawings
FIG. 1 is a schematic flow diagram of the present invention;
FIG. 2 is a model block diagram of the present invention;
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is explained below with reference to the accompanying drawings and examples, but is not limited thereto:
step 1, collecting images containing targets to be detected, and preprocessing the collected images for network training.
Selecting the types of targets to be detected, then collecting a large number of images containing the targets of the types, and marking the targets, namely marking the boundary frame and the type information of each target to be detected appearing in each image;
when the number of the collected images is small, the existing images are utilized to carry out data enhancement operation. The method of turning, translating, rotating or adding noise and the like is adopted to create more images, so that the trained neural network has better effect;
uniformly converting the image resolution into 224 × 224 to adapt to the input size;
and optimizing the image based on the number of positive and negative samples, and dividing to obtain a training image set and a testing image set.
And 2, inputting the image obtained after the preprocessing in the step 1 into a depth separable expansion convolutional neural network for feature extraction to obtain feature maps with different resolutions.
In stage 1, the input image at 224 × 224 is down-sampled by a standard convolution with 7 × 7 at step 2, outputting a feature map at 112 × 64.
In stage 2, the input feature maps of 112 × 64 are down-sampled using 3 × 3 max pooling layers, feature extraction is performed through 3 depth separable convolution layers, and feature maps of 56 × 256 are output.
In stage 3, the input feature map of 56 × 256 is down-sampled with 3 × 3 depth separable convolutional layers of step 2, and feature extraction is performed through the 3 depth separable convolutional layers, outputting a feature map of 28 × 512.
In stage 4, the 28 x 512 input feature map is downsampled with 3 x 3 depth separable convolutional layers with step 2, feature extraction is performed through 5 depth separable convolutional layers, and a 14 x 1024 feature map is output.
In stage 5, 14 × 14 input feature maps are convolved using the depth separable layer with the expansion rate of 2, and the feature maps are output 14 × 141024 while the spatial resolution of the feature maps is kept constant while the receptive field is expanded.
And 3, selecting the different resolution characteristic graphs obtained in the step 2, inputting the different resolution characteristic graphs into a characteristic pyramid network for characteristic fusion, and generating a fusion characteristic graph carrying more abundant information.
And (5) respectively carrying out 1-by-1 convolution on the feature graphs finally output in the stages 2-5 to unify the number of channels into 256.
And (4) fusing the feature map A with the 14 × 14 feature map B output by the stage 4 through 1 × 1 convolution to obtain a 14 × 14 feature map AB.
And (4) upsampling the feature map AB to obtain 28 × 28 feature maps, and then fusing the 28 × 28 feature maps C output by the stage 3 to obtain a feature map ABC.
The feature map ABC is up-sampled to obtain 56 × 56 feature maps, and then fused with the 56 × 56 feature map D output in stage 2 to obtain a feature map ABCD.
And 4, inputting the fusion characteristic diagram generated in the step 3 into a detection network to classify and position the target to be detected, and finally performing non-maximum suppression to obtain an optimal target detection result.
And (4) taking the fused feature map obtained in the step (3) as an input, generating 4 default frames for each pixel of the input feature map, and then respectively detecting by the positioning sub-network and the classification sub-network. The detection value contains two parts: bounding box position and category confidence;
the positioning sub-network generates a prediction box for each default box; the classification sub-network predicts for each default box the confidence of all its classes;
and inhibiting the confidence degrees of the object types in the plurality of prediction frames and the position offset of the prediction frames relative to the default frame by using non-maximum value inhibition, and selecting the prediction frame with the minimum target loss function as the optimal prediction frame to obtain the object type and the position of the prediction frame in the optimal prediction frame.
Wherein the target loss function L (x, L, c, g) is defined by the classification loss function Lconf(x, c) and a localization loss function Lloc(x, l, g) composition:
Figure BDA0002100789170000081
wherein x is a default frame on the feature map, L is a prediction frame, c is a confidence prediction value of the default frame on the feature map on each category, g is a real frame, and L is a confidence value of the default frame on the feature map on each categoryconf(x, c) denotes the softmax classification loss function of the default box on the feature map over the set of category scores c, Lloc(x, l, g) represents a position loss function, N represents the number of default boxes matched with the real boxes, and the weight coefficient α is set to 1 by cross validation.
The above embodiments are merely illustrative of the technical ideas of the present invention, and the technical ideas of the present invention can not be limited thereto, and any modifications based on the technical ideas of the present invention are within the scope of the present invention.

