CN110245706A - A kind of lightweight target detection network for Embedded Application - Google Patents
A kind of lightweight target detection network for Embedded Application Download PDFInfo
- Publication number
- CN110245706A CN110245706A CN201910516354.4A CN201910516354A CN110245706A CN 110245706 A CN110245706 A CN 110245706A CN 201910516354 A CN201910516354 A CN 201910516354A CN 110245706 A CN110245706 A CN 110245706A
- Authority
- CN
- China
- Prior art keywords
- convolution
- class
- convolution module
- module
- convolutional layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to deep learnings, object detection field, it is related to a kind of lightweight target detection network for Embedded Application, the network is made of two parts of lightweight basis convolutional neural networks and lightweight branch prediction network, lightweight basis convolutional neural networks include three kinds of convolution modules, lightweight branch prediction network includes a kind of convolution module, the parameter scale and operand that network is reduced while guaranteeing the accuracy rate of target detection, improve detection speed of the network under embedded environment.
Description
Technical field
The invention belongs to deep learning, object detection field more particularly to a kind of lightweight mesh for Embedded Application
Mark detection network, reduces the parameter scale and operand of network, to be promoted while guaranteeing the accuracy rate of target detection
Detection speed of the network under embedded environment.
Background technique
The task of target detection is to whether there is the target of certain particular categories in determining image, is video analysis, image
Key technology in the complexity visual task such as semantic understanding, the accuracy of object detection results will have a direct impact on high level computer
The effect of visual task.Target detection technique is widely used in many fields, such as the retail of video monitoring, wisdom.
Have much that the intelligent use based on target detection technique needs to be deployed on embedded type terminal equipment at present, and is based on
The algorithm of target detection of convolutional neural networks is more demanding to the operational capability and memory headroom of computing platform, hinders algorithm and exists
Use on intelligent terminal.Therefore, light-weight technologg is carried out to the algorithm of target detection based on convolutional neural networks, reduced
Computing cost required for algorithm, enables algorithm fast and accurately to run on embeded processor, with important practical
Value.
Summary of the invention
In order to enable algorithm of target detection fast and accurately to run on embeded processor, the invention proposes one
Kind is directed to the lightweight target detection network of Embedded Application, and the parameter scale and operand of the network are small.
The technical scheme is that providing a kind of lightweight target detection network for Embedded Application, special character
It is: including lightweight basis convolutional neural networks and lightweight branch prediction network；
Lightweight basis convolutional neural networks include convolutional layer, the first convolution unit, the second convolution unit and third
Convolution unit；
First convolution unit includes at least one first kind convolution module and at least one second class convolution module；
Second convolution unit include at least one second class convolution module or at least one second class convolution module with
At least one first kind convolution module；
The third convolution unit include at least one second class convolution module or at least one second class convolution module with
At least one first kind convolution module；
The first kind convolution module output channel number is identical as input channel number, by a convolution kernel having a size of t × t,
The convolutional layer and bypass composition that the depth convolutional layer and a convolution kernel that step-length is 1 are 1 having a size of 1 × 1, step-length；Wherein t is positive
Integer；
The second class convolution module output channel number is 2 times of input channel number, by a convolution kernel having a size of l × l,
Having a size of 1 × 1, the convolutional layer that step-length is 1 forms the depth convolutional layer and a convolution kernel that step-length is 2；Wherein l is positive integer；
The input of the convolutional layer is testing image, for obtaining the feature of testing image；
The image of convolutional layer output exports a characteristic image after the first convolution unit convolution algorithm；
A characteristic image exports b characteristic image after being input to the second convolution unit convolution algorithm；
B characteristic image exports c characteristic image after being input to third convolution unit convolution algorithm；
The lightweight branch prediction network includes at least one the 4th class convolution module and a convolutional layer c；
The 4th class convolution module output channel number is identical as input channel number, by a convolution kernel having a size of t × t,
The convolutional layer composition that the depth convolutional layer and a convolution kernel that step-length is 1 are 1 having a size of 1 × 1, step-length；
Wherein the 4th class convolution module I receives the c characteristic image of lightweight basis convolutional neural networks output；
C characteristic image exports first group of prediction after exporting after I convolution algorithm of the 4th class convolution module to convolutional layer c operation
Tensor.
Further, in order to further improve calculating speed, first kind convolution module, the second class convolution module, third
Add BN layers and active coating in class convolution module and the 4th class convolution module behind depth convolutional layer and convolutional layer；Active coating uses
Activation primitive be Relu function.
