CN109977947A - A kind of image characteristic extracting method and device - Google Patents
A kind of image characteristic extracting method and device Download PDFInfo
- Publication number
- CN109977947A CN109977947A CN201910187905.7A CN201910187905A CN109977947A CN 109977947 A CN109977947 A CN 109977947A CN 201910187905 A CN201910187905 A CN 201910187905A CN 109977947 A CN109977947 A CN 109977947A
- Authority
- CN
- China
- Prior art keywords
- characteristic
- remarkable picture
- image
- picture
- extracting method
- 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
- 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
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image characteristic extracting method and device, method includes: that S1. is based on convolutional neural networks and target image to be detected, generates initial characteristics figure.S2., initial characteristics figure is input to the characteristic information for further learning Higher Order Abstract in initial module, obtains Enhanced feature figure.S3. it is based on Enhanced feature figure, generates characteristic remarkable picture of the length and width as Enhanced feature figure using supervised learning.S4. characteristic remarkable picture is normalized.S5. initial characteristics figure is multiplied with the characteristic remarkable picture after normalization, is purified characteristic pattern.Noise in energy inhibitory character figure of the present invention, prominent target information, lift scheme detect the ability of Small object.
Description
Technical field
The present invention relates to field of machine vision, in particular to a kind of image characteristic extracting method and device.
Background technique
Convolutional neural networks introduce after object detection field, and machine, which has the recognition capability of image object, greatly to be mentioned
It rises, numerous scholars has also been attracted to participate in the research in the field.Small target deteection is always one in field of image detection
Research theme full of challenge and temperature.The difficult point of small target deteection is mainly that object features information is few, to noise-sensitive.At present
The method of detection Small object can be roughly divided into two classes, and one kind is by enlarged image come amplification target, to increase Small object
Information, such methods bring are improved limited；It is another kind of be using or fusion convolutional neural networks in multilayer feature figure obtain
Sufficient Small object characteristic information is taken, but these methods are not all handled the noise in figure.Small target deteection is to making an uproar
Acoustic sensing, identification of the excessive influence of noise to Small object information.
Summary of the invention
In order to solve interference of the noise to small target deteection, the present invention proposes that pixel pays attention to network, is a kind of energy enhancing
Target information, while the image characteristic extracting method and device of non-targeted information interference are weakened again.The technical solution is as follows:
S1. it is based on convolutional neural networks and target image to be detected, generates initial characteristics figure.
S2., initial characteristics figure is input to the characteristic information for further learning Higher Order Abstract in initial module, is enhanced
Characteristic pattern.
S3. it is based on Enhanced feature figure, generates characteristic remarkable of the length and width as Enhanced feature figure using supervised learning
Figure.
S4. characteristic remarkable picture is normalized.
S5. initial characteristics figure is multiplied with the characteristic remarkable picture after normalization, is purified characteristic pattern.
Further, in step s 2, there are 4 branches in the initial module, each branch uses different size of more
A asymmetric convolution kernel extracts the Higher Order Abstract feature of different level of abstractions.Then by the output of each branch on channel dimension
It is spliced together.
Further, in step s3, during according to supervised learning thought training pattern generates characteristic remarkable picture,
Need according to labeled data generate distinguish whether be object binary map.Again by continue to optimize binary map and characteristic remarkable picture it
Between intersection entropy loss carry out guidance model study and generate correct characteristic remarkable picture.
Further, in step s3, the intersection entropy loss between binary map and characteristic remarkable picture is defined as follows:
Wherein hyper parameter λ indicates that pixel pays attention to network losses function LattIt is accounted in entire target detection model loss function
Specific gravity, w, h respectively indicate the length and width of characteristic pattern,Indicate the calibration value of i-th j pixel in binary map, uijIndicate the i-th j
The predicted value of a pixel
Further, in step s3, characteristic remarkable picture can have 2 channels, can also there was only 1 channel.
Further, in step s 5, a Channel elements in characteristic remarkable picture is optionally taken each to lead to initial characteristics figure
Element multiplication in road, is purified characteristic pattern.
A kind of image characteristics extraction device, including processor and memory are stored with computer program in the memory；
The computer program can realize described in any item methods as above when being executed by the processor.
Compared with prior art, the invention has the advantages that
1, multiple branches, the asymmetric convolution kernel comprising multiple and different sizes are used in initial module.On the one hand it reduces
Parameter mitigates over-fitting, has on the other hand then added the ability to express of nonlinear extensions model, and asymmetrical convolutional coding structure can be located
Reason increases the diversity for extracting feature to space characteristics richer in mapping.
