CN109145983A - A kind of real-time scene image, semantic dividing method based on lightweight network - Google Patents
A kind of real-time scene image, semantic dividing method based on lightweight network Download PDFInfo
- Publication number
- CN109145983A CN109145983A CN201810952416.1A CN201810952416A CN109145983A CN 109145983 A CN109145983 A CN 109145983A CN 201810952416 A CN201810952416 A CN 201810952416A CN 109145983 A CN109145983 A CN 109145983A
- Authority
- CN
- China
- Prior art keywords
- image
- semantic
- lightweight network
- scene image
- real
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Abstract
The invention discloses a kind of real-time scene image, semantic dividing method based on lightweight network, comprising the following steps: S1, according to scene image data collection, one lightweight network class model of training;S2, depth convolutional neural networks model is constructed based on lightweight network class model;S3, scene image data concentration training data are input to depth convolutional neural networks, export forecast image, and the semantic tagger image comparison concentrated with scene image data, and calculate and intersect entropy loss as objective function, obtain trained image, semantic parted pattern;S4, real-time scene image to be tested is input in image, semantic parted pattern, obtains image, semantic segmentation result.The present invention passes through the MobileNetV2 that will be modified as basic network, can efficiently extract characteristics of image, in upper sampling process, with quick link block, keep parameter utilization more efficient, further improve the speed of semantic segmentation model.
Description
Technical field
The invention belongs to image, semantic segmentation technologies, and in particular to a kind of real-time scene figure based on lightweight network
As semantic segmentation method.
Background technique
Scene Semantics segmentation should belong to the application that image, semantic is segmented on scene image.Scene Semantics are divided to subsequent
Computer Vision Task is of crucial importance, such as the differentiation of unmanned middle pedestrian and vehicle etc..
Semantic segmentation is in the important component of many practical application scenes, such as machine vision, automatic Pilot, and movement
Calculate etc., scene is extremely important to the decision of practical application around accurate understanding, and therefore, runing time is assessment semantic segmentation system
Key factor of the system in practical application scene.Currently, the development of depth convolutional neural networks achieved in semantic segmentation it is aobvious
Write progress, but most of correlative studys all concentrate on improve segmentation precision rather than in the computational efficiency of model, these networks
Validity is largely the design depending on complicated depth and width model, this needs is related to much operations and ginseng
Number.Then, a large amount of practical application scene such as automated driving system is typically based on embedded device, calculates and storage resource phase
To limited.It has been more than some movements or embedded system institute energy that network for semantic segmentation, which requires very high computing resource,
It provides, causes that accuracy rate is higher, but the situation that speed is far from enough.MobileNetV2 be it is a kind of for mobile or
The limited neural network structure of mobile resources, he can be by substantially reducing the quantity of operation and memory, while keeping identical essence
Degree.
In recent years, most current optimal image, semantic dividing methods are all based on depth convolutional neural networks.
Typical semantic segmentation network structure is the structure based on coder-decoder, and encoder is an image drop sampling process, is born
The coarse semantic feature of abstract image is blamed, a decoder is followed by, a picture up-sampling process of decoder is responsible for
Up-sampling is carried out to down-sampled obtained characteristics of image and is restored to the original dimension of input picture.Lightweight network is appointed in image classification
Present very outstanding in business as a result, coming lightweight network foundation network (i.e. encoder) quickly to extract scene image
Feature can also promote speed while not sacrificing accuracy rate.
Summary of the invention
For above-mentioned deficiency in the prior art, the real-time scene image, semantic provided by the invention based on lightweight network
Dividing method solves in the prior art, realizes the slow problem of image, semantic splitting speed.
In order to achieve the above object of the invention, the technical solution adopted by the present invention are as follows: a kind of based on the real-time of lightweight network
Scene image semantic segmentation method, comprising the following steps:
S1, according to scene image data collection, lightweight network class model of the training one by image to class label;
S2, depth convolutional neural networks model is constructed based on lightweight network class model;
S3, scene image data concentration training data are input to depth convolutional neural networks, export forecast image, and with
The semantic tagger image comparison that scene image data is concentrated, and calculate and intersect entropy loss as objective function, it is trained
Image, semantic parted pattern;
S4, real-time scene image to be tested is input in image, semantic parted pattern, obtains image, semantic segmentation knot
Fruit.