Further, in order to promote detection accuracy, the lightweight branch prediction network further includes three the 4th class convolution
Module, respectively the 4th class convolution module II, the 4th class convolution module III, IV, warp lamination b of the 4th class convolution module with
One convolutional layer b；
4th class convolution module III receives the b characteristic image of lightweight basis convolutional neural networks output, to b characteristic image
Carry out convolution algorithm；
Characteristic image c is also sequentially input after I convolution algorithm of the 4th class convolution module to warp lamination b and the 4th class convolution
Module II carries out convolution algorithm；
4th class convolution module II exports after being added with the output characteristic pattern of the 4th class convolution module III to the 4th class convolution
Module IV carries out convolution algorithm；
The output characteristic pattern of 4th class convolution module IV exports second group of prediction tensor after being input to convolutional layer b operation.
Further, in order to preferably promote detection accuracy, the lightweight branch prediction network further includes three the 4th
VII, class convolution module, respectively the 4th class convolution module V, the 4th class convolution module VI, the 4th class convolution module warp
A lamination a and convolutional layer a；
4th class convolution module VI receives a characteristic image of lightweight basis convolutional neural networks output；To a characteristic image
Carry out convolution algorithm；
The output characteristic pattern of 4th class convolution module IV be also sequentially output to warp lamination a and the 4th class convolution module V into
Row convolution algorithm；
4th class convolution module V is sequentially input after being added with the output characteristic pattern of the 4th class convolution module VI to the 4th class
Output third group predicts tensor after convolution module VII and convolutional layer a operation.
Further, in order to increase characteristic pattern port number, and bulk is not suffered a loss, and first convolution unit is also wrapped
Include third class convolution module；
The third class convolution module output channel number is 2 times of input channel number, by a convolution kernel having a size of m × m,
The convolutional layer composition that the depth convolutional layer and a convolution kernel that step-length is 1 are 1 having a size of 1 × 1, step-length；Wherein m is positive integer.
Further, first kind convolution module and the second class convolution module are two in first convolution unit, the
Three classes convolution module is one；Respectively first kind convolution module I, first kind convolution module II, the second class convolution module I,
Two class convolution modules II and third class convolution module I；
Third class convolution module I receives the detection image of convolutional layer input and sequentially inputs after carrying out convolution algorithm to second
A is exported after class convolution module I, II convolution algorithm of first kind convolution module I, the second class convolution module II and first kind convolution module
Characteristic image.
Further, the second class convolution module is one in second convolution unit, and first kind convolution module is five,
Respectively the second class convolution module III, the first kind III~first kind of convolution module convolution module VII；Second class convolution module III connects
It receives a characteristic image and carries out exporting b characteristic pattern after sequentially inputting to five first kind convolution module convolution algorithms after convolution algorithm
Picture；
Or,
Second convolution unit only includes a second class convolution module, and the second class convolution module III receives a characteristic pattern
Picture simultaneously carries out output b characteristic image after convolution algorithm.
Further, the second class convolution module is one in the third convolution unit, and first kind convolution module is one；
Respectively the second class convolution module IV, first kind convolution module VIII；Second class convolution module IV receives b characteristic image and is rolled up
C characteristic image is exported after sequentially inputting after product operation to VIII convolution algorithm of first kind convolution module；
Or,
The third convolution unit only includes a second class convolution module, and the second class convolution module IV receives b characteristic pattern
Picture simultaneously carries out output c characteristic image after convolution algorithm.
Further, the t is equal to 3, l and is equal to 3, m equal to 3.
The beneficial effects of the present invention are:
1, the lightweight target detection network that the present invention is constructed by four kinds of convolution modules, parameter amount is less, detects speed
Faster, high-precision target detection can be realized on embedded device；
2, the step-length of depth convolutional layer and convolutional layer is all 1 in first kind convolution module of the invention, and convolutional layer
Convolution nucleus number is identical as the port number of input feature vector figure, therefore the port number of characteristic pattern will not change with bulk.With
Prior art convolution module reduces 5 times compared to parameter, and the thought of residual error has also been used for reference in the design of convolution module, will be defeated by bypass
Enter characteristic pattern to be connected in output, increases the stability and training for promotion effect of network；
3, the second class convolution module structure of the invention is simple, is realized under characteristic pattern by the depth convolution that step-length is 2
Sampling, then the convolutional layer for being 2 times of input channel number by convolution nucleus number increase output channel number, the use of information effect actually calculated
Rate is higher；
4, the step-length of depth convolutional layer and convolutional layer is all 1 in the 4th class convolution module of the invention, and convolutional layer
Convolution nucleus number is identical as the port number of input feature vector figure, therefore the port number of characteristic pattern will not change with bulk, and
The interference to prediction tensor is generated is reduced by removal bypass.
Detailed description of the invention
Fig. 1 is the structure chart of one lightweight target detection network of embodiment；
Fig. 2 is the structure chart of one lightweight basis convolutional neural networks of embodiment；
Fig. 3 is first kind convolution module in embodiment one；
Fig. 4 is the second class convolution module in embodiment one；
Fig. 5 is third class convolution module in embodiment one；
Fig. 6 is lightweight branch prediction network structure in embodiment one；
Fig. 7 is the 4th class convolution module in embodiment one；
Fig. 8 is the structure chart that embodiment the second light industry bureau quantified goal detects network；
Fig. 9 is the structure chart of three lightweight target detection network of embodiment.
