CN106845507A - A kind of blocking object detection method based on notice - Google Patents
A kind of blocking object detection method based on notice Download PDFInfo
- Publication number
- CN106845507A CN106845507A CN201510887751.4A CN201510887751A CN106845507A CN 106845507 A CN106845507 A CN 106845507A CN 201510887751 A CN201510887751 A CN 201510887751A CN 106845507 A CN106845507 A CN 106845507A
- Authority
- CN
- China
- Prior art keywords
- network
- input
- picture
- fully
- connected network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
- G06F18/2111—Selection of the most significant subset of features by using evolutionary computational techniques, e.g. genetic algorithms
-
- 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
-
- 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
- G06N3/08—Learning methods
Abstract
A kind of blocking object detection method based on notice, first to a width picture, determines focus, is randomly assigned first, finds zonule interested, then as center, creates three picture blocks of proportional size, then zooms to same size;The three width pictures that will be obtained are input to recurrent neural network, and then Recursive Networks produce two outputs, and an output enters positioning network and produces location information, for determining the interesting target focus in picture again;Another input fully-connected network is used to determine whether this picture block for producing is an object, is if it is fed back to 1, if not being fed back to 0, as the signal of enhancing study.
Description
Technical field
The invention belongs to image information data processing technology field, it is related to image information data treatment technology in deep learning, video
Analysis, the application of object detection field.
Background technology
Depth convolutional neural networks achieve achievement best at present in object monitoring identification field, are designed to process Multidimensional numerical
Data, the attribute of natural sign is utilized using 4 ideas of key:Local connection, shared weights, pond and Multi net voting
The use of layer.
In target detection, candidate mainly is extracted to there may be target area using methods such as selective search,
Then convolutional neural networks are input into, because every pictures will produce thousands of candidates, each candidate are then input into convolutional Neural net
Network, it is significantly slack-off in such speed.
The deepmind team of Google proposes a kind of learning process for imitating human vision, is processed one by one according to notice, learns
Habit process strengthens the feedback of correct study by continuous trial and error, to try to achieve maximal rewards, letter, number is only applied at present
The simple applications such as word identification.
The content of the invention
Present object identification field, the fast-rcnn models that Microsoft proposes, effect is best on image-net data sets, but
It is fixed dimension to need input picture, it is an object of the invention to provide a kind of blocking object detection method based on notice,
Using enhancing study extension faster-rcnn so that input dimension of picture can be with arbitrary size.
Technical scheme is as follows:
A kind of blocking object detection method based on notice, it is characterised in that:
(1) receptor network is built;The central point for being input into picture is defined as (0,0), and (- 1, -1) is orientated in the upper left corner as, adopts
Collect a target of picture, constitute three picture blocks of multiresolution, be input to fully-connected network, while it is directly defeated to position target
Enter a fully-connected network, two outputs of network are input to next fully-connected network simultaneously, exported, as lower step
The input of recurrent neural network;
(2) recurrent neural network is built:Obtain the output of receptor network, and last time Recursive Networks output as input,
Into internal state, then connection fully-connected network is exported, obtain positioning output, impression is again flowed into together with picture is originally inputted
Device network, internal state will also flow into fully-connected network and judge whether containing object, and used as reward, it is straight that circulation carries out this process
Predetermined number of times has all been found or reached to all objects;
(3) receptor network is input to fast-rcnn models by repeatedly circulating many target candidates for obtaining, it is determined that
Classification and position, wherein fast-rcnn models carry out pre-training on image-net.
The present invention is combined based on current main flows such as notice, recurrent neural network, enhancing study, convolutional Neural neutral nets
Model and method, the defect of the faster-rcnn of solution:Input picture needs fixed size, speed to greatly promote, while ensureing
Accuracy rate.
Specific embodiment
Large-sized picture is processed using convolutional neural networks, amount of calculation is very huge, and human eye treatment picture is based on attention
Power, so there is the identification of emphasis, rather than the treatment of the width figure integral part emphasis to seeing.Meanwhile, the study of people
Journey or the continuous correct understanding of reinforcement, the process that mistake is gradually reduced, enhancing study is exactly to imitate this principle.