Further, in the step S1,
The lightweight network class model include sequentially connected 1 conv2d unit, 17 bottleneck units,
11 × 1 conv2d, unit 17 × 7 avgpool unit and 11 × 1 conv2d unit;
Each bottleneck unit includes the first INPLACE-ABN layers, the 2nd INPLACE-ABN layers and one
Projection layer.
Further, in the lightweight network class model:
When step-length is 1, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the 2nd INPLACE-
ABN layers and a projection layer connection in sequential series, input terminal simultaneously with the first INPLACE-ABN layer and projection layer series connection,
Output end of the projection layer as bottleneck unit;
When step-length is 2, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the 2nd INPLACE-
ABN layers and a projection layer connection in sequential series, input terminal only with the first INPLACE-ABN layers connect, projection layer conduct
The output end of bottleneck unit.
Further, in the step S3:
Image, semantic parted pattern is encoder-decoder network structure;
The encoder is lightweight network class model, for extracting characteristics of image;
The decoder includes sequentially connected quick link block and one 1 × 1 convolutional layer, for restoring image resolution
Rate.
Further, the quick link block includes that sequentially connected 11 × 1 convolutional layer, 13 × 3 depth can
Separate convolution unit, 11 × 1 convolutional layer and 1 quick connection;
It includes sequentially connected depth convolutional layer and point-by-point convolutional layer that the depth, which separates convolution unit,.
Further, in the step S3, the depth convolutional neural networks training process are as follows:
S31, the training data image that scene image data is concentrated is pre-processed;
S32, using the parameter value of trained lightweight network class model as depth convolutional neural networks model just
Initial value;
S33, data amplification processing is carried out to training data image;
S34, expanded using data after training data image each pixel intersection entropy loss sum as loss function,
The training to depth convolutional neural networks model is completed using multinomial learning strategy using stochastic gradient descent method;
Further, in the step S34 multinomial learning strategy learning rate lr are as follows:
Wherein, baselr is initial learning rate;
Iter is current iteration number;
Total_iter is total the number of iterations;
Subscript power is polynomial power.
Further,
Training data image carries out pretreatment as the size of image is cut to 224 × 224 in the step S31;
Data amplification processing includes being overturn to image, being contracted at random at random between 0.5 to 2 times in the step S33
Put image and the Random-Rotation image between -10 degree and 10 degree.
The invention has the benefit that the real-time scene image, semantic segmentation side provided by the invention based on lightweight network
Method passes through the MobileNetV2 that will be modified as basic network, can efficiently extract characteristics of image, in upper sampling process,
The utilization of quick link block, keeps parameter utilization more efficient, further improves the speed of semantic segmentation model.
Detailed description of the invention
Fig. 1 is the realtime graphic semantic segmentation method implementation process based on lightweight network in embodiment provided by the invention
Figure.
Fig. 2 is two kinds of bottlencke schematic diagram of a layer structure in embodiment provided by the invention.
Fig. 3 is depth convolutional neural networks training flow chart in embodiment provided by the invention.
Specific embodiment
A specific embodiment of the invention is described below, in order to facilitate understanding by those skilled in the art this hair
It is bright, it should be apparent that the present invention is not limited to the ranges of specific embodiment, for those skilled in the art,
As long as various change is in the spirit and scope of the present invention that the attached claims limit and determine, these variations are aobvious and easy
See, all are using the innovation and creation of present inventive concept in the column of protection.
As shown in Figure 1, a kind of real-time scene image, semantic dividing method based on lightweight network, comprising the following steps:
S1, according to scene image data collection, lightweight network class model of the training one by image to class label;
Above-mentioned scene image data integrates as Cityscapes avenue contextual data collection, (contains 1 comprising 20 classification marks
A background classification), cover European 50 cities, totally 5000 data sets finely marked, using wherein 2975 as trained number
According to collection, 500 are used as validation data set, and 1525 are used as test data set.