2, characteristic remarkable picture is multiplied with initial characteristics figure, can weaken the noise in initial characteristics figure, sharpening target side
Boundary, it is opposite to enhance target information, be conducive to target detection.In addition, characteristic remarkable picture is a kind of continuous characteristic pattern, non-targeted letter
Breath will not be completely eliminated, this is conducive to retain certain contextual information, improves the robustness of network.
3, the present invention is improved in feature extraction phases, and the purge feature figure of generation can be directly as different type mesh
The input of mark detection network, promotes the performance of network detection Small object.The present invention is transplanted using simply, is had a wide range of application.
Detailed description of the invention
Fig. 1 is feature extraction overview flow chart.
Fig. 2 is based on Faster RCNN target detection model of the invention.
Fig. 3 is the internal structure of initial module.
Fig. 4 is effect of optimization figure of the invention.
Specific embodiment
Below in conjunction with Figure of description and specific preferred embodiment, the invention will be further described, but not therefore and
It limits the scope of the invention.
The image characteristic extracting method of the present embodiment, comprising: S1. is based on convolutional neural networks and target figure to be detected
Picture generates initial characteristics figure.S2., initial characteristics figure is input to the feature letter for further learning Higher Order Abstract in initial module
Breath, obtains Enhanced feature figure.S3. it is based on Enhanced feature figure, generates a length and width as Enhanced feature figure using supervised learning
Characteristic remarkable picture.S4. characteristic remarkable picture is normalized.S5. the feature after initial characteristics figure and normalization is shown
It writes figure to be multiplied, is purified characteristic pattern.
As shown in Fig. 2, showing in this example, optimize Faster RCNN in conjunction with the feature extracting method that the present invention puts forward
Detect the ability of Small object.Image, which is input in ResNet101, carries out preliminary feature extraction.We will have in ResNet101
There is the convolutional layer of identical dimensional to regard a stage as, ResNet101 then there are 5 stages, can be denoted as C1 respectively to C5.It is preferred that C3 makees
It is unfavorable for the study of subsequent initial module, too because the abstracted information for including in too shallow network layer is very little for initial characteristics figure
Deep network layer have passed through multiple pond, and the Small object of reservation is considerably less, is also unfavorable for based on this subsequent small
Target detection.
Initial characteristics figure is input to progress further feature extraction in initial module.General promotion network performance is most direct
Method be exactly to increase network depth and width, this also means that the parameter of flood tide.Flood tide parameter not only brings longer meter
Time-consuming is calculated, over-fitting is also easy to produce.Break network symmetry and improve learning ability, traditional Web vector graphic is random dilute
Dredge connection.But computer software and hardware is very poor to the computational efficiency of non-homogeneous sparse data.In order to balance the two, research shows that
Sparse matrix can be clustered is more intensive submatrix to improve calculated performance, has both maintained the sparse of network structure in this way
Property, and the high calculated performance of dense matrix is utilized.Initial module is exactly such a structure.There are 4 branches in initial module,
Each branch extracts the Higher Order Abstract feature of different level of abstractions using different size of multiple asymmetric convolution kernels.It then will be every
The output of a branch is spliced together on channel dimension.In this example, the internal structure of initial module is as shown in Figure 3.Each
Branch uses different size of convolution kernel, and different size of receptive field available in this way can be more when extracting abstract characteristics
The object of all size is adapted to well.Branch is used for dimensionality reduction near the average pondization of preceding 1x1 convolution sum 3x3 in example, can
Efficiently reduce calculation amount.By 3x3 convolution, 5x5 convolution sum 7x7 convolution be split as respectively 3x1 convolution, 1x3 convolution, 5x1 convolution,
1x5 convolution, 7x1 convolution sum 1x7 convolution reduce calculation amount also under the premise of keeping sufficient to feature extraction.If
After determining convolution step-length stride=1, as long as pad=0,1,2 are set separately, then the available phase of different branches after convolution
With the characteristic pattern of dimension, then these characteristic patterns are directly cascaded on channel dimension to obtain Enhanced feature figure.Series connection
Different branches mean to merge the feature of different scale, and Fusion Features are conducive to lift scheme to the detection energy of Small object
Power.
