CN110516091A - A kind of image data retrieval method - Google Patents
A kind of image data retrieval method Download PDFInfo
- Publication number
- CN110516091A CN110516091A CN201910815340.2A CN201910815340A CN110516091A CN 110516091 A CN110516091 A CN 110516091A CN 201910815340 A CN201910815340 A CN 201910815340A CN 110516091 A CN110516091 A CN 110516091A
- Authority
- CN
- China
- Prior art keywords
- layer
- layers
- network
- pool
- conv
- 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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/53—Querying
-
- 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
- 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 image data retrieval method, belongs to Computer Science and Technology field.In order to solve the problems, such as some existing models to need to rely on caused by triple is trained that training process is complicated and the serious problem not high by precision caused by the influence of gradient problem of existing prototype network depth.A convolution sub-network and a Hash sub-network are connected after INPUT layers of the present invention；Hash sub-network includes sliced layer, connects layer unit entirely, normalization layer unit, activation layer unit, merges layer and thresholding layer；Then the feature scribing that sliced layer exports convolution sub-network respectively enters each full articulamentum and handles respectively, then handles after normalizing layer using activation primitive respectively；It is laminated simultaneously that feature after activation primitive enters merging；Into thresholding layer, the thresholding number of plies exports Hash codes；For retrieval image, the retrieval of destination image data is carried out using trained depth Hash model.The present invention is suitable for image data retrieval.
Description
Technical field
The present invention relates to a kind of search methods of image, belong to Computer Science and Technology field.
Background technique
With the development of computer technology, the especially development of artificial intelligence field technology.Face recognition, image procossing,
Each practical application area such as natural language processing starts with deep neural network all to realize, or even has relied in depth
Neural network.For nearest neighbor search problem, need then to need to find in space according to a given inquiry (query)
The point nearest from it, image retrieval are also typically used as a kind of neighbor search also to handle.
In recent years, image retrieval had obtained very extensive research, such as the teacher Pan Yan study group of Zhongshan University and Yan Shui
Cooperate at teacher, the paper delivered on American Association for Artificial Inte annual meeting (AAAI 2014) proposes a kind of entitled CNNH
The method of (Convolutional Neural Network Hashing), before the depth hash algorithm based on CNN has been shifted onto
Platform.Later teacher Yan Shuicheng has used the network knot that more Network in Network is firmly got than the network in CNNH
Structure, referred to as NINH (NIN Hashing) or DNNH (Deep Neural Network Hashing).Its Web vector graphic three is opened
The triple of image construction is trained.Good effect is achieved, but training process needs to spend more training image
The time of processing, and operate comparatively laborious.Other scientific research personnel ask CNN and depth hash algorithm to solve picture search
The technology of topic also expands a large amount of research and experiment, and the deep neural network that everybody builds is different, the effect of acquirement
Respectively there is superiority and inferiority.There is more classical model structure during this, but the model of various classics is also due to the scope of application
With itself existing for it is specific continuously improved and updated, but some improve not only cannot more excellent or even effect can be deteriorated, this is all
Caused by being the characteristic having due to deep neural network itself, because the structure of deep neural network can drastically influence place
Rationality energy.Moreover, relaxation problem, the influence of loss function (or objective function), training process gradient descent procedures
Etc. can all seriously affect the process performance of deep neural network.
Current deep neural network has been able to obtain preferable search result, but how to determine and query
More matched search result still needs to be further improved.
Summary of the invention
The present invention in order to solve some existing models need to rely on triple be trained caused by training process it is complicated
The problem of and existing prototype network depth seriously influenced by gradient problem caused by the not high problem of precision.