The network structure of above-mentioned lightweight network class model is as shown in table 1:
Table 1: lightweight network class prototype network structure table
layer | Input | Operator | t | c | n | s |
1 | 2242×3 | Conv2d | - | 32 | 1 | 2 |
2 | 1122×32 | bottleneck | 1 | 16 | 1 | 1 |
3 | 1122×16 | bottleneck | 6 | 24 | 2 | 2 |
4 | 562×24 | bottleneck | 6 | 32 | 3 | 2 |
5 | 282×32 | bottleneck | 6 | 64 | 4 | 2 |
6 | 142×64 | bottleneck | 6 | 96 | 3 | 1 |
7 | 142×96 | bottleneck | 6 | 160 | 3 | 2 |
8 | 72×160 | bottleneck | 6 | 320 | 1 | 1 |
9 | 72×320 | Conv2d 1×1 | - | 1280 | 1 | 1 |
10 | 72×1280 | Avgpool7×7 | - | - | 1 | - |
11 | 1×1×1280 | Conv2d 1×1 | - | k | - |
In table, t indicates that ' expansion ' multiple, c indicate that output channel number, n indicate number of repetition, and s indicates step-length.
The lightweight network class model include sequentially connected 1 conv2d unit, 17 bottleneck units,
11 × 1 conv2d unit 17 × 7 avgpool unit and 11 × 1 conv2d unit;
Each bottleneck unit includes the first INPLACE-ABN layers, the 2nd INPLACE-ABN layers and one
Projection layer.INPLACE-ABN is the new method for efficiently reducing deep neural network training memory consumption, can replace tradition
Batch normalization and active coating, bring better semantic segmentation effect.
As shown in Fig. 2, when step-length is 1, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the
Two INPLACE-ABN layers and a projection layer connection in sequential series, input terminal simultaneously with the first INPLACE-ABN layers and projection layer
It is connected in series, output end of the projection layer as bottleneck unit;
When step-length is 2, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the 2nd INPLACE-
ABN layers and a projection layer connection in sequential series, input terminal only with the first INPLACE-ABN layers connect, projection layer conduct
The output end of bottleneck unit.
S2, depth convolutional neural networks model is constructed based on lightweight network class model;
By the characteristic pattern of the layer3 output in table 1, the characteristic pattern having a size of 1122 × 16, layer4 output, size 562
The characteristic pattern of × 24, layer5 output, having a size of 282The characteristic pattern of × 32, layer7 output, having a size of 142× 96, make respectively
It is characterized first layer, the second layer, third layer, the 4th layer for extracting network (encoder), is denoted as Encoder_1 respectively,
Encoder_2, Encoder_3, Encoder_4.
Characteristic pattern after Encoder_4 and Encoder_3 up-sampling is input in quick link block, Decoder_ is exported
1.Characteristic pattern after Decoder_1 and Encoder_2 up-sampling is input in quick link block, Decoder_2 is exported.It will
Decoder_2 is input to the characteristic pattern after Encoder_1 up-sampling and fast connect in fast, exports Decoder_3.Finally will
Encoder_4, Decoder_1, Decoder_2, Decoder_3 are up-sampled to input picture size, then by obtain four
Characteristic pattern is connected, and 11 × 1 convolution is finally passed through, and the characteristic pattern obtained and semantic segmentation mark image calculate loss letter
Number, error back propagation update weight, obtain semantic segmentation network model.
S3, scene image data concentration training data are input to depth convolutional neural networks, export forecast image, and with
The semantic tagger image comparison that scene image data is concentrated, and calculate and intersect entropy loss as objective function, it is trained
Image, semantic parted pattern;
In above-mentioned steps S3:
Image, semantic parted pattern is encoder-decoder network structure;
The encoder is lightweight network class model, for extracting characteristics of image;In order to retain the space letter of image
Breath, removes the full articulamentum of the lightweight network, and as encoder.
The decoder includes sequentially connected quick link block and one 1 × 1 convolutional layer, for restoring image resolution
Rate.Characteristic pattern is up-sampled, finally by each spy of the output of decoder in conjunction with the characteristic pattern of encoder using quick link block
Sign figure is up-sampled to original image size and is together in series, using 1 × 1 convolution, the characteristic pattern finally obtained and semantic segmentation mark
It infuses image and carries out error back propagation, obtain neural network model.
Wherein, quick link block to include sequentially connected be 11 × 1 convolutional layer, 13 × 3 separable volume of depth
Product unit, 11 × 1 convolutional layer and 1 quick connection, wherein 3 × 3 depth, which separates convolution unit, up-samples feature
Figure.
