CN109308458A - A method of small target deteection precision is promoted based on characteristic spectrum change of scale - Google Patents
A method of small target deteection precision is promoted based on characteristic spectrum change of scale Download PDFInfo
- Publication number
- CN109308458A CN109308458A CN201811010695.6A CN201811010695A CN109308458A CN 109308458 A CN109308458 A CN 109308458A CN 201811010695 A CN201811010695 A CN 201811010695A CN 109308458 A CN109308458 A CN 109308458A
- Authority
- CN
- China
- Prior art keywords
- characteristic spectrum
- spectrum
- characteristic
- dimension
- scale
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
- G06V10/7515—Shifting the patterns to accommodate for positional errors
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a kind of methods for promoting small target deteection precision based on characteristic spectrum change of scale, belong to target detection technique field.The present invention is on the basis of conventional target detection algorithm SSD, by to characteristic spectrum channel, width and high reorganization operation, the width of further feature spectrum and height are put and be twice, then obtained new characteristic spectrum and shallow-layer characteristic spectrum are subjected to cascading and obtain new characteristic spectrum, predict position and the classification of target on this basis.The present invention does not have to through additional calculating, and the mode that characteristic spectrum recombinates is carried out change of scale amplification, realizes the promotion of small target deteection effect.Compared to traditional SSD detection method, the program can more effectively promote the precision of small target deteection, also be able to maintain good precision effect simultaneously for larger target.
Description
Technical field
The invention belongs to target detection technique fields, and in particular to one kind composes change of scale based on target detection network characterization
Method.
Background technique
Target detection is a basic research topic in computer vision field, and target detection is a large amount of high-level visions
The indispensable premise of task, including activity or event recognition, scene content understanding etc..And target detection is also applied to many realities
Border task, such as intelligent video monitoring, content-based image retrieval, robot navigation and augmented reality etc..Target detection pair
Computer vision field and practical application are of great significance, and large quantities of researchers is motivated to pay close attention to simultaneously in the past few decades
Input research.And with the development of powerful machine Learning Theory and Signature Analysis Techique, nearly more than ten years target detection project
Relevant research activities is growing on and on, and has newest research achievement and practical application to deliver and announce every year.Nevertheless, working as
The Detection accuracy of front method is still lower, and the detection effect of especially Small object is not satisfactory, may not apply to practical general
Detection task.Therefore, target detection is also solved perfectly far away, is still important the research topic of challenge.
The algorithm of target detection of mainstream is mainly based upon deep learning model at present, is segmented into two major classes: (1) two rank
The detection algorithm of section, the problem of will test are divided into two stages, first generation candidate region, then classify to candidate region,
The Typical Representative of this kind of algorithm is R-CNN (the Regions with Convolutional Neural extracted based on candidate frame
Network) serial algorithm, such as R-CNN, Fast R-CNN, Faster R-CNN etc.;(2) one stage detection algorithms, do not need
Candidate frame extracts the stage, directly generates the class probability and position coordinate value of object, for example than more typical algorithm YOLO (You
Only Look Once) and SSD (Single Shot MultiBox Detector).The main performance of target detection model refers to
Mark is detection accuracy and speed, and for detection accuracy, target detection will consider the positioning accuracy of object, and is not merely classification
Accuracy.In practical applications, although SSD algorithm has preferable effect, SSD algorithm pair in detection speed and precision
In the detection effect of Small object and bad, the present invention is directed to optimization aims to detect SSD algorithm, promote the detection effect of Small object.
Summary of the invention
Goal of the invention of the invention is: using for traditional algorithm of target detection SSD and directly predicts in characteristic spectrum
Defect present in the classification of target and position, the present invention is on the basis of the characteristic spectrum of SSD method, using the side of change of scale
Characteristic spectrum scale is amplified and carries out characteristic spectrum fusion by method, then carries out target prediction again.