Claims (1)

1. A lightweight depth network image target detection method suitable for a raspberry pi is characterized by comprising the following steps:
(1) collecting images containing targets to be detected, and preprocessing the collected images for network training;
(2) inputting the image obtained after the preprocessing in the step (1) into a depth separable expansion convolution neural network for feature extraction to obtain feature maps with different resolutions;
(3) selecting the different resolution characteristic graphs obtained in the step (2) to input into a characteristic pyramid network for characteristic fusion, and generating a fusion characteristic graph carrying more abundant information;
(4) inputting the fusion characteristic diagram generated in the step (3) into a detection network to classify and position the target to be detected, and finally performing non-maximum value inhibition to obtain an optimal target detection result;
the specific process of the step (1) is as follows:
(A) firstly, carrying out primary feature extraction on the 224 × 224 input image through a 7 × 7 standard volume block to obtain a 112 × 64 feature map, wherein 64 represents the channel number of the feature map;
(B) sequentially extracting the depth features of the 3 depth separable volume blocks from the 112 × 64 feature maps obtained in the step (a) to obtain feature maps of 56 × 256, 28 × 512 and 14 × 1024 respectively;
(C) performing final feature extraction on the 14 × 1024 feature map obtained in the step (B) through a depth separable extended volume block to obtain a feature map with a resolution of 14 × 1024;
the specific process of the step (2) is as follows:
(I) respectively carrying out 1 × 1 convolution operation on the 28 × 512 and 14 × 1024 characteristic graphs obtained by the characteristic extraction in the step (1), and unifying the number of channels into 256 to obtain 28 × 256 and 14 × 256 characteristic graphs;
and (II) adjusting the multiple feature maps with different spatial resolutions obtained in the step (I) to the same resolution through upsampling, then performing splicing processing to generate fused feature maps 56 x 256, 28 x 256 and 14 x 256 which carry more abundant information, and completing target detection by using the multi-scale fused feature maps.
CN201910534572.0A 2019-06-20 2019-06-20 Lightweight depth network image target detection method suitable for raspberry pi Active CN110287849B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910534572.0A CN110287849B (en) 2019-06-20 2019-06-20 Lightweight depth network image target detection method suitable for raspberry pi

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910534572.0A CN110287849B (en) 2019-06-20 2019-06-20 Lightweight depth network image target detection method suitable for raspberry pi

Publications (2)

Publication Number Publication Date
CN110287849A CN110287849A (en) 2019-09-27
CN110287849B true CN110287849B (en) 2022-01-07

Family

ID=68004845

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910534572.0A Active CN110287849B (en) 2019-06-20 2019-06-20 Lightweight depth network image target detection method suitable for raspberry pi

Country Status (1)

Country Link
CN (1) CN110287849B (en)

Families Citing this family (30)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008562B (en) * 2019-10-31 2023-04-18 北京城建设计发展集团股份有限公司 Human-vehicle target detection method with feature map depth fusion
CN111047630B (en) * 2019-11-13 2023-06-13 芯启源(上海)半导体科技有限公司 Neural network and target detection and depth prediction method based on neural network
CN110991305B (en) * 2019-11-27 2023-04-07 厦门大学 Airplane detection method under remote sensing image and storage medium
CN111191508A (en) * 2019-11-28 2020-05-22 浙江省北大信息技术高等研究院 Face recognition method and device
CN111028282A (en) * 2019-11-29 2020-04-17 浙江省北大信息技术高等研究院 Unsupervised pose and depth calculation method and system
CN111199227A (en) * 2019-12-20 2020-05-26 广西柳州联耕科技有限公司 High-precision image identification method
CN111242122B (en) * 2020-01-07 2023-09-08 浙江大学 Lightweight deep neural network rotating target detection method and system
CN111199220B (en) * 2020-01-21 2023-04-28 北方民族大学 Light-weight deep neural network method for personnel detection and personnel counting in elevator
CN111204452B (en) * 2020-02-10 2021-07-16 北京建筑大学 Target detection system based on miniature aircraft
CN111340141A (en) * 2020-04-20 2020-06-26 天津职业技术师范大学(中国职业培训指导教师进修中心) Crop seedling and weed detection method and system based on deep learning
KR102497361B1 (en) * 2020-05-20 2023-02-10 한국전자통신연구원 Object detecting system and method
CN111666836B (en) * 2020-05-22 2023-05-02 北京工业大学 High-resolution remote sensing image target detection method of M-F-Y type light convolutional neural network
CN112115970B (en) * 2020-08-12 2023-03-31 南京理工大学 Lightweight image detection agricultural bird repelling method and system based on hierarchical regression
CN112183203B (en) * 2020-08-26 2024-05-28 北京工业大学 Real-time traffic sign detection method based on multi-scale pixel feature fusion
CN112132001B (en) * 2020-09-18 2023-09-08 深圳大学 Automatic tracking and quality control method for iPSC and terminal equipment
CN112183291B (en) * 2020-09-22 2024-09-10 蜜度科技股份有限公司 Method and system for detecting minimum object in image, storage medium and terminal
CN112115914B (en) * 2020-09-28 2023-04-07 北京市商汤科技开发有限公司 Target detection method and device, electronic equipment and storage medium
CN112347936A (en) * 2020-11-07 2021-02-09 南京天通新创科技有限公司 Rapid target detection method based on depth separable convolution
CN112435236B (en) * 2020-11-23 2022-08-16 河北工业大学 Multi-stage strawberry fruit detection method
CN112507872B (en) * 2020-12-09 2021-12-28 中科视语(北京)科技有限公司 Positioning method and positioning device for head and shoulder area of human body and electronic equipment
CN113270156B (en) * 2021-04-29 2022-11-15 甘肃路桥建设集团有限公司 Detection modeling and detection method and system of machine-made sandstone powder based on image processing
CN113468992B (en) * 2021-06-21 2022-11-04 四川轻化工大学 Construction site safety helmet wearing detection method based on lightweight convolutional neural network
CN113420651B (en) * 2021-06-22 2023-05-05 四川九洲电器集团有限责任公司 Light weight method, system and target detection method for deep convolutional neural network
CN113487551B (en) * 2021-06-30 2024-01-16 佛山市南海区广工大数控装备协同创新研究院 Gasket detection method and device for improving dense target performance based on deep learning
CN113642662B (en) * 2021-08-24 2024-02-20 凌云光技术股份有限公司 Classification detection method and device based on lightweight classification model
CN113971731A (en) * 2021-10-28 2022-01-25 燕山大学 Target detection method and device and electronic equipment
CN114170526B (en) * 2021-11-22 2024-09-10 中国电子科技集团公司第十五研究所 Remote sensing image multi-scale target detection and identification method based on lightweight network
CN114841307A (en) * 2022-03-01 2022-08-02 北京交通大学 Training method for binaryzation target detection neural network structure and model
CN114462555B (en) * 2022-04-13 2022-08-16 国网江西省电力有限公司电力科学研究院 Multi-scale feature fusion power distribution network equipment identification method based on raspberry group
CN115719445A (en) * 2022-12-20 2023-02-28 齐鲁工业大学 Seafood identification method based on deep learning and raspberry type 4B module