It includes sequentially connected depth convolutional layer and point-by-point convolutional layer that the depth, which separates convolution unit,;Depth convolutional layer
The filtering of lightweight is realized by applying a convolution filter in each input channel;Second is 11 × 1 convolution
Layer, i.e., point-by-point convolutional layer establish new feature by calculating the linear combination of input channel.Depth separates convolution and realizes
Decoupling between space and channel achievees the purpose that model accelerates, is widely used in lightweight network.
Wherein, cross entropy loss function are as follows:
In formula, y indicates sample label,Indicate prediction output.
As shown in figure 3, in above-mentioned steps S3, the depth convolutional neural networks training process are as follows:
S31, the training data image that scene image data is concentrated is pre-processed;
Training data image is subjected to pretreatment for the size of image is cut to 224 × 224;
S32, using the parameter value of trained lightweight network class model as depth convolutional neural networks model just
Initial value;
S33, data amplification processing is carried out to training data image;
Data amplification processing includes being overturn at random to image, between 0.5 to 2 times in the step S43, random to contract
It puts image and is spent between 10 degree -10, Random-Rotation image.
S34, expanded using data after training data image each pixel intersection entropy loss sum as loss function,
The training to depth convolutional neural networks model is completed using multinomial learning strategy using stochastic gradient descent method;
The learning rate lr of multinomial learning strategy in the step S34 are as follows:
Wherein, baselr is initial learning rate;It is set as 0.001;
Iter is current iteration number;
Total_iter is total the number of iterations;
Subscript power is polynomial power, is set as 0.9.
S4, real-time scene image to be tested is input in image, semantic parted pattern, obtains image, semantic segmentation knot
Fruit.
The invention has the benefit that the real-time scene image, semantic segmentation side provided by the invention based on lightweight network
Method passes through the MobileNetV2 that will be modified as basic network, can efficiently extract characteristics of image, in upper sampling process,
The utilization of quick link block, keeps parameter utilization more efficient, further improves the speed of semantic segmentation model.
Claims (8)
1. a kind of real-time scene image, semantic dividing method based on lightweight network, which comprises the following steps:
S1, according to scene image data collection, lightweight network class model of the training one by image to class label;
S2, depth convolutional neural networks model is constructed based on lightweight network class model;
S3, scene image data concentration training data are input to depth convolutional neural networks, export forecast image, and and scene
The semantic tagger image comparison that image data is concentrated, and calculate and intersect entropy loss as objective function, obtain trained figure
As semantic segmentation model;
S4, real-time scene image to be tested is input in image, semantic parted pattern, obtains image, semantic segmentation result.
2. the real-time scene image, semantic dividing method according to claim 1 based on lightweight network, which is characterized in that
In the step S1,
The lightweight network class model includes sequentially connected 1 conv2d unit, 17 bottleneck units, 11
× 1 conv2d unit, 17 × 7 avgpool unit and 11 × 1 conv2d unit;
Each bottleneck unit includes the first INPLACE-ABN layers, the 2nd INPLACE-ABN layers and a projection
Layer.
3. the real-time scene image, semantic dividing method according to claim 2 based on lightweight network, which is characterized in that
In the lightweight network class model:
When step-length is 1, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the 2nd INPLACE-ABN
Layer and a projection layer connection in sequential series, input terminal with the first INPLACE-ABN layers and projection layer series connection, project simultaneously
Output end of the layer as bottleneck unit;
When step-length is 2, the structure of the bottleneck unit are as follows: the first INPLACE-ABN layers, the 2nd INPLACE-ABN
Layer and a projection layer connection in sequential series, input terminal only with the first INPLACE-ABN layers connect, projection layer conduct
The output end of bottleneck unit.
4. the real-time scene image, semantic dividing method according to claim 1 based on lightweight network, which is characterized in that
In the step S3:
Image, semantic parted pattern is encoder-decoder network structure;
The encoder is lightweight network class model, for extracting characteristics of image;
The decoder includes sequentially connected quick link block and one 1 × 1 convolutional layer, for restoring image resolution ratio.
5. the realtime graphic semantic segmentation method according to claim 4 based on lightweight network, which is characterized in that
The quick link block includes that sequentially connected 11 × 1 convolutional layer, 13 × 3 depth separate convolution unit, 1
A 1 × 1 convolutional layer and 1 quick connection;
It includes sequentially connected depth convolutional layer and point-by-point convolutional layer that the depth, which separates convolution unit,.