The method that small target deteection precision is promoted based on characteristic spectrum change of scale of the invention, including the following steps:
Step 1: to image to be detected carry out SSD method object detection process, obtain shallow-layer to deep layer characteristic spectrum,
The width and height of middle later layer characteristic spectrum are preceding layer feature spectrum width and high half;
Step 2: scale is carried out to the characteristic spectrum for obtaining the non-first floor and carries out conversion process:
Step 201: on the spectrum channel dimension of characteristic spectrum to be transformed, it is logical that port number C being divided into C/4 group in order
Road obtains the subcharacter that C/4 dimension is 4*W*H and composes, and wherein W indicates that the width of characteristic spectrum, H indicate the height of characteristic spectrum;
Step 202: the region of each 4*1*1 being converted into 1*2*2 dimension in sequence in each subcharacter spectrum;
Step 203: the 1*2*2 feature converted in each subcharacter spectrum is combined to obtain according to original relative positional relationship
The subcharacter that one new dimension size is 1* (W*2) * (H*2) is composed;
Step 204: C/4 sub- characteristic spectrums being subjected to cascading in order, obtaining dimension size is (C/4) * (W*2) *
(H*2) characteristic spectrum;
Step 3: to transformed characteristic spectrum, carrying out cascaded series on the spectrum channel dimension of channel with one layer of characteristic spectrum thereon
It closes, obtains new characteristic spectrum;
Based on the new characteristic spectrum, target detection is carried out to specified Small object, obtains object detection results.
In conclusion by adopting the above-described technical solution, the beneficial effects of the present invention are:
Proposed by the invention can effectively promote Small object inspection based on a kind of method based on characteristic spectrum change of scale
The precision of survey, will be deep by characteristic spectrum channel, width and high reorganization operation on the basis of conventional target detection algorithm SSD
The width and height of layer characteristic spectrum are put and are twice, and obtained new characteristic spectrum and shallow-layer characteristic spectrum are then carried out cascading and obtained newly
Characteristic spectrum, on this basis predict target position and classification.The present invention does not have to through additional calculating, by characteristic spectrum weight
The mode of group carries out change of scale amplification, realizes the promotion of small target deteection effect.Compared to traditional SSD detection method, the party
Case can more effectively promote the precision of small target deteection, also be able to maintain good precision effect simultaneously for larger target.
Detailed description of the invention
Fig. 1: feature of present invention composes scale transformation method figure;
Fig. 2: the prediction flow chart of feature of present invention spectrum fusion.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below with reference to embodiment and attached drawing, to this hair
It is bright to be described in further detail.
The present invention is put characteristic spectrum scale using the method for change of scale on the basis of the characteristic spectrum that SSD method obtains
Greatly and characteristic spectrum fusion is carried out, then carries out target prediction again, specific implementation step process is as follows:
Firstly, carrying out the object detection process of SSD method to image to be detected, characteristic spectrum of the shallow-layer to deep layer, example are obtained
The SSD method such as exported for 6 layers, the characteristic spectrum of corresponding shallow-layer to deep layer are followed successively by n1, n2, n3, n4, n5, n6;
Then to the scale of the characteristic spectrum of the obtained non-first floor carry out conversion process (i.e. change of scale object is kth layer,
Middle k ≠ 1), it is substantially a kind of characteristic spectrum channel, width and high reorganization operation.Dimension with characteristic spectrum is that 512*19*19 is
Example, operates by change of scale, and the present invention can obtain the new characteristic spectrums that dimension is 128*38*38, specific to grasp referring to Fig. 1
As:
The dimension of characteristic spectrum to be transformed is indicated with C*W*H, wherein C indicates that port number, W indicate that wide, H indicates high;
The present invention divides on the port number C of characteristic spectrum channel dimension according to scale factor 4, i.e., C channel exists
Channel dimension obtains the subcharacter that C/4 dimension is 4*W*H and composes by C/4 group subchannel is sequentially divided into from left to right;
Then in each subcharacter spectrum, the value in the region of each 4*1*1 is pressed into sequence from left to right in channel dimension,
It is successively placed on matrixIn r1, r2, r3, the position r4 is thus converted into the feature of 1*2*2 dimension;It is special in every height again
In sign spectrum, the 1*2*2 feature converted is stitched together to obtain according to the relative positional relationship on the wide W and high H of former characteristic spectrum
The subcharacter that one dimension size is 1* (W*2) * (H*2) is composed, finally by C/4 sub- characteristic spectrums by sequentially carrying out group from left to right
It closes, obtains the characteristic spectrum that dimension size is (C/4) * (W*2) * (H*2), that is, complete the transformation of characteristic spectrum scale.