Appended drawing reference in figure are as follows: 1- third class convolution module I, the second class of 2- convolution module I, 3- first kind convolution module I,
The second class of 4- convolution module II, 5- first kind convolution module II, the second class of 6- convolution module III, 7- first kind convolution module III,
8- first kind convolution module IV, 9- first kind convolution module V, 10- first kind convolution module VI, 11- first kind convolution module
The second class convolution module IV of VII, 12-, 13- first kind convolution module VIII, the 4th class convolution module I of 14-, the 4th class convolution mould of 15-
Block II, the 4th class convolution module III of 16-, the 4th class convolution module IV of 17-, the 4th class convolution module V of 18-, the 4th class of 19- volume
Volume module VI, the 4th class convolution module VII of 20-.
Specific embodiment
The basic idea of the invention is that the parameter scale and operand of target detection network are reduced, to reach embedded
The purpose of high-precision target detection is realized in equipment.
The present invention be directed to Embedded Application lightweight target detection network, including lightweight basis convolutional neural networks and
Lightweight branch prediction network；
Lightweight basis convolutional neural networks include convolutional layer, the first convolution unit, the second convolution unit and third convolution
Unit；Wherein the first convolution unit includes at least one first kind convolution module and at least one second class convolution module, may be used also
To include at least one third class convolution module；Second convolution unit include at least one second class convolution module and at least one
First kind convolution module, or only include at least one second class convolution module；Third convolution unit includes at least one second class
Convolution module and at least one first kind convolution module, or only include at least one second class convolution module；First kind convolution mould
Block output channel number is identical as input channel number, by a convolution kernel having a size of t × t, the depth convolutional layer that step-length is 1 and one
The convolutional layer and bypass composition that convolution kernel is 1 having a size of 1 × 1, step-length；Wherein t is positive integer；The output of second class convolution module is logical
Road number is 2 times of input channel number, by a convolution kernel having a size of l × l, the depth convolutional layer and a convolution kernel that step-length is 2
Having a size of 1 × 1, the convolutional layer that step-length is 1 is formed；Wherein l is positive integer；Third class convolution module output channel number is that input is logical
2 times of road number, by a convolution kernel having a size of m × m, the depth convolutional layer that step-length is 1 and a convolution kernel having a size of 1 × 1, step
A length of 1 convolutional layer composition；Wherein m is positive integer.
Lightweight branch prediction network includes at least one the 4th class convolution module and a convolutional layer c；4th class convolution
Module output channel number is identical as input channel number, by a convolution kernel having a size of t × t, the depth convolutional layer and one that step-length is 1
A convolution kernel is formed having a size of the convolutional layer that 1 × 1, step-length is 1.
The input of convolutional layer is testing image；The image of convolutional layer output is defeated after by the first convolution unit convolution algorithm
A characteristic image out；A characteristic image exports b characteristic image after being input to the second convolution unit convolution algorithm；The input of b characteristic image
C characteristic image is exported after to third convolution unit convolution algorithm；Lightweight branch prediction network receives a characteristic image, b characteristic pattern
Picture and/or c characteristic image export each prediction tensor after carrying out convolution algorithm.
The present invention is further described through below in conjunction with drawings and the specific embodiments.
Embodiment one
As shown in Figure 1, the present embodiment lightweight target detection network is by lightweight basis convolutional neural networks and lightweight
Two parts of branch prediction network are formed.
The present embodiment lightweight basis convolutional neural networks structure is as shown in Fig. 2, the cromogram that its input is 416 × 416
Picture.S in Fig. 2 in various convolution modules refers to step-length when convolution algorithm, and × N indicates that the port number of output characteristic pattern is N.
It can be seen that the convolution module in Fig. 2 is divided into three classes, respectively first kind convolution module, the second class convolution module and third class are rolled up
Volume module.Wherein, first kind convolution module includes first kind convolution module I 3, first kind convolution module II 5, first kind convolution mould
III 7~first kind of block convolution module VII 11 and first kind convolution module VIII 13；Second class convolution module includes the second class convolution
Module I 2, the second class convolution module II 4, the second class convolution module III 6, the second class convolution module IV 12；Third class convolution module
Including third class convolution module I 1.
In the present embodiment, third class convolution module I 1, the second class convolution module I 2, first kind convolution module I 3, second
Class convolution module II 4 and first kind convolution module II 5 successively sort, and constitute the first convolution unit；Second class convolution module III 6 and
The first kind III 7~first kind of convolution module convolution module VII 11 constitutes the second convolution unit；Second class convolution module IV 12 and
A kind of convolution module VIII 13 constitutes third convolution unit.The image of convolutional layer output is after by the first convolution unit convolution algorithm
Export a characteristic image；A characteristic image exports b characteristic image after being input to the second convolution unit convolution algorithm；B characteristic image is defeated
C characteristic image is exported after entering to third convolution unit convolution algorithm.