A kind of image data retrieval method, comprising the following steps:
Depth Hash model is as follows:
A convolution sub-network is connected after INPUT layers, connects a Hash sub-network after convolution sub-network；The Hash
Sub-network includes sliced layer, connects layer unit entirely, normalization layer unit, activation layer unit, merges layer and thresholding layer；By convolution
The feature of sub-network output is denoted as x, and feature x is divided into n piece by sliced layer, and every feature is xi, i=1,2 ... n；
It is m/n per a piece of intrinsic dimensionality for including, m is characterized the dimension of x；
The full connection layer unit of Hash sub-network has n full articulamentums, and each full articulamentum respectively corresponds one x of processingi,
Full connection processing is expressed as fi=Wixi+bi, wherein WiFor the weight matrix of i-th of full articulamentum, biFor corresponding biasing；
Normalization layer unit has n normalization layer, each one f of normalization layer alignment processingi, indicate after normalized
For gi, range is [- 1,1],
Activation layer unit has n activation primitive layer, each one g of activation primitive layer alignment processingi, activation primitive isWherein β is smooth control parameter, e natural constant；giIt is q by activation primitive postscripti；Active coating is compiled
Code enters Softmax classifier；
All qiInto layer is merged, merge layer for qiMerge into a n-dimensional vector q=(q1,q2,…,qn)T；
Subsequently into thresholding layer, thresholding function are as follows:The thresholding number of plies exports Hash
Code；
The loss function of depth Hash model is
V=(v (q1),v(q2),…,v(qn))TFor the Hash codes of thresholding layer output；Indicate the L1 norm of q-1
2 powers；L is weight factor；
Depth Hash model obtains trained depth Hash model by training set after training；
For retrieval image, the retrieval of destination image data is carried out using trained depth Hash model.
Further, in the normalization layerfjFor f1To fnThe middle maximum feature of mould.
The beneficial effects of the present invention are:
Training process of the invention can directly use individual image, and be trained without relying on triple, energy
The time largely handled training image is enough saved, and training process is simple, is easily achieved；And it can save a large amount of
Time for being directed to training set itself and being handled, i.e. training set can be used directly, time saving and energy saving.
The present invention normalizes the normalization operation in layer unit, avoids because double tangent activation primitives exist due to saturation
The problem of gradient of generation disappears can not only guarantee that the present invention can have neural network deep enough and build ability in this way,
And it can also ensure that can there is relatively good optimizing effect in gradient derivation process.
The loss function of depth Hash model of the present invention makes the present invention have preferably study and effect of optimization.
Detailed description of the invention
Fig. 1 is depth Hash model structure schematic diagram.
Specific embodiment
Specific embodiment 1:
Present embodiment is a kind of image data retrieval method, specifically includes the following steps:
As shown in Figure 1, depth Hash model is as follows:
A convolution sub-network is connected after INPUT layers, connects a Hash sub-network after convolution sub-network；The Hash
Sub-network includes sliced layer, connects layer unit entirely, normalization layer unit, activation layer unit, merges layer and thresholding layer；It is described
INPUT layers are input layer.
The feature that convolution sub-network exports is denoted as x, feature x is divided into n piece by sliced layer, and every feature is xi, i=1,
2,…n；It is m/n per a piece of intrinsic dimensionality for including, m is characterized the dimension of x；It is integer that strict control m/n is wanted in the present invention, can
To utilize FC layers of controlling feature dimension of the last layer in convolution sub-network, so that m/n is integer；Here after operation will affect
Continue normalized process, to influence active coating and then influence the effect of overall model；
The full connection layer unit of Hash sub-network has n full articulamentums, and each full articulamentum respectively corresponds one x of processingi,
Full connection processing is expressed as fi=Wixi+bi, wherein WiFor the weight matrix of i-th of full articulamentum, biFor corresponding biasing；
Normalization layer unit has n normalization layer, each one f of normalization layer alignment processingi, indicate after normalized
For gi, range is [- 1,1],
In some embodiments,fjFor f1To fnThe middle maximum feature of mould；It can guarantee characteristic symbol not in this way
Become, and in the normalized section to [- 1,1], to avoid because double tangent activation primitives have the ladder generated due to saturation
The problem of degree disappears can not only guarantee that the present invention can have neural network deep enough and build ability in this way, and can also
Enough guarantee there can be relatively good optimizing effect in gradient derivation process.