6. the real-time scene image, semantic dividing method according to claim 4 based on lightweight network, which is characterized in that
In the step S3, the depth convolutional neural networks training process are as follows:
S31, the training data image that scene image data is concentrated is pre-processed;
S32, using the parameter value of trained lightweight network class model as the initial value of depth convolutional neural networks model;
S33, data amplification processing is carried out to training data image;
S34, expanded using data after training data image each pixel intersection entropy loss sum as loss function, use
Stochastic gradient descent method completes the training to depth convolutional neural networks model using multinomial learning strategy.
7. the real-time scene image, semantic dividing method according to claim 6 based on lightweight network, which is characterized in that
The learning rate lr of multinomial learning strategy in the step S34 are as follows:
Wherein, baselr is initial learning rate;
Iter is current iteration number;
Total_iter is total the number of iterations;
Subscript power is polynomial power.
8. the realtime graphic semantic segmentation method according to claim 7 based on lightweight network, which is characterized in that
Training data image carries out pretreatment as the size of image is cut to 224 × 224 in the step S31;
In the step S33 data amplification processing include image is overturn at random, the scaling figure at random between 0.5 to 2 times
Picture and the Random-Rotation image between -10 degree and 10 degree.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810952416.1A CN109145983A (en) | 2018-08-21 | 2018-08-21 | A kind of real-time scene image, semantic dividing method based on lightweight network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810952416.1A CN109145983A (en) | 2018-08-21 | 2018-08-21 | A kind of real-time scene image, semantic dividing method based on lightweight network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109145983A true CN109145983A (en) | 2019-01-04 |
Family
ID=64790373
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810952416.1A Pending CN109145983A (en) | 2018-08-21 | 2018-08-21 | A kind of real-time scene image, semantic dividing method based on lightweight network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109145983A (en) |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009015A (en) * | 2019-03-25 | 2019-07-12 | 西北工业大学 | EO-1 hyperion small sample classification method based on lightweight network and semi-supervised clustering |
CN110096202A (en) * | 2019-04-23 | 2019-08-06 | 华中师范大学 | A kind of lightweight image automatic cutting system and method based on deeply study |
CN110147794A (en) * | 2019-05-21 | 2019-08-20 | 东北大学 | A kind of unmanned vehicle outdoor scene real time method for segmenting based on deep learning |
CN110276766A (en) * | 2019-06-27 | 2019-09-24 | 广州久邦世纪科技有限公司 | A kind of method and device of portrait segmentation |
CN110287837A (en) * | 2019-06-17 | 2019-09-27 | 上海大学 | Sea obstacle detection method based on prior estimate network and space constraint mixed model |
CN110414428A (en) * | 2019-07-26 | 2019-11-05 | 厦门美图之家科技有限公司 | A method of generating face character information identification model |
CN110427821A (en) * | 2019-06-27 | 2019-11-08 | 高新兴科技集团股份有限公司 | A kind of method for detecting human face and system based on lightweight convolutional neural networks |
CN110490858A (en) * | 2019-08-21 | 2019-11-22 | 西安工程大学 | A kind of fabric defect Pixel-level classification method based on deep learning |
CN110531774A (en) * | 2019-09-16 | 2019-12-03 | 京东数字科技控股有限公司 | Obstacle Avoidance, device, robot and computer readable storage medium |
CN110570429A (en) * | 2019-08-30 | 2019-12-13 | 华南理工大学 | Lightweight real-time semantic segmentation method based on three-dimensional point cloud |
CN110597086A (en) * | 2019-08-19 | 2019-12-20 | 深圳元戎启行科技有限公司 | Simulation scene generation method and unmanned system test method |
CN110824481A (en) * | 2019-10-28 | 2020-02-21 | 兰州大方电子有限责任公司 | Quantitative precipitation prediction method based on radar reflectivity extrapolation |