Different from the method amplification characteristic spectrum using deconvolution and linear interpolation, the method for change of scale of the invention does not have
Additional calculating is carried out, the dimensional information in characteristic spectrum has only been subjected to permutation and combination again, has not increased additional meter
Calculation amount, will not reduce the speed of service of object detection method, while can be good at keeping the semantic information of characteristic spectrum.
After the change of scale for completing to compose specific characteristic, by the preceding layer characteristic spectrum of transformed characteristic spectrum and current layer
Carry out the fusion treatment of characteristic spectrum.For example, further feature is composed the shallow-layer of characteristic spectrum and preceding layer that n6 is converted by the present invention
Characteristic spectrum n5 is recombinated to obtain new characteristic spectrum.Further feature composes the width of n6 and height is preceding layer shallow-layer characteristic spectrum n5 wide and high
Half, i.e.,Therefore the dimension of shallow-layer characteristic spectrum n5 is C5n*W5n*H5n, further feature spectrum
The dimension for the characteristic spectrum that n6 is obtained by change of scale is (C6n/4)*(W6n*2)*(H6n* 2), thus the width of two groups of characteristic spectrums and
It is high the same, therefore the present invention is cascaded in this dimension of the channel of characteristic spectrum to get to new characteristic spectrum, dimension is (C5n+
C6n/4)*W5n*H5n, new characteristic spectrum has merged more semantic informations, therefore has better table for the information of Small object
Sign.The present invention carries out position and the classification of prediction target using new characteristic spectrum.
Then fusion feature to be composed again and carries out object detection process, any conventional techniques can be used in specific detection mode, this
Invention does not limit this.
Embodiment
It, can be with by the object detection process of the characteristic spectrum to the last layer when carrying out target detection based on SSD method
Obtain position and the classification of target to be detected;In order to promote the detection accuracy to part Small object, the present invention is to the last layer
Characteristic spectrum carries out change of scale processing (dimension C*W*H), realizes the process of refinement to specified Small object, referring to fig. 2, specifically
Processing step includes:
Step S1: on the spectrum channel dimension of characteristic spectrum to be transformed, being divided into C/4 group subchannel for port number C in order,
The subcharacter that C/4 dimension is 4*W*H is obtained to compose;
Step S2: the region of each 4*1*1 is converted into 1*2*2 dimension in sequence in each subcharacter spectrum;
Step S3: the 1*2*2 feature converted in each subcharacter spectrum is combined to obtain according to original relative positional relationship
The subcharacter that one new dimension size is 1* (W*2) * (H*2) is composed;
Step S4: C/4 sub- characteristic spectrums are subjected to cascading in order, obtaining dimension size is (C/4) * (W*2) *
(H*2) characteristic spectrum.
Step S5: to upper one layer of shallow-layer characteristic spectrum of the characteristic spectrum converted, dimension C1*W*H, with C2*W*H table
Show transformed characteristic dimension, cascading is carried out on the spectrum channel dimension of channel to the two to get new characteristic spectrum is arrived;
Step S6: utilizing new characteristic spectrum, carries out target detection to specified Small object, obtains object detection results.Such as
Predict target position and the classification of specified Small object.
The above description is merely a specific embodiment, any feature disclosed in this specification, except non-specifically
Narration, can be replaced by other alternative features that are equivalent or have similar purpose;Disclosed all features or all sides
Method or in the process the step of, other than mutually exclusive feature and/or step, can be combined in any way.