In other embodiments, the demand different for realization, the first convolution unit can only include the second class convolution module
With first kind convolution module and quantity it is unlimited.
As shown in figure 3, the step-length of first kind convolution module is 1, output channel number is identical as input channel number.This first
Class convolution module is by a convolution kernel having a size of 3 × 3 depth convolutional layer (step-length 1) and one 1 × 1 convolutional layer (step-length
For 1) along with bypass composition, bypass is accomplished that the operation that characteristic pattern is added.Add BN behind depth convolutional layer and convolutional layer
Layer and active coating.The activation primitive that active coating in the first kind convolution module uses is Relu function, the operation of Relu function
Simple and calculating speed is fast, the negative fraction of input data only need to be set to 0, i.e. y=max (0, x) is more suitable in embedded ring
It is used under border.
The step-length of depth convolutional layer and convolutional layer is all 1 in the first kind convolution module, and the convolution nucleus number of convolutional layer
It is identical as the port number of input feature vector figure, therefore the port number of characteristic pattern will not change with bulk.It will by bypass
Input feature vector figure is connected in output, can increase the stability and training for promotion effect of network.The ginseng of the first kind convolution module
Quantity is 3 × 3 × n+n × 1 × 1 × n, and n is the port number (effect) of input feature vector figure.As n=128, the parameter of the module
Amount is 17536.
Darknet-53 in YOLOv3 network includes the convolution module that step-length is 1, and the convolution module is by two groups of convolution lists
Member composition, in the case where not considering BN layer parameter, its number of parameters isWherein n is input
The port number of characteristic pattern.Convolution module parameter in first kind convolution module of the present invention and Darknet-53 reduces multiple such as following formula
It is shown.
As shown in figure 4, the step-length of the second class convolution module is 2, output channel number is 2 times of input channel number, this
Two class convolution modules (are walked by a convolution kernel having a size of 3 × 3 depth convolutional layer (step-length 2) and one 1 × 1 convolutional layer
It is a length of 1) to form.Add BN layers and active coating behind depth convolutional layer and convolutional layer.The depth convolution for being wherein 2 by step-length is real
The down-sampling of existing characteristic pattern, then the convolutional layer for being 2 times of input channel number by convolution nucleus number increase output channel number.The convolution mould
The parameter amount of block is 9n+2n2, n is the port number of input feature vector figure.As n=128, which is 33920.
As shown in figure 5, the step-length of third class convolution module is 1, output channel number is 2 times of input channel number, the third
It is 2 that class convolution module, which is by step-length, the depth convolution in double lightweight convolution module i.e. the second class convolution module of port number
Layer step-length is set as 1.The parameter amount of the convolution module is 9n+2n2, n is the port number of input feature vector figure.As n=128,
The value is 33920.
The present embodiment lightweight branch prediction network structure as shown in fig. 6, by 7 step-lengths be 1 and input channel number with it is defeated
The 4th equal class convolution module of port number out, 2 warp volume modules, the convolutional layer and two spies that 31 × 1 step-lengths are 1
Sign figure phase add operation is formed.The input of lightweight branch prediction network is three features of lightweight basis convolutional network output
Figure.
7 the 4th class convolution modules are respectively the 4th class convolution module I 14, the 4th class convolution module II 15, the 4th class volume
Volume module III 16, the 4th class convolution module IV 17, the 4th class convolution module V 18, the 4th class convolution module VI 19 and the 4th class volume
Volume module VII 20；2 warp volume modules are respectively warp lamination b and warp lamination a；3 convolutional layers are respectively convolutional layer a, volume
Lamination b and convolutional layer c.
Wherein the 4th class convolution module I 14, the 4th class convolution module III 16 and the 4th class convolution module VI 19 receive respectively
C characteristic image, b characteristic image and a characteristic image of lightweight basis convolutional neural networks output；Characteristic image c is through the 4th class
First group of prediction tensor is exported after output to convolutional layer c operation after I 14 convolution algorithm of convolution module；Characteristic image c is through the 4th class
It is also sequentially input after I 14 convolution algorithm of convolution module to warp lamination b and the 4th class convolution module II 15 and carries out convolution algorithm；It is special
Sign image b is exported after III 16 convolution algorithm of the 4th class convolution module；4th class convolution module II 15 and the 4th class convolution module
Output to the 4th class convolution module IV 17 carries out convolution algorithm after III 16 output characteristic pattern is added；4th class convolution module IV 17
Output characteristic pattern be input to convolutional layer b operation after export second group of prediction tensor；The output of 4th class convolution module IV 17 is special
Sign figure is also sequentially output to warp lamination and the 4th class convolution module V 18 and carries out convolution algorithm；Characteristic image a is rolled up through the 4th class
Volume module VI 19 exports after carrying out convolution algorithm；The output feature of 4th class convolution module V 18 and the 4th class convolution module VI 19
Figure is sequentially input after being added to the 4th class convolution module VII 20 and output third group prediction tensor after convolutional layer a operation.