Activation layer unit has n activation primitive layer, each one g of activation primitive layer alignment processingi, activation primitive isWherein β is smooth control parameter, e natural constant；giIt is q by activation primitive postscripti；Active coating is compiled
Code enters Softmax classifier；
All qiInto layer is merged, merge layer for qiMerge into a n-dimensional vector q=(q1,q2,…,qn)T；
Subsequently into thresholding layer, thresholding function are as follows:
The thresholding number of plies exports Hash codes；
The loss function of depth Hash model is
V=(v (q1),v(q2),…,v(qn))TFor the Hash codes of thresholding layer output；Indicate the L1 norm of q-1
2 powers；L is weight factor；In the present inventionIt is corresponding when coboundary minimum for the coboundary of quantization error loss
Target in boundary also can be minimum, so that the present invention has preferably study and effect of optimization.
Depth Hash model obtains trained depth Hash model by training set after training；
For retrieval image, the retrieval of destination image data is carried out using trained depth Hash model.
Specific embodiment 2:
Present embodiment is a kind of image data retrieval method, and convolution sub-network described in present embodiment includes at least four
It is Conv layers a, at least four Pool layers and at least two FC layers；Conv layers of expression convolutional layer, Pool layers of expression pond layer, FC layers
Indicate full articulamentum；
In some embodiments, the convolution sub-network includes there are four at Conv layers and four Pool layers, and structure is as follows:
First Conv layers of Pool layers of connection the first, the first Pool layers of Conv layers of connection the 2nd, the 2nd Conv layers of connection second
Pool layers, the 2nd Pool layers of Conv layers of connection the 3rd, the 3rd Conv layers of Pool layers of connection the 3rd, the 3rd Pool layers of connection the 4th
Conv layers, the 4th Conv layers of Pool layers of connection the 4th；Four Pool layers of primary two FC layers of connection.Each Conv layers is made using RELU
For activation primitive.
The loss function of existing model determines that its model depth cannot be too deep.Due to loss function form of the invention
The convolution sub-network and Hash sub-network suitable for deeper can be allowed the invention to, will not be influenced because of gradient problem
The depth of model；But also can reduce influence caused by relaxation, so that the effect of model is more preferable.The actually present invention
Need four layers of convolution-pondization that can obtain the effect well retrieved, whole network model especially at this time is simple, so that
Whole arithmetic speed faster, it is more efficient, nor will appear gradient and disappear or the problem of gradient is exploded, so that whole
The Model Practical of body is stronger.Certainly model of the invention can be adapted for the network model of deeper completely, and with net
The intensification effect of network model can further increase.
Other steps and depth Hash model structure are same as the specific embodiment one.
Embodiment
Tested according to the scheme of specific embodiment one and two, wherein convolution sub-network include there are four Conv layer with
At four Pool layers, structure is as follows:
First Conv layers of Pool layers of connection the first, the first Pool layers of Conv layers of connection the 2nd, the 2nd Conv layers of connection second
Pool layers, the 2nd Pool layers of Conv layers of connection the 3rd, the 3rd Conv layers of Pool layers of connection the 3rd, the 3rd Pool layers of connection the 4th
Conv layers, the 4th Conv layers of Pool layers of connection the 4th；Four Pool layers of primary two FC layers of connection.Each Conv layers is made using RELU
For activation primitive.
The weight factor l=0.35.
It is tested with CIFAR-10 data set, experiment extracts 1000 images from one kind every in CIFAR-10 data set
Sample is as experimental data, wherein 900 image patterns are as training data, remaining 100 image patterns are as test data.
The present invention and existing classical way compare explanation, are shown in Table 1.Wherein the data of classical way are existing
Data disclosed in file, data of the invention are the result that actual experiment obtains.