Claims (1)
1. a kind of method for promoting small target deteection precision based on characteristic spectrum change of scale, characterized in that it comprises the following steps:
Step 1: to image to be detected carry out SSD method object detection process, obtain shallow-layer to deep layer characteristic spectrum, wherein after
The wide and height of one layer of characteristic spectrum is preceding layer feature spectrum width and high half;
Step 2: scale is carried out to the characteristic spectrum for obtaining the non-first floor and carries out conversion process:
Step 201: on the spectrum channel dimension of characteristic spectrum to be transformed, port number C being divided into C/4 group subchannel in order, is obtained
It is composed to the subcharacter that C/4 dimension is 4*W*H, wherein W indicates that the width of characteristic spectrum, H indicate the height of characteristic spectrum;
Step 202: the region of each 4*1*1 being converted into 1*2*2 dimension in sequence in each subcharacter spectrum;
Step 203: combining the 1*2*2 feature converted in each subcharacter spectrum to obtain one according to original relative positional relationship
The subcharacter that new dimension size is 1* (W*2) * (H*2) is composed;
Step 204: C/4 sub- characteristic spectrums being subjected to cascading in order, obtaining dimension size is (C/4) * (W*2) * (H*2)
Characteristic spectrum;
Step 3: to transformed characteristic spectrum, carrying out cascading on the spectrum channel dimension of channel with one layer of characteristic spectrum thereon, obtain
To new characteristic spectrum;
Based on the new characteristic spectrum, target detection is carried out to specified Small object, obtains object detection results.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811010695.6A CN109308458B (en) | 2018-08-31 | 2018-08-31 | Method for improving small target detection precision based on characteristic spectrum scale transformation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811010695.6A CN109308458B (en) | 2018-08-31 | 2018-08-31 | Method for improving small target detection precision based on characteristic spectrum scale transformation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109308458A true CN109308458A (en) | 2019-02-05 |
CN109308458B CN109308458B (en) | 2022-03-15 |
Family
ID=65224699
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811010695.6A Active CN109308458B (en) | 2018-08-31 | 2018-08-31 | Method for improving small target detection precision based on characteristic spectrum scale transformation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109308458B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110619356A (en) * | 2019-08-28 | 2019-12-27 | 电子科技大学 | Target detection method based on regional suggestion attention |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095836A (en) * | 2014-05-14 | 2015-11-25 | 中国科学院深圳先进技术研究院 | Skin texture detecting method and apparatus based on Gabor features |
CN106419909A (en) * | 2016-09-12 | 2017-02-22 | 西安电子科技大学 | Multi-class motion imagination EEG signal classification method based on characteristic recombination and wavelet transformation |
CN107527031A (en) * | 2017-08-22 | 2017-12-29 | 电子科技大学 | A kind of indoor objects detection method based on SSD |
CN107564009A (en) * | 2017-08-30 | 2018-01-09 | 电子科技大学 | Outdoor scene Segmentation of Multi-target method based on depth convolutional neural networks |
CN107689036A (en) * | 2017-09-01 | 2018-02-13 | 深圳市唯特视科技有限公司 | A kind of Real-time image enhancement method based on the bilateral study of depth |
CN107871119A (en) * | 2017-11-01 | 2018-04-03 | 西安电子科技大学 | A kind of object detection method learnt based on object space knowledge and two-stage forecasting |
CN108460403A (en) * | 2018-01-23 | 2018-08-28 | 上海交通大学 | The object detection method and system of multi-scale feature fusion in a kind of image |
-
2018
- 2018-08-31 CN CN201811010695.6A patent/CN109308458B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095836A (en) * | 2014-05-14 | 2015-11-25 | 中国科学院深圳先进技术研究院 | Skin texture detecting method and apparatus based on Gabor features |