As shown in fig. 7, the 4th class convolution module by a convolution kernel having a size of 3 × 3 depth convolutional layer (step-length 1) with
One 1 × 1 convolutional layer (step-length 1) composition.Add BN layers and active coating behind depth convolutional layer and convolutional layer.The convolution mould
The step-length of depth convolutional layer and convolutional layer is all 1 in block, and the port number phase of the convolution nucleus number of convolutional layer and input feature vector figure
Together, therefore the port number of characteristic pattern will not change with bulk.It is light that the convolution module, which does not use the reason of bypass,
The main function for quantifying branch prediction network is different with basis convolutional neural networks, primarily as a prediction network, this
When again by input feature vector figure with output characteristic pattern be directly added, can interference prediction tensor generation.
Deconvolution, also referred to as transposition convolution have been widely used in the tasks such as scene cut, generation model at present and have worked as
In.Deconvolution operates the size that can restore characteristic pattern, but cannot restore the numerical value before convolution operation.With directly to image
Carry out up-sampling compare, deconvolution can also complete up-sampling function, and also can learning parameter, can be more by training
The good information using characteristic pattern.Light-weighted convolution module will first pass through depth convolution followed by 1 × 1 convolution, and deeply
Degree convolution can not carry out the mixing of interchannel information.It, can be in low-level feature figure using add operation for depth convolution
The information of upper fusion high-level characteristic figure, so that there are also high-level semantics features, Jin Erti for the existing minutia of fused characteristic pattern
Rise the detection effect of Small object.
By following experiment, the detection accuracy and speed of the present embodiment target detection network are tested, at the same with it is existing
Some target detection networks compare.
Use MPSoC ZCU106 development board as embedded testing platform.Include in ZU7EV device on the development boardCortex-A53 application processor.Cortex-A53 application processor is the four cores processing based on Armv8-A framework
Device, its main feature is that it is low in energy consumption, efficiency is high, be mainly directed towards the application scenarios such as top box of digital machine, tablet computer, DTV.
The positioning of Cortex-A53 processor is the middle-end processor of balance quality and efficiency, can be represented in practical intelligent terminal application
The operational capability of processor used.The maximum operation frequency of Cortex-A53 processor on ZCU106 development board is 1.20GHz,
Running memory is 4GB.Therefore, it is based onCortexTM- A53 application processor is to lightweight mesh proposed by the invention
The detection speed of mark detection network is tested.Experimental method is as follows:
1) lightweight target is examined using VOC2007 and 2012 data sets and COCO2017 data set in GPU platform
Survey grid network is trained；Optimization process uses the prior art.
2) it is converted again by format and the network model trained is deployed on arm processor.
3) lightweight target detection network is tested using 2007 data set of VOC and COCO2017 data set.It uses
Ncnn deep learning frame, test program are based on C++ programming language and write.With the comparison result such as table of other target detection networks
1, table 2, shown in table 3.Tables 1 and 2 the result shows that, the detection accuracy of lightweight target detection network proposed by the invention compared with
It is high.From table 3 it can be seen that detection speed ratio YOLOv3-416 network of the present networks under embedding assembly platform fast 14.87
Times, it was demonstrated that practicability of the network structure under embedding assembly platform.
Detection accuracy of the table 1 based on 2007 target detection network of PASCAL VOC
Experimental result of the table 2 based on COCO data set
| Network model | MAP@IOU=0.5:0.95 | MAP@IOU=0.5 | MAP@IOU=0.75 |
| SSD300 | 23.2 | 41.2 | 23.4 |
| SSD512 | 26.8 | 46.5 | 27.8 |
| YOLOv2-416 | 21.6 | 44.0 | 19.2 |
| YOLOv3-416 | 30.7 | 55.3 | 30.9 |
| YOLOv3-608 | 33.0 | 57.9 | 34.4 |
| YOLOv3-Tiny | — | 33.1 | — |
| Present networks | 24.0 | 43.7 | 23.5 |
Model velocity test experiments result of the table 3 based on embedded platform
Embodiment two
As shown in figure 8, being the present embodiment target detection network, it can be seen that is different from the first embodiment is that in the implementation
In example, lightweight branch prediction network only includes the 4th class convolution module I 14 and convolutional layer c, corresponding 4th class convolution module I
14 receive the c characteristic image of lightweight basis convolutional neural networks output, and c characteristic image is through I 14 convolution of the 4th class convolution module
First group of prediction tensor is exported after output to convolutional layer c operation after operation.When to detection accuracy require do not need Tai Gao and image
In target to be detected be big target when, the target detection network of the embodiment can be chosen, with faster speed obtain target inspection
Survey result.