1 mAP Comparative result of table
By table 1 it can be seen that mAP result of the invention is much higher than other several methods, here it may also be noticed that this hair
It is bright be only with simple four groups Conv-Pool layer add that two two FC layers of convolution sub-network obtains as a result, if using
Classical Conv-Pool layers of group or deeper neural network, effect of the invention can be further improved.Simultaneously because this hair
Bright Hash sub-network structure and loss function determine that the present invention can be adapted for deeper neural network completely, and will not be by
To the influence of gradient problem, while influence caused by relaxation can also be controlled to a certain extent, so the present invention selects deeper
Neural network, precision can be higher.And if may also be noticed that training set data be more often trained, the present invention
MAP value can also further increase.
Even if having been had using simplest four groups Conv-Pool layers plus two FC layers the convolution sub-network present invention
Good effect, in this case, since structure of the invention is simple, so operation efficiency is higher, compared to by simple five groups
Conv-Pool layers plus two FC layers convolution sub-networks, efficiency of the invention can be improved 10% or so.
Claims (5)
1. a kind of image data retrieval method, which comprises the following steps:
Depth Hash model is as follows:
A convolution sub-network is connected after INPUT layers, connects a Hash sub-network after convolution sub-network；The Hash subnet
Network includes sliced layer, connects layer unit entirely, normalization layer unit, activation layer unit, merges layer and thresholding layer；
The feature that convolution sub-network exports is denoted as x, feature x is divided into n piece by sliced layer, and every feature is xi, i=1,2 ...
n；It is m/n per a piece of intrinsic dimensionality for including, m is characterized the dimension of x；
The full connection layer unit of Hash sub-network has n full articulamentums, and each full articulamentum respectively corresponds one x of processingi, Quan Lian
It connects processing and is expressed as fi=Wixi+bi, wherein WiFor the weight matrix of i-th of full articulamentum, biFor corresponding biasing；
Normalization layer unit has n normalization layer, each one f of normalization layer alignment processingi, g is expressed as after normalizedi,
Range is [- 1,1],
Activation layer unit has n activation primitive layer, each one g of activation primitive layer alignment processingi, activation primitive isWherein β is smooth control parameter, e natural constant；giIt is q by activation primitive postscripti；Active coating is compiled
Code enters Softmax classifier；
All qiInto layer is merged, merge layer for qiMerge into a n-dimensional vector q=(q1,q2,…,qn)T；
Subsequently into thresholding layer, thresholding function are as follows:
The thresholding number of plies exports Hash codes；
The loss function of depth Hash model is
V=(v (q1),v(q2),…,v(qn))TFor the Hash codes of thresholding layer output；Indicate 2 times of the L1 norm of q-1
Side；L is weight factor；
Depth Hash model obtains trained depth Hash model by training set after training；
For retrieval image, the retrieval of destination image data is carried out using trained depth Hash model.
2. a kind of image data retrieval method according to claim 1, which is characterized in that in the normalization layerfjFor f1To fnThe middle maximum feature of mould.
3. a kind of image data retrieval method according to claim 1, which is characterized in that the weight factor l=0.35.
4. a kind of image data retrieval method according to claim 1,2 or 3, which is characterized in that the convolution subnet
Network includes at least four Conv layers, at least four Pool layers and at least two FC layers；Conv layers of expression convolutional layer, Pool layers of expression
Pond layer, FC layers indicate full articulamentum.