CN111062950A (en) * | 2019-11-29 | 2020-04-24 | 南京恩博科技有限公司 | Method, storage medium and equipment for multi-class forest scene image segmentation |
CN111079649A (en) * | 2019-12-17 | 2020-04-28 | 西安电子科技大学 | Remote sensing image ground feature classification method based on lightweight semantic segmentation network |
CN111144418A (en) * | 2019-12-31 | 2020-05-12 | 北京交通大学 | Railway track area segmentation and extraction method |
CN111598095A (en) * | 2020-03-09 | 2020-08-28 | 浙江工业大学 | Deep learning-based urban road scene semantic segmentation method |
CN111696110A (en) * | 2020-06-04 | 2020-09-22 | 山东大学 | Scene segmentation method and system |
CN111814736A (en) * | 2020-07-23 | 2020-10-23 | 上海东普信息科技有限公司 | Express bill information identification method, device, equipment and storage medium |
CN111950572A (en) * | 2019-05-14 | 2020-11-17 | 北京字节跳动网络技术有限公司 | Method, apparatus, electronic device and computer-readable storage medium for training classifier |
CN112800998A (en) * | 2021-02-05 | 2021-05-14 | 南京邮电大学 | Multi-mode emotion recognition method and system integrating attention mechanism and DMCCA |
CN112819000A (en) * | 2021-02-24 | 2021-05-18 | 长春工业大学 | Streetscape image semantic segmentation system, streetscape image semantic segmentation method, electronic equipment and computer readable medium |
CN112825121A (en) * | 2019-11-20 | 2021-05-21 | 北京眼神智能科技有限公司 | Deep convolutional neural network initialization and training method, device, medium and equipment |
CN113052189A (en) * | 2021-03-30 | 2021-06-29 | 电子科技大学 | Improved MobileNet V3 feature extraction network |
CN113361505A (en) * | 2021-08-10 | 2021-09-07 | 杭州一知智能科技有限公司 | Non-specific human sign language translation method and system based on contrast decoupling element learning |
CN113822921A (en) * | 2021-11-22 | 2021-12-21 | 四川大学 | Side film intelligent head shadow measuring method based on deep neural network |
WO2022188154A1 (en) * | 2021-03-12 | 2022-09-15 | 深圳市大疆创新科技有限公司 | Front view to top view semantic segmentation projection calibration parameter determination method and adaptive conversion method, image processing device, mobile platform, and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3171297A1 (en) * | 2015-11-18 | 2017-05-24 | CentraleSupélec | Joint boundary detection image segmentation and object recognition using deep learning |
CN107480726A (en) * | 2017-08-25 | 2017-12-15 | 电子科技大学 | A kind of Scene Semantics dividing method based on full convolution and shot and long term mnemon |
CN107704866A (en) * | 2017-06-15 | 2018-02-16 | 清华大学 | Multitask Scene Semantics based on new neural network understand model and its application |
CN107766794A (en) * | 2017-09-22 | 2018-03-06 | 天津大学 | The image, semantic dividing method that a kind of Fusion Features coefficient can learn |
US20180181864A1 (en) * | 2016-12-27 | 2018-06-28 | Texas Instruments Incorporated | Sparsified Training of Convolutional Neural Networks |
-
2018
- 2018-08-21 CN CN201810952416.1A patent/CN109145983A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3171297A1 (en) * | 2015-11-18 | 2017-05-24 | CentraleSupélec | Joint boundary detection image segmentation and object recognition using deep learning |
US20180181864A1 (en) * | 2016-12-27 | 2018-06-28 | Texas Instruments Incorporated | Sparsified Training of Convolutional Neural Networks |
CN107704866A (en) * | 2017-06-15 | 2018-02-16 | 清华大学 | Multitask Scene Semantics based on new neural network understand model and its application |
CN107480726A (en) * | 2017-08-25 | 2017-12-15 | 电子科技大学 | A kind of Scene Semantics dividing method based on full convolution and shot and long term mnemon |
CN107766794A (en) * | 2017-09-22 | 2018-03-06 | 天津大学 | The image, semantic dividing method that a kind of Fusion Features coefficient can learn |
Non-Patent Citations (3)
Title |
---|
MARK SANDLER, ET AL: "MobileNetV2: Inverted Residuals and Linear Bottlenecks", 《ARXIV PREPRINT ARXIV: 1801.04381V3》 * |
MENNATULLAH SIAM, ET AL: "RTSEG: REAL-TIME SEMANTIC SEGMENTATION COMPARATIVE STUDY", 《ARXIV PREPRINT ARXIV: 1803.02758 V1》 * |