First to a width picture, focus is determined, be randomly assigned first, zonule interested is found, then as in
The heart, creates three picture blocks of proportional size, then zooms to same size, this namely human eye quick glance.Three for obtaining
Width picture is input to recurrent neural network, and then Recursive Networks produce two outputs, an output to enter positioning network and produce positioning
Information Lt, for determining the interesting target focus in picture again;Another input fully-connected network is used to determine that this is produced
Picture block whether be an object, 1 is if it is fed back to, if not being fed back to 0, as the signal of enhancing study.
Enhancing learning process of the invention is, according to former picture focus, the focus of next time to be selected, so as to generate whether one contain
There is the reward of object, the target of study is to ask for reward to maximize.Reward is much contained thing after circulating several times
The picture block of body, as the input of fast-rcnn models, exports target classification and position coordinates, and fast-rcnn exists before this
Pre-training has been done above image-net.
The inside shape body of recurrent neural network is exactly that a kind of coding schedule of picture block and location information Lt reaches, and two outputs are exactly to increase
The strong randomized policy to be learnt, moves to certain next position location, and the reward for being obtained.
(1) receptor network is built.The central point for being input into picture is defined as (0,0), upper left corner positioning (- 1, -1) collection picture
One target, constitutes three picture blocks of multiresolution, is input to fully-connected network, while positioning target directly inputs one entirely
Two outputs of network are input to next fully-connected network by connection network simultaneously, are exported, used as lower step recurrence god
Through the input of network, i.e. g=Rect (Linear (hg)+Linear (hl)).
(2) recurrent neural network is built.Obtain the output of receptor network, and last time Recursive Networks output as input,
Into internal state, then connection fully-connected network is exported, obtain positioning output, impression is again flowed into together with picture is originally inputted
Device network, internal state will also flow into fully-connected network and judge whether containing object, used as reward, R=sum (r1+r2+r3
+ ...), circulation carries out this process until all objects have all found or reached predetermined number of times.
(3) receptor network is input to fast-rcnn models by repeatedly circulating many target candidates for obtaining, and determines class
Other and position, wherein fast-rcnn models carry out pre-training on image-net.
Claims (1)
1. a kind of blocking object detection method based on notice, it is characterised in that:
(1) receptor network is built;The central point for being input into picture is defined as (0,0), and (- 1, -1) is orientated in the upper left corner as, adopts
Collect a target of picture, constitute three picture blocks of multiresolution, be input to fully-connected network, while it is directly defeated to position target
Enter a fully-connected network, two outputs of network are input to next fully-connected network simultaneously, exported, as lower step
The input of recurrent neural network;
(2) recurrent neural network is built:Obtain the output of receptor network, and last time Recursive Networks output as input,
Into internal state, then connection fully-connected network is exported, obtain positioning output, impression is again flowed into together with picture is originally inputted
Device network, internal state will also flow into fully-connected network and judge whether containing object, and used as reward, it is straight that circulation carries out this process
Predetermined number of times has all been found or reached to all objects;
(3) receptor network is input to fast-rcnn models by repeatedly circulating many target candidates for obtaining, it is determined that
Classification and position.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510887751.4A CN106845507A (en) | 2015-12-07 | 2015-12-07 | A kind of blocking object detection method based on notice |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510887751.4A CN106845507A (en) | 2015-12-07 | 2015-12-07 | A kind of blocking object detection method based on notice |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106845507A true CN106845507A (en) | 2017-06-13 |
Family
ID=59151137
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510887751.4A Pending CN106845507A (en) | 2015-12-07 | 2015-12-07 | A kind of blocking object detection method based on notice |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106845507A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108230294A (en) * | 2017-06-14 | 2018-06-29 | 北京市商汤科技开发有限公司 | Image detecting method, device, electronic equipment and storage medium |