Based on Enhanced feature figure, one 1 channel identical with Enhanced feature figure length and width of training in the way of supervised learning
Characteristic remarkable picture.The learning objective of characteristic remarkable picture is the binary map generated according to the mark of training data.By reducing two-value
Intersection entropy loss between figure and characteristic remarkable picture carrys out guidance model study and generates correct characteristic remarkable picture.Intersect entropy loss letter
Number is as follows:
Wherein hyper parameter λ indicates that pixel pays attention to network losses function LattIt is accounted in entire target detection model loss function
Specific gravity, w, h respectively indicate the length and width of characteristic pattern,Indicate the calibration value of i-th j pixel in binary map, uijIndicate the i-th j
The predicted value of a pixel
It is normalized in this example using characteristic remarkable picture of the softmax function to generation.Spy after normalization
Each element in sign notable figure indicates that the element of corresponding position in figure in initial characteristics correctly characterizes the general of target signature
Rate.Then, it is multiplied with characteristic remarkable picture with each of initial characteristics figure channel layer, is purified characteristic pattern.Purge feature
Scheme the input as Faster RCNN, is input in region candidate network RPN (region proposal network), for mentioning
Take more accurate object candidate area (proposals).Then the pond RoI is carried out to various sizes of target candidate frame, by it
Zoom to identical size, be finally output in full articulamentum, final testing result be calculated.
The training data used in the present invention that Fig. 4 is shown, and use present invention front and back target's feature-extraction effect
Variation.The input that pixel pays attention to network, that is, initial characteristics figure is shown in Fig. 4 (b).What solid circles came out in figure is target
Feature, the part that dotted line goes out entirely are noises, and noise accumulation is easy misjudged break as target when more.Box marks in Fig. 4 (b)
Part is 5 compact arranged objects, and Fig. 4 (a) is exaggerated this part, it can be seen that has very more make an uproar between object
Sound, this leads to the obscurity boundary between object, is unfavorable for the recurrence of object space.Fig. 4 (d) is characteristic remarkable picture, it and Fig. 4
(b) it is multiplied and obtains Fig. 4 (c), i.e. purge feature figure.Fig. 4 (f) is labeled data, can be with according to target area and nontarget area
Fig. 4 (f) is become into the binary map such as Fig. 4 (e), network can be trained to generate Fig. 4 (d) using Fig. 4 (e).It is shown from Fig. 4 (c)
The characteristics of image extracted by the present invention can effectively inhibit noise it can be found that extracting object features with the present invention, make mesh
Mark object has the boundary being more clear, and is conducive to identifying and positioning for target.
A kind of image characteristics extraction device, including processor and memory are stored with computer program in the memory；
The computer program can realize described in any item methods as above when being executed by the processor.
Above-mentioned only presently preferred embodiments of the present invention, is not intended to limit the present invention in any form.Although of the invention
It has been disclosed in a preferred embodiment above, however, it is not intended to limit the invention.Therefore, all without departing from technical solution of the present invention
Content, technical spirit any simple modifications, equivalents, and modifications made to the above embodiment, should all fall according to the present invention
In the range of technical solution of the present invention protection.
Claims (7)
1. a kind of image characteristic extracting method, it is characterised in that:
S1. it is based on convolutional neural networks and target image to be detected, generates initial characteristics figure.
S2., initial characteristics figure is input to the characteristic information for further learning Higher Order Abstract in initial module, obtains Enhanced feature
Figure.
S3. it is based on Enhanced feature figure, generates characteristic remarkable picture of the length and width as Enhanced feature figure using supervised learning.
S4. characteristic remarkable picture is normalized.
S5. initial characteristics figure is multiplied with the characteristic remarkable picture after normalization, is purified characteristic pattern.
2. image characteristic extracting method according to claim 1, it is characterised in that: in step s 2, the initial module
In have 4 branches, each branch extracts the Higher Order Abstract of different level of abstractions using different size of multiple asymmetric convolution kernels
Feature.Then the output of each branch is spliced together on channel dimension.
3. according to claim 1 to 2 described in any item image characteristic extracting methods, it is characterised in that: in step s3, according to
Supervised learning thought training pattern generate characteristic remarkable picture during, need according to labeled data generate distinguish whether be object
The binary map of body.Learn to generate come guidance model by continuing to optimize the intersection entropy loss between binary map and characteristic remarkable picture again
Correct characteristic remarkable picture.
4. image characteristic extracting method according to claim 3, it is characterised in that: in step s3, binary map and feature
Intersection entropy loss between notable figure is defined as follows:
Wherein hyper parameter λ indicates that pixel pays attention to network losses function LattThe ratio accounted in entire target detection model loss function
Weight, w, h respectively indicate the length and width of characteristic pattern,Indicate the calibration value of i-th j pixel in binary map, uijIndicate i-th j picture
The predicted value of element.
5. image characteristic extracting method according to claim 4, it is characterised in that: in step s3, characteristic remarkable picture can
To have 2 channels, can also there was only 1 channel.
6. image characteristic extracting method according to any one of claims 1 to 5, it is characterised in that: in step s 5, optionally
The element multiplication in characteristic remarkable picture in a Channel elements and each channel of initial characteristics figure is taken, characteristic pattern is purified.