Embodiment three
As shown in figure 9, being the present embodiment target detection network, it can be seen that is different from the first embodiment is that in the implementation
In example, lightweight branch prediction network includes the 4th class convolution module I 14, the 4th class convolution module II 15, the 4th class convolution mould
Block III 16, the 4th class convolution module IV 17, warp lamination b, convolutional layer c and convolutional layer b.
4th class convolution module III 16 receives the b characteristic image of lightweight basis convolutional neural networks output, to b characteristic pattern
As carrying out convolution algorithm；4th class convolution module I 14 receives the c characteristic image of lightweight basis convolutional neural networks output, and c is special
Sign image exports first group of prediction tensor after exporting after I 14 convolution algorithm of the 4th class convolution module to convolutional layer c operation；Feature
Image c also sequentially input after I 14 convolution algorithm of the 4th class convolution module to warp lamination b and the 4th class convolution module II 15 into
Row convolution algorithm；4th class convolution module II 15 exports after being added with the output characteristic pattern of the 4th class convolution module III 16 to the 4th
Class convolution module IV 17 carries out convolution algorithm；After the output characteristic pattern of 4th class convolution module IV 17 is input to convolutional layer b operation
Second group of prediction tensor is exported, embodiment detection speed with higher compared with embodiment one has compared with embodiment two
There is higher detection accuracy.
Claims (9)
1. a kind of lightweight target detection network for Embedded Application, it is characterised in that: including lightweight basis convolution mind
Through network and lightweight branch prediction network；
Lightweight basis convolutional neural networks include convolutional layer, the first convolution unit, the second convolution unit and third convolution
Unit；
First convolution unit includes at least one first kind convolution module and at least one second class convolution module；
Second convolution unit include at least one second class convolution module or at least one second class convolution module at least
One first kind convolution module；
The third convolution unit include at least one second class convolution module or at least one second class convolution module at least
One first kind convolution module；
The first kind convolution module output channel number is identical as input channel number, by a convolution kernel having a size of t × t, step-length
The convolutional layer and bypass composition for being 1 having a size of 1 × 1, step-length for 1 depth convolutional layer and a convolution kernel；Wherein t is positive whole
Number；
The second class convolution module output channel number is 2 times of input channel number, by a convolution kernel having a size of l × l, step-length
It is 2 depth convolutional layer and a convolution kernel having a size of 1 × 1, the convolutional layer that step-length is 1 forms；Wherein l is positive integer；
The input of the convolutional layer is testing image, for obtaining the feature of testing image；
The image of convolutional layer output exports a characteristic image after the first convolution unit convolution algorithm；
A characteristic image exports b characteristic image after being input to the second convolution unit convolution algorithm；
B characteristic image exports c characteristic image after being input to third convolution unit convolution algorithm；
The lightweight branch prediction network includes at least one the 4th class convolution module and a convolutional layer c；
The 4th class convolution module output channel number is identical as input channel number, by a convolution kernel having a size of t × t, step-length
The convolutional layer for being 1 having a size of 1 × 1, step-length for 1 depth convolutional layer and a convolution kernel forms；
Wherein the 4th class convolution module I (14) receives the c characteristic image of lightweight basis convolutional neural networks output；
C characteristic image exports after the 4th class convolution module I (14) convolution algorithm to convolutional layer c, exported after operation first group it is pre-
Survey tensor.
2. a kind of lightweight target detection network for Embedded Application according to claim 1, it is characterised in that: the
In depth convolutional layer and volume in a kind of convolution module, the second class convolution module, third class convolution module and the 4th class convolution module
Add BN layers and active coating behind lamination；The activation primitive that active coating uses is Relu function.
3. a kind of lightweight target detection network for Embedded Application according to claim 1, it is characterised in that: institute
Stating lightweight branch prediction network further includes three the 4th class convolution modules, respectively the 4th class convolution module II (15), the 4th
Class convolution module III (16), the 4th class convolution module IV (17), a warp lamination b and a convolutional layer b；
4th class convolution module III (16) receives the b characteristic image of lightweight basis convolutional neural networks output, to b characteristic image
Carry out convolution algorithm；
Characteristic image c is also sequentially input after the 4th class convolution module I (14) convolution algorithm to warp lamination b and the 4th class convolution
Module II (15) carries out convolution algorithm；
4th class convolution module II (15) exports after being added with the output characteristic pattern of the 4th class convolution module III (16) to the 4th class
Convolution module IV (17) carries out convolution algorithm；
The output characteristic pattern of 4th class convolution module IV (17) is input to convolutional layer b, and second group of prediction tensor is exported after operation.