CN109117876A (en) * | 2018-07-26 | 2019-01-01 | 成都快眼科技有限公司 | A kind of dense small target deteection model building method, model and detection method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5706400A (en) * | 1995-03-08 | 1998-01-06 | Nec Research Institute, Inc. | Fault-tolerant implementation of finite-state automata in recurrent neural networks |
US20070011119A1 (en) * | 2005-05-07 | 2007-01-11 | Thaler Stephen L | Device for the autonomous bootstrapping of useful information |
CN104463172A (en) * | 2014-12-09 | 2015-03-25 | 中国科学院重庆绿色智能技术研究院 | Face feature extraction method based on face feature point shape drive depth model |
-
2015
- 2015-12-07 CN CN201510887751.4A patent/CN106845507A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5706400A (en) * | 1995-03-08 | 1998-01-06 | Nec Research Institute, Inc. | Fault-tolerant implementation of finite-state automata in recurrent neural networks |
US20070011119A1 (en) * | 2005-05-07 | 2007-01-11 | Thaler Stephen L | Device for the autonomous bootstrapping of useful information |
CN104463172A (en) * | 2014-12-09 | 2015-03-25 | 中国科学院重庆绿色智能技术研究院 | Face feature extraction method based on face feature point shape drive depth model |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108230294A (en) * | 2017-06-14 | 2018-06-29 | 北京市商汤科技开发有限公司 | Image detecting method, device, electronic equipment and storage medium |
WO2018228336A1 (en) * | 2017-06-14 | 2018-12-20 | 北京市商汤科技开发有限公司 | Image detection method and apparatus, electronic device, and storage medium |
CN108230294B (en) * | 2017-06-14 | 2020-09-29 | 北京市商汤科技开发有限公司 | Image detection method, image detection device, electronic equipment and storage medium |
US11200416B2 (en) | 2017-06-14 | 2021-12-14 | Beijing Sensetime Technology Development Co., Ltd | Methods and apparatuses for image detection, electronic devices and storage media |
CN109117876A (en) * | 2018-07-26 | 2019-01-01 | 成都快眼科技有限公司 | A kind of dense small target deteection model building method, model and detection method |
CN109117876B (en) * | 2018-07-26 | 2022-11-04 | 成都快眼科技有限公司 | Dense small target detection model construction method, dense small target detection model and dense small target detection method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109543606B (en) | Human face recognition method with attention mechanism | |
Ale et al. | Deep learning based plant disease detection for smart agriculture | |
CN107679522B (en) | Multi-stream LSTM-based action identification method | |
WO2022001623A1 (en) | Image processing method and apparatus based on artificial intelligence, and device and storage medium | |
Chung et al. | Comparative analysis of skeleton-based human pose estimation | |
CN107506712A (en) | Method for distinguishing is known in a kind of human behavior based on 3D depth convolutional networks | |
Yang et al. | Emotion recognition for multiple context awareness | |
CN110363210A (en) | A kind of training method and server of image, semantic parted pattern | |
CN108572733B (en) | A kind of eye movement behavior visual search target prediction method based on condition random field | |
CN105808610A (en) | Internet picture filtering method and device | |
CN107451607A (en) | A kind of personal identification method of the typical character based on deep learning | |
Wang et al. | Learning multimodal word representation via dynamic fusion methods | |
CN109377441A (en) | Tongue with privacy protection function is as acquisition method and system | |
CN109740585A (en) | A kind of text positioning method and device | |
CN109740752A (en) | Depth model training method and device, electronic equipment and storage medium | |
WO2020033345A1 (en) | Action localization using relational features | |
CN108596256A (en) | One kind being based on RGB-D object identification grader building methods | |
CN106355574A (en) | Intra-abdominal adipose tissue segmentation method based on deep learning | |
CN106845507A (en) | A kind of blocking object detection method based on notice | |
Xiao et al. | Novel convolutions for semantic segmentation of remote sensing images | |
Aliakbarian et al. | Deep action-and context-aware sequence learning for activity recognition and anticipation | |
Liang | Improved EfficientDET algorithm for basketball players’ upper limb movement trajectory recognition | |
CN111652168B (en) | Group detection method, device, equipment and storage medium based on artificial intelligence | |
Wan et al. | CellT-Net: A composite transformer method for 2-D cell instance segmentation | |
CN105046193B (en) | A kind of human motion recognition method based on fusion rarefaction representation matrix |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170613 |