7. a kind of image characteristics extraction device, including processor and memory, it is characterised in that: be stored with meter in the memory
Calculation machine program；The computer program can be realized when being executed by the processor such as side as claimed in any one of claims 1 to 6
Method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910187905.7A CN109977947A (en) | 2019-03-13 | 2019-03-13 | A kind of image characteristic extracting method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910187905.7A CN109977947A (en) | 2019-03-13 | 2019-03-13 | A kind of image characteristic extracting method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109977947A true CN109977947A (en) | 2019-07-05 |
Family
ID=67078674
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910187905.7A Pending CN109977947A (en) | 2019-03-13 | 2019-03-13 | A kind of image characteristic extracting method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109977947A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110866897A (en) * | 2019-10-30 | 2020-03-06 | 上海联影智能医疗科技有限公司 | Image detection method and computer readable storage medium |
| CN111091122A (en) * | 2019-11-22 | 2020-05-01 | 国网山西省电力公司大同供电公司 | Training and detecting method and device for multi-scale feature convolutional neural network |
| CN116384448A (en) * | 2023-04-10 | 2023-07-04 | 中国人民解放军陆军军医大学 | CD severity grading system based on hybrid high-order asymmetric convolution network |
-
2019
- 2019-03-13 CN CN201910187905.7A patent/CN109977947A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110866897A (en) * | 2019-10-30 | 2020-03-06 | 上海联影智能医疗科技有限公司 | Image detection method and computer readable storage medium |
| CN111091122A (en) * | 2019-11-22 | 2020-05-01 | 国网山西省电力公司大同供电公司 | Training and detecting method and device for multi-scale feature convolutional neural network |
| CN111091122B (en) * | 2019-11-22 | 2024-01-05 | 国网山西省电力公司大同供电公司 | Training and detecting method and device for multi-scale characteristic convolutional neural network |
| CN116384448A (en) * | 2023-04-10 | 2023-07-04 | 中国人民解放军陆军军医大学 | CD severity grading system based on hybrid high-order asymmetric convolution network |
| CN116384448B (en) * | 2023-04-10 | 2023-09-12 | 中国人民解放军陆军军医大学 | CD severity grading system based on hybrid high-order asymmetric convolution network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images | |
| US10229346B1 (en) | Learning method, learning device for detecting object using edge image and testing method, testing device using the same | |
| CN109086824B (en) | Seabed substrate sonar image classification method based on convolutional neural network | |
| WO2021212736A1 (en) | Feature fusion block, convolutional neural network, person re-identification method, and related device | |
| US20220230324A1 (en) | Camouflaged object segmentation method with distraction mining | |
| US20200242451A1 (en) | Method, system and apparatus for pattern recognition | |
| CN109977947A (en) | A kind of image characteristic extracting method and device | |
| CN107239736A (en) | Method for detecting human face and detection means based on multitask concatenated convolutional neutral net | |
| CN109828251A (en) | Radar target identification method based on feature pyramid light weight convolutional neural networks | |
| CN113822209B (en) | Hyperspectral image recognition method and device, electronic equipment and readable storage medium | |
| CN112598643A (en) | Depth counterfeit image detection and model training method, device, equipment and medium | |
| WO2023272995A1 (en) | Person re-identification method and apparatus, device, and readable storage medium | |
| Verma et al. | Residual squeeze CNDS deep learning CNN model for very large scale places image recognition | |
| Liu et al. | Deep convolutional neural networks-based age and gender classification with facial images | |
| CN112017192A (en) | Glandular cell image segmentation method and system based on improved U-Net network | |
| Li et al. | Two-b-real net: Two-branch network for real-time salient object detection | |
| Le et al. | An efficient hand detection method based on convolutional neural network | |
| CN108363962B (en) | Face detection method and system based on multi-level feature deep learning | |
| CN112669343A (en) | Zhuang minority nationality clothing segmentation method based on deep learning | |
| CN115410081A (en) | Multi-scale aggregated cloud and cloud shadow identification method, system, equipment and storage medium | |
| Shen et al. | ICAFusion: Iterative cross-attention guided feature fusion for multispectral object detection | |
| Bao et al. | An improved DenseNet model to classify the damage caused by cotton aphid | |
| CN108629405A (en) | The method and apparatus for improving convolutional neural networks computational efficiency | |
| EP3832542A1 (en) | Device and method with sensor-specific image recognition | |
| CN108921017A (en) | Method for detecting human face and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190705 |
|
| WD01 | Invention patent application deemed withdrawn after publication |