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018042388A1 (en) * 2016-09-02 2018-03-08 Artomatix Ltd. Systems and methods for providing convolutional neural network based image synthesis using stable and controllable parametric models, a multiscale synthesis framework and novel network architectures
CN108288075B (en) * 2018-02-02 2019-06-14 沈阳工业大学 A kind of lightweight small target detecting method improving SSD
CN108229442B (en) * 2018-02-07 2022-03-11 西南科技大学 Method for rapidly and stably detecting human face in image sequence based on MS-KCF
CN109214406B (en) * 2018-05-16 2021-07-09 长沙理工大学 Image classification method based on D-MobileNet neural network
CN109344821A (en) * 2018-08-30 2019-02-15 西安电子科技大学 Small target detecting method based on Fusion Features and deep learning
CN109784298A (en) * 2019-01-28 2019-05-21 南京航空航天大学 A kind of outdoor on-fixed scene weather recognition methods based on deep learning

Also Published As

Publication number Publication date
CN110287849A (en) 2019-09-27

Similar Documents

Publication Publication Date Title
CN110287849B (en) Lightweight depth network image target detection method suitable for raspberry pi
CN112287940B (en) Semantic segmentation method of attention mechanism based on deep learning
CN113052210B (en) Rapid low-light target detection method based on convolutional neural network
CN110210551B (en) Visual target tracking method based on adaptive subject sensitivity
CN109886066B (en) Rapid target detection method based on multi-scale and multi-layer feature fusion
CN115063573B (en) Multi-scale target detection method based on attention mechanism
CN109035251B (en) Image contour detection method based on multi-scale feature decoding
CN113240691A (en) Medical image segmentation method based on U-shaped network
CN113239954B (en) Attention mechanism-based image semantic segmentation feature fusion method
CN112183203A (en) Real-time traffic sign detection method based on multi-scale pixel feature fusion
CN113222124B (en) SAUNet + + network for image semantic segmentation and image semantic segmentation method
CN112950780B (en) Intelligent network map generation method and system based on remote sensing image
CN113554032B (en) Remote sensing image segmentation method based on multi-path parallel network of high perception
CN114612660A (en) Three-dimensional modeling method based on multi-feature fusion point cloud segmentation
CN114048822A (en) Attention mechanism feature fusion segmentation method for image
CN110097047B (en) Vehicle detection method based on deep learning and adopting single line laser radar
CN114820655A (en) Weak supervision building segmentation method taking reliable area as attention mechanism supervision
CN116109678A (en) Method and system for tracking target based on context self-attention learning depth network
CN110633706B (en) Semantic segmentation method based on pyramid network
CN117727046A (en) Novel mountain torrent front-end instrument and meter reading automatic identification method and system
Zuo et al. A remote sensing image semantic segmentation method by combining deformable convolution with conditional random fields
CN112668662B (en) Outdoor mountain forest environment target detection method based on improved YOLOv3 network
CN114119669A (en) Image matching target tracking method and system based on Shuffle attention
CN113361496A (en) City built-up area statistical method based on U-Net
CN117788810A (en) Learning system for unsupervised semantic segmentation

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
TR01 Transfer of patent right

Effective date of registration: 20231009

Address after: No. 202-126, 2nd Floor, Building 18, Zone 17, No. 188 South Fourth Ring West Road, Fengtai District, Beijing, 100070

Patentee after: China Industrial Internet (Beijing) Technology Group Co.,Ltd.

Address before: 100124 No. 100 Chaoyang District Ping Tian Park, Beijing

Patentee before: Beijing University of Technology

TR01 Transfer of patent right