4. a kind of lightweight target detection network for Embedded Application according to claim 3, it is characterised in that:
The lightweight branch prediction network further includes three the 4th class convolution modules, respectively the 4th class convolution module V
(18), the 4th class convolution module VI (19), the 4th class convolution module VII (20), a warp lamination a and a convolutional layer a；
4th class convolution module VI (19) receives a characteristic image of lightweight basis convolutional neural networks output；To a characteristic image
Carry out convolution algorithm；
The output characteristic pattern of 4th class convolution module IV (17) is also sequentially output to warp lamination a and the 4th class convolution module V
(18) convolution algorithm is carried out；
4th class convolution module V (18) is sequentially input after being added with the output characteristic pattern of the 4th class convolution module VI (19) to
Four class convolution modules VII (20) and convolutional layer a export third group and predict tensor after operation.
5. a kind of lightweight target detection network for Embedded Application according to claim 4, it is characterised in that:
First convolution unit further includes third class convolution module；
The third class convolution module output channel number is 2 times of input channel number, by a convolution kernel having a size of m × m, step-length
The convolutional layer for being 1 having a size of 1 × 1, step-length for 1 depth convolutional layer and a convolution kernel forms；Wherein m is positive integer.
6. a kind of lightweight target detection network for Embedded Application according to claim 5, it is characterised in that:
First kind convolution module and the second class convolution module are two in first convolution unit, and third class convolution module is
One；Respectively first kind convolution module I (3), first kind convolution module II (5), the second class convolution module I (2), the second class volume
Volume module II (4) and third class convolution module I (1)；
Third class convolution module I (1) receives the detection image of convolutional layer input and sequentially inputs after carrying out convolution algorithm to second
Class convolution module I (2), first kind convolution module I (3), the second class convolution module II (4) and first kind convolution module II (5) volume
A characteristic image is exported after product operation.
7. a kind of lightweight target detection network for Embedded Application according to claim 6, it is characterised in that:
The second class convolution module is one in second convolution unit, and first kind convolution module is five, respectively the second class
Convolution module III (6), first kind convolution module III (7)~first kind convolution module VII (11)；Second class convolution module III (6) connects
It receives a characteristic image and carries out exporting b characteristic pattern after sequentially inputting to five first kind convolution module convolution algorithms after convolution algorithm
Picture；
Or,
Second convolution unit only includes a second class convolution module, and the second class convolution module III (6) receives a characteristic image
And b characteristic image is exported after carrying out convolution algorithm.
8. a kind of lightweight target detection network for Embedded Application according to claim 7, it is characterised in that:
The second class convolution module is one in the third convolution unit, and first kind convolution module is one；Respectively the second class
Convolution module IV (12), first kind convolution module VIII (13)；Second class convolution module IV (12) receives b characteristic image and is rolled up
C characteristic image is exported after sequentially inputting after product operation to first kind convolution module VIII (13) convolution algorithm；
Or,
The third convolution unit only includes a second class convolution module, and the second class convolution module IV (12) receives b characteristic pattern
Picture simultaneously carries out output c characteristic image after convolution algorithm.
9. a kind of lightweight target detection network for Embedded Application according to any one of claims 1 to 8, feature
Be: the t is equal to 3, l and is equal to 3, m equal to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910516354.4A CN110245706B (en) | 2019-06-14 | 2019-06-14 | Lightweight target detection method for embedded application |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910516354.4A CN110245706B (en) | 2019-06-14 | 2019-06-14 | Lightweight target detection method for embedded application |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110245706A true CN110245706A (en) | 2019-09-17 |
| CN110245706B CN110245706B (en) | 2021-03-26 |
Family
ID=67887235
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910516354.4A Active CN110245706B (en) | 2019-06-14 | 2019-06-14 | Lightweight target detection method for embedded application |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110245706B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008994A (en) * | 2019-11-14 | 2020-04-14 | 山东万腾电子科技有限公司 | Moving target real-time detection and tracking system and method based on MPSoC |
| CN114529798A (en) * | 2022-02-21 | 2022-05-24 | 山东浪潮科学研究院有限公司 | Production line product quality inspection implementation method based on TinyML and auxiliary system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846446A (en) * | 2018-07-04 | 2018-11-20 | 国家新闻出版广电总局广播科学研究院 | The object detection method of full convolutional network is merged based on multipath dense feature |