5. a kind of image data retrieval method according to claim 4, it is characterised in that；The convolution sub-network includes
At four Conv layers and four Pool layers, structure is as follows:
First Conv layers of Pool layers of connection the first, the first Pool layers of Conv layers of connection the 2nd, the 2nd Conv layers of the 2nd Pool of connection
Layer, the 2nd Pool layers of Conv layers of connection the 3rd, the 3rd Conv layers of Pool layers of connection the 3rd, the 3rd Pool layers of the 4th Conv of connection
Layer, the 4th Conv layers of Pool layers of connection the 4th；Four Pool layers of primary two FC layers of connection.Each Conv layers using RELU as sharp
Function living.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910815340.2A CN110516091A (en) | 2019-08-30 | 2019-08-30 | A kind of image data retrieval method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910815340.2A CN110516091A (en) | 2019-08-30 | 2019-08-30 | A kind of image data retrieval method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110516091A true CN110516091A (en) | 2019-11-29 |
Family
ID=68628547
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910815340.2A Pending CN110516091A (en) | 2019-08-30 | 2019-08-30 | A kind of image data retrieval method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110516091A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512289A (en) * | 2015-12-07 | 2016-04-20 | 郑州金惠计算机系统工程有限公司 | Image retrieval method based on deep learning and Hash |
| CN107092661A (en) * | 2017-03-28 | 2017-08-25 | 桂林明辉信息科技有限公司 | A kind of image search method based on depth convolutional neural networks |
| CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
-
2019
- 2019-08-30 CN CN201910815340.2A patent/CN110516091A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105512289A (en) * | 2015-12-07 | 2016-04-20 | 郑州金惠计算机系统工程有限公司 | Image retrieval method based on deep learning and Hash |
| CN107092661A (en) * | 2017-03-28 | 2017-08-25 | 桂林明辉信息科技有限公司 | A kind of image search method based on depth convolutional neural networks |
| CN109165306A (en) * | 2018-08-09 | 2019-01-08 | 长沙理工大学 | Image search method based on the study of multitask Hash |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109142946A (en) | Transformer fault detection method based on ant group algorithm optimization random forest | |
| CN109741318A (en) | The real-time detection method of single phase multiple dimensioned specific objective based on effective receptive field | |
| CN110851566A (en) | Improved differentiable network structure searching method | |
| CN110321566A (en) | Chinese name entity recognition method, device, computer equipment and storage medium | |
| CN108021658A (en) | A kind of big data intelligent search method and system based on whale optimization algorithm | |
| CN107403191A (en) | A kind of semi-supervised learning machine sorting technique that transfinites with depth structure | |
| CN112732864B (en) | Document retrieval method based on dense pseudo query vector representation | |
| CN110287282A (en) | The Intelligent dialogue systems response method and Intelligent dialogue system of calculation are assessed based on tree | |
| Cao et al. | Improved crow search algorithm optimized extreme learning machine based on classification algorithm and application | |
| CN106096052A (en) | A kind of consumer's clustering method towards wechat marketing | |
| CN104731811B (en) | A kind of clustering information evolution analysis method towards extensive dynamic short text | |
| CN108399268A (en) | A kind of increment type isomery figure clustering method based on game theory | |
| CN107578101B (en) | Data stream load prediction method | |
| Cheng et al. | Remote sensing image classification based on optimized support vector machine | |
| Jothi et al. | Soft set based quick reduct approach for unsupervised feature selection | |
| CN116960991B (en) | Probability-oriented power load prediction method based on graph convolution network model | |
| CN110516091A (en) | A kind of image data retrieval method | |
| CN103440332B (en) | A kind of image search method strengthening expression based on relational matrix regularization | |
| CN109886454A (en) | A kind of fresh water environment wawter bloom prediction technique based on self-organizing deepness belief network and Method Using Relevance Vector Machine | |
| Qiu et al. | Feature selection using a set based discrete particle swarm optimization and a novel feature subset evaluation criterion | |
| JP2011039977A (en) | Word clustering device, method and program, and recording medium storing the program | |
| CN101339615B (en) | Method of image segmentation based on similar matrix approximation | |
| Zhai et al. | Deep product quantization for large-scale image retrieval | |
| Kurasova et al. | Integration of the self-organizing map and neural gas with multidimensional scaling | |
| Yue et al. | Study on the deep neural network of intelligent image detection and the improvement of elastic momentum on image recognition |
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 |