CN106419909A (en) * | 2016-09-12 | 2017-02-22 | 西安电子科技大学 | Multi-class motion imagination EEG signal classification method based on characteristic recombination and wavelet transformation |
CN107527031A (en) * | 2017-08-22 | 2017-12-29 | 电子科技大学 | A kind of indoor objects detection method based on SSD |
CN107564009A (en) * | 2017-08-30 | 2018-01-09 | 电子科技大学 | Outdoor scene Segmentation of Multi-target method based on depth convolutional neural networks |
CN107689036A (en) * | 2017-09-01 | 2018-02-13 | 深圳市唯特视科技有限公司 | A kind of Real-time image enhancement method based on the bilateral study of depth |
CN107871119A (en) * | 2017-11-01 | 2018-04-03 | 西安电子科技大学 | A kind of object detection method learnt based on object space knowledge and two-stage forecasting |
CN108460403A (en) * | 2018-01-23 | 2018-08-28 | 上海交通大学 | The object detection method and system of multi-scale feature fusion in a kind of image |
Non-Patent Citations (3)
Title |
---|
GUIMEI CAO等: "Feature-Fused SSD: Fast Detection for Small Objects", 《NINTH INTERNATIONAL CONFERENCE ON GRAPHIC AND IMAGE PROCESSING (ICGIP 2017)》 * |
何慕威: "自然场景下多目标的检测方法研究", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
郭会文: "监控视频中的快速多目标检测与跟踪研究", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110619356A (en) * | 2019-08-28 | 2019-12-27 | 电子科技大学 | Target detection method based on regional suggestion attention |
CN110619356B (en) * | 2019-08-28 | 2023-03-28 | 电子科技大学 | Target detection method based on regional suggestion attention |
Also Published As
Publication number | Publication date |
---|---|
CN109308458B (en) | 2022-03-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhou et al. | TSNet: Three-stream self-attention network for RGB-D indoor semantic segmentation | |
Piergiovanni et al. | Temporal gaussian mixture layer for videos | |
Zhang et al. | MCnet: Multiple context information segmentation network of no-service rail surface defects | |
CN106874688B (en) | Intelligent lead compound based on convolutional neural networks finds method | |
CN109190442B (en) | Rapid face detection method based on deep cascade convolution neural network | |
CN112884712B (en) | Method and related device for classifying defects of display panel | |
Ren et al. | A novel squeeze YOLO-based real-time people counting approach | |
CN107844785A (en) | A kind of method for detecting human face based on size estimation | |
CN107808129A (en) | A kind of facial multi-characteristic points localization method based on single convolutional neural networks | |
CN103646257B (en) | A kind of pedestrian detection and method of counting based on video monitoring image | |
CN109919059B (en) | Salient object detection method based on deep network layering and multi-task training | |
CN112347964B (en) | Behavior detection method and device based on graph network | |
CN105513066B (en) | It is a kind of that the generic object detection method merged with super-pixel is chosen based on seed point | |
CN111598175B (en) | Detector training optimization method based on online difficult case mining mode | |
CN111709331B (en) | Pedestrian re-recognition method based on multi-granularity information interaction model | |
Li et al. | Depthwise nonlocal module for fast salient object detection using a single thread | |
CN111339874A (en) | Single-stage face segmentation method | |
CN114529581A (en) | Multi-target tracking method based on deep learning and multi-task joint training | |
CN108537825A (en) | A kind of method for tracking target based on transfer learning Recurrent networks | |
Liu et al. | A lightweight and accurate recognition framework for signs of X-ray weld images | |
Wu et al. | Learning comprehensive motion representation for action recognition | |
CN109308458A (en) | A method of small target deteection precision is promoted based on characteristic spectrum change of scale | |
Zhai et al. | Automatic pavement crack detection using multimodal features fusion deep neural network | |
Huang et al. | Tao: A trilateral awareness operation for human parsing | |
CN115861610A (en) | Improved CondInst-based sandstone aggregate image segmentation processing method |
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 |