| CN108960230A (en) * | 2018-05-31 | 2018-12-07 | 中国科学院自动化研究所 | Lightweight target identification method and device based on rotation rectangle frame |
| CN109034215A (en) * | 2018-07-09 | 2018-12-18 | 东北大学 | A kind of safety cap wearing detection method based on depth convolutional neural networks |
| WO2019040168A1 (en) * | 2017-08-25 | 2019-02-28 | Microsoft Technology Licensing, Llc | Object detection based on deep neural network |
| CN109447066A (en) * | 2018-10-18 | 2019-03-08 | 中国人民武装警察部队海警学院 | A kind of quick accurately single phase object detection method and device |
| CN109685017A (en) * | 2018-12-26 | 2019-04-26 | 中山大学 | A kind of ultrahigh speed real-time target detection system and detection method based on light weight neural network |
-
2019
- 2019-06-14 CN CN201910516354.4A patent/CN110245706B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019040168A1 (en) * | 2017-08-25 | 2019-02-28 | Microsoft Technology Licensing, Llc | Object detection based on deep neural network |
| CN108960230A (en) * | 2018-05-31 | 2018-12-07 | 中国科学院自动化研究所 | Lightweight target identification method and device based on rotation rectangle frame |
| CN108846446A (en) * | 2018-07-04 | 2018-11-20 | 国家新闻出版广电总局广播科学研究院 | The object detection method of full convolutional network is merged based on multipath dense feature |
| CN109034215A (en) * | 2018-07-09 | 2018-12-18 | 东北大学 | A kind of safety cap wearing detection method based on depth convolutional neural networks |
| CN109447066A (en) * | 2018-10-18 | 2019-03-08 | 中国人民武装警察部队海警学院 | A kind of quick accurately single phase object detection method and device |
| CN109685017A (en) * | 2018-12-26 | 2019-04-26 | 中山大学 | A kind of ultrahigh speed real-time target detection system and detection method based on light weight neural network |
Non-Patent Citations (4)
| Title |
|---|
| JOSEPH REDMON 等: "YOLOv3: an incremental improvement", 《ARXIV:1804.02767V1》 * |
| MARK SANDLER等: "MobileNetV2: Inverted Residuals and Linear Bottlenecks", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
| PEDOEEM, JONATHAN 等: "YOLO-LITE: A Real-Time Object Detection Algorithm Optimized for Non-GPU Computers", 《ARXIV:1811.05588V1》 * |
| 王晓青 等: "应用于嵌入式图形处理器的实时目标检测方法", 《光学学报》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008994A (en) * | 2019-11-14 | 2020-04-14 | 山东万腾电子科技有限公司 | Moving target real-time detection and tracking system and method based on MPSoC |
| CN114529798A (en) * | 2022-02-21 | 2022-05-24 | 山东浪潮科学研究院有限公司 | Production line product quality inspection implementation method based on TinyML and auxiliary system |
| CN114529798B (en) * | 2022-02-21 | 2024-05-21 | 山东浪潮科学研究院有限公司 | TinyML-based production line product quality inspection implementation method and auxiliary system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110245706B (en) | 2021-03-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105046279B (en) | A kind of analog circuit fault method for classifying modes | |
| CN110378222B (en) | Method and device for detecting vibration damper target and identifying defect of power transmission line | |
| CN109831478A (en) | Rule-based and model distributed processing intelligent decision system and method in real time | |
| CN107220217A (en) | Characteristic coefficient training method and device that logic-based is returned | |
| CN105825511A (en) | Image background definition detection method based on deep learning | |
| CN110245754A (en) | A kind of knowledge distillating method based on position sensing figure | |
| Yu et al. | Real-time object detection towards high power efficiency | |
| CN109376763A (en) | Sample classification method, system and medium based on multisample reasoning neural network | |
| CN110245706A (en) | A kind of lightweight target detection network for Embedded Application | |
| CN110458096A (en) | A kind of extensive commodity recognition method based on deep learning | |
| CN106295670A (en) | Data processing method and data processing equipment | |
| Wu et al. | Multi-sensor signal fusion for a compound fault diagnosis method with strong generalization and noise-tolerant performance | |
| Tamosiunas et al. | Investigating cosmological GAN emulators using latent space interpolation | |
| CN116152254A (en) | Industrial leakage target gas detection model training method, detection method and electronic equipment | |
| Zhang et al. | IDD-Net: industrial defect detection method based on deep-learning | |
| CN111368707A (en) | Face detection method, system, device and medium based on feature pyramid and dense block | |
| Jeon et al. | Target capacity filter pruning method for optimized inference time based on YOLOv5 in embedded systems | |
| CN110263808A (en) | A kind of Image emotional semantic classification method based on LSTM network and attention mechanism | |
| Zeng et al. | Energy-efficient hardware acceleration of shallow machine learning applications | |
| Lorentz et al. | Explaining defect detection with saliency maps | |
| CN105760478A (en) | Large-scale distributed data clustering method based on machine learning | |
| Zhang et al. | A spatial–spectral adaptive learning model for textile defect images recognition with few labeled data | |
| Li et al. | Fpga-based object detection acceleration architecture design | |
| Yang et al. | Convolutional autoencoder-based sensor fault classification | |
| Long et al. | Cascaded feature enhancement network model for real-time video monitoring of power system |
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 |