SAMUEL ROTA BULÒ, ET AL: "In-Place Activated BatchNorm for Memory-Optimized Training of DNNs", 《ARXIV PREPRINT ARXIV: 1712.02616V2》 * |
Cited By (38)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009015A (en) * | 2019-03-25 | 2019-07-12 | 西北工业大学 | EO-1 hyperion small sample classification method based on lightweight network and semi-supervised clustering |
CN110096202A (en) * | 2019-04-23 | 2019-08-06 | 华中师范大学 | A kind of lightweight image automatic cutting system and method based on deeply study |
CN110096202B (en) * | 2019-04-23 | 2020-11-20 | 华中师范大学 | Automatic lightweight image clipping system and method based on deep reinforcement learning |
CN111950572A (en) * | 2019-05-14 | 2020-11-17 | 北京字节跳动网络技术有限公司 | Method, apparatus, electronic device and computer-readable storage medium for training classifier |
CN110147794A (en) * | 2019-05-21 | 2019-08-20 | 东北大学 | A kind of unmanned vehicle outdoor scene real time method for segmenting based on deep learning |
CN110287837A (en) * | 2019-06-17 | 2019-09-27 | 上海大学 | Sea obstacle detection method based on prior estimate network and space constraint mixed model |
CN110276766A (en) * | 2019-06-27 | 2019-09-24 | 广州久邦世纪科技有限公司 | A kind of method and device of portrait segmentation |
CN110427821A (en) * | 2019-06-27 | 2019-11-08 | 高新兴科技集团股份有限公司 | A kind of method for detecting human face and system based on lightweight convolutional neural networks |
CN110414428A (en) * | 2019-07-26 | 2019-11-05 | 厦门美图之家科技有限公司 | A method of generating face character information identification model |
CN110597086B (en) * | 2019-08-19 | 2023-01-13 | 深圳元戎启行科技有限公司 | Simulation scene generation method, unmanned driving system test method and device |
CN110597086A (en) * | 2019-08-19 | 2019-12-20 | 深圳元戎启行科技有限公司 | Simulation scene generation method and unmanned system test method |
CN110490858A (en) * | 2019-08-21 | 2019-11-22 | 西安工程大学 | A kind of fabric defect Pixel-level classification method based on deep learning |
CN110490858B (en) * | 2019-08-21 | 2022-12-13 | 西安工程大学 | Fabric defective pixel level classification method based on deep learning |
CN110570429A (en) * | 2019-08-30 | 2019-12-13 | 华南理工大学 | Lightweight real-time semantic segmentation method based on three-dimensional point cloud |
CN110570429B (en) * | 2019-08-30 | 2021-12-17 | 华南理工大学 | Lightweight real-time semantic segmentation method based on three-dimensional point cloud |
CN110531774A (en) * | 2019-09-16 | 2019-12-03 | 京东数字科技控股有限公司 | Obstacle Avoidance, device, robot and computer readable storage medium |
CN110824481A (en) * | 2019-10-28 | 2020-02-21 | 兰州大方电子有限责任公司 | Quantitative precipitation prediction method based on radar reflectivity extrapolation |
CN112825121A (en) * | 2019-11-20 | 2021-05-21 | 北京眼神智能科技有限公司 | Deep convolutional neural network initialization and training method, device, medium and equipment |
CN111062950A (en) * | 2019-11-29 | 2020-04-24 | 南京恩博科技有限公司 | Method, storage medium and equipment for multi-class forest scene image segmentation |
CN111079649A (en) * | 2019-12-17 | 2020-04-28 | 西安电子科技大学 | Remote sensing image ground feature classification method based on lightweight semantic segmentation network |
CN111144418B (en) * | 2019-12-31 | 2022-12-02 | 北京交通大学 | Railway track area segmentation and extraction method |
CN111144418A (en) * | 2019-12-31 | 2020-05-12 | 北京交通大学 | Railway track area segmentation and extraction method |
CN111598095B (en) * | 2020-03-09 | 2023-04-07 | 浙江工业大学 | Urban road scene semantic segmentation method based on deep learning |
CN111598095A (en) * | 2020-03-09 | 2020-08-28 | 浙江工业大学 | Deep learning-based urban road scene semantic segmentation method |
CN111696110B (en) * | 2020-06-04 | 2022-04-01 | 山东大学 | Scene segmentation method and system |
CN111696110A (en) * | 2020-06-04 | 2020-09-22 | 山东大学 | Scene segmentation method and system |
CN111814736B (en) * | 2020-07-23 | 2023-12-29 | 上海东普信息科技有限公司 | Express delivery face list information identification method, device, equipment and storage medium |
CN111814736A (en) * | 2020-07-23 | 2020-10-23 | 上海东普信息科技有限公司 | Express bill information identification method, device, equipment and storage medium |
CN112800998A (en) * | 2021-02-05 | 2021-05-14 | 南京邮电大学 | Multi-mode emotion recognition method and system integrating attention mechanism and DMCCA |
CN112800998B (en) * | 2021-02-05 | 2022-07-29 | 南京邮电大学 | Multi-mode emotion recognition method and system integrating attention mechanism and DMCCA |
CN112819000A (en) * | 2021-02-24 | 2021-05-18 | 长春工业大学 | Streetscape image semantic segmentation system, streetscape image semantic segmentation method, electronic equipment and computer readable medium |
WO2022188154A1 (en) * | 2021-03-12 | 2022-09-15 | 深圳市大疆创新科技有限公司 | Front view to top view semantic segmentation projection calibration parameter determination method and adaptive conversion method, image processing device, mobile platform, and storage medium |
CN113052189B (en) * | 2021-03-30 | 2022-04-29 | 电子科技大学 | Improved MobileNet V3 feature extraction network |
CN113052189A (en) * | 2021-03-30 | 2021-06-29 | 电子科技大学 | Improved MobileNet V3 feature extraction network |
CN113361505B (en) * | 2021-08-10 | 2021-12-07 | 杭州一知智能科技有限公司 | Non-specific human sign language translation method and system based on contrast decoupling element learning |
CN113361505A (en) * | 2021-08-10 | 2021-09-07 | 杭州一知智能科技有限公司 | Non-specific human sign language translation method and system based on contrast decoupling element learning |
CN113822921B (en) * | 2021-11-22 | 2022-03-04 | 四川大学 | Side film intelligent head shadow measuring method based on deep neural network |
CN113822921A (en) * | 2021-11-22 | 2021-12-21 | 四川大学 | Side film intelligent head shadow measuring method based on deep neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109145983A (en) | A kind of real-time scene image, semantic dividing method based on lightweight network | |
CN109871532B (en) | Text theme extraction method and device and storage medium | |
CN109086722A (en) | Mix licence plate recognition method, device, electronic equipment | |
CN108197294A (en) | A kind of text automatic generation method based on deep learning | |
CN109285162A (en) | A kind of image, semantic dividing method based on regional area conditional random field models | |
CN110111334A (en) | A kind of crack dividing method, device, electronic equipment and storage medium | |
CN111325664B (en) | Style migration method and device, storage medium and electronic equipment | |
CN110084274A (en) | Realtime graphic semantic segmentation method and system, readable storage medium storing program for executing and terminal | |
CN111523546A (en) | Image semantic segmentation method, system and computer storage medium | |
CN105139041A (en) | Method and device for recognizing languages based on image | |
CN110210620A (en) | A kind of channel pruning method for deep neural network | |
CN106779055B (en) | Image characteristic extracting method and device | |
CN110287806A (en) | A kind of traffic sign recognition method based on improvement SSD network | |
CN107506792A (en) | A kind of semi-supervised notable method for checking object | |
CN111143578A (en) | Method, device and processor for extracting event relation based on neural network | |
Sagar | Dmsanet: Dual multi scale attention network | |
CN103136757A (en) | SAR image segmentation method based on manifold distance two-stage clustering algorithm | |
CN112084911A (en) | Human face feature point positioning method and system based on global attention | |
CN112200310B (en) | Intelligent processor, data processing method and storage medium | |
CN116740362A (en) | Attention-based lightweight asymmetric scene semantic segmentation method and system | |
Li et al. | LPCCNet: A lightweight network for point cloud classification | |
Jia et al. | MobileNetV3 with CBAM for bamboo stick counting | |
CN114782720A (en) | Method, device, electronic device, medium, and program product for determining matching of document | |
CN115995029A (en) | Image emotion analysis method based on bidirectional connection | |
Si et al. | Image semantic segmentation based on improved DeepLab V3 model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190104 |
|
RJ01 | Rejection of invention patent application after publication |