CN105740916A - Image feature coding method and device - Google Patents

Abstract

The application discloses an image feature coding method and device. The method comprises the steps that the normalization features of any one detection image are extracted by utilizing a neural network which is trained in advance; the normalization features of each sample image are extracted by utilizing the neural network; first similarity of the detection image and each sample image is calculated according to the normalization features of the detection image and each sample image; an optimal coding threshold is searched so that binary features obtained according to the optimal coding threshold through coding and second similarity of the detection image and each sample image obtained through calculation are enabled to be closest to the first similarity of the detection image and each sample image; and the normalization features of the detection image are coded into the binary features by utilizing the optimal coding threshold. Accuracy of the image expressed by the binary features can be enhanced and fidelity of information can be enhanced by the image feature coding method and device.

Description

Characteristics of image coded method and device

Technical field

The application belongs to technical field of image processing, specifically, relates to a kind of characteristics of image coded method and device.

Background technology

Along with the development of science and technology, the image procossing such as picture search and image classification obtains everybody attention day by day as a kind of important technology application.

When carrying out the image processing operations such as picture search or image classification, it is necessary to the similarity between expression and the image of characteristics of image is calculated, therefore, how the Pixel-level information of image is converted to characteristics of image, become the key to the issue of high-precision image processing.

Current research shows, the characteristics of image obtained based on the study of the neutral net degree of depth can significantly increase the ability expressing image, but the characteristics of image often dimension being based on the study acquisition of the neutral net degree of depth is relatively big and is floating number, calculates the similarity between them and extremely takies calculating room and time.

A kind of effective solution is exactly that the characteristic in characteristics of image is carried out binary coding, a kind of coded system of the prior art is one layer of hidden layer of addition in the neutral net trained, the neurode of this layer is trained study, allow to the image feature data of extraction is normalized to 0 to 1 data interval, then utilize fixed threshold 0.5 that the characteristic ranging for 0 to 1 obtained is carried out binaryzation, it is encoded to binary data, obtain binary features, such as namely the characteristic more than 0.5 is encoded to binary data 1, namely characteristic less than 0.5 is encoded to binary data 0.

But, existing this binary coding mode, fixed threshold is adopted to make binary data lose partial information in the process expressing normalization data, fidelity of information is relatively low, the accuracy that the image in image data base is carried out the image procossing such as picture search or image classification reduces binary features and expresses the degree of accuracy of image, thus will be affected.

Summary of the invention

In view of this, technical problems to be solved in this application there is provided characteristics of image coded method and device, improves binary features and expresses the degree of accuracy of image, improves fidelity of information.

In order to solve above-mentioned technical problem, this application discloses a kind of characteristics of image coded method, including:

The neutral net utilizing training in advance extracts the normalization characteristic of any one detection image；

Described neutral net is utilized to extract the normalization characteristic of each sample image；

Normalization characteristic according to described detection image and each sample image, calculates the first similarity of described detection image and each sample image；

Find optimum code threshold value so that according to the binary features that described optimum code threshold coding obtains, calculate the second similarity of described detection image and each sample image obtained, and described detection image is closest with the first similarity of each sample image；

Utilize described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

Preferably, described searching optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and the first similarity of described detection image and each sample image is closest to including:

The binary features of described detection image and each sample image will be utilized, calculate the second similarity of described detection image and each sample image obtained as unknown data；

The first similarity and the second similarity that utilize described detection image and each sample image arrange target optimizing function；

From initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

Preferably, described target optimizing function is:

$map=\frac{1}{n}\mathrm{\Σ}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small_iRepresent the second similarity ranking result that i-th sample image obtains according to the second similarity ranking from big to small；

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

Preferably, described optimum code threshold value includes M；Utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰；

Described utilize described optimum code threshold value, the normalization characteristic of described detection image be encoded to binary features and include:

M optimum code threshold value is arranged according to numerical values recited, it is thus achieved that M+1 interval；

According to numerical value order from big to small, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

A kind of characteristics of image coded method, including:

The neutral net utilizing training in advance extracts the normalization characteristic of each detection image obtained from image data base；

Described neutral net is utilized to extract the normalization characteristic of each sample image；

Normalization characteristic according to each detection image and each sample image, calculates first similarity of each detection image and each sample image；

Find optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image；

Utilize described optimum code threshold value, the normalization characteristic of each detection image described is encoded to binary features.

A kind of characteristics of image code device, including:

First extraction module, for utilizing the neutral net of training in advance to extract the normalization characteristic of any one detection image；

Second extraction module, for utilizing described neutral net to extract the normalization characteristic of each sample image；

First computing module, for the normalization characteristic according to described detection image and each sample image, calculates the first similarity of described detection image and each sample image；

First finds module, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；

First coding module, is used for utilizing described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

Preferably, the first searching module includes:

First arranges unit, for the binary features that will utilize described detection image and each sample image, calculates the second similarity of described detection image and each sample image obtained as unknown data；

Second arranges unit, arranges target optimizing function for the first similarity and the second similarity utilizing described detection image and each sample image；

Find unit, for from initial code threshold value, find optimum code threshold value so that described target optimizing function obtains best optimizing result, and described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

Preferably, described target optimizing function is:

$map=\frac{1}{n}\mathrm{\Σ}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small_iRepresent the second similarity ranking result that i-th sample image obtains according to the second similarity ranking from big to small；

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

Preferably, described optimum code threshold value includes M；Utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰；

Described first coding module specifically for:

M optimum code threshold value is arranged according to numerical values recited, it is thus achieved that M+1 interval；

According to numerical value order from big to small, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

A kind of characteristics of image code device, including:

3rd extraction module, for utilizing the neutral net of training in advance to extract the normalization characteristic of each detection image obtained from image data base；

4th extraction module, for utilizing described neutral net to extract the normalization characteristic of each sample image；

Second computing module, for the normalization characteristic according to each detection image and each sample image, calculates first similarity of each detection image and each sample image；

Second finds module, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image；

Second coding module, is used for utilizing described optimum code threshold value, and the normalization characteristic of each detection image described is encoded to binary features.

Compared with prior art, the application can obtain and include techniques below effect:

For any one detection image, sample image is utilized to find the optimum code threshold value that detection image is corresponding, make the binary features utilizing described optimum code threshold coding to obtain, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；Utilize the detection image that optimum code threshold coding obtains binary features can accurate detection of expression image, improve the degree of accuracy of binary features, improve fidelity of information.

Certainly, the arbitrary product implementing the application must be not necessarily required to reach all the above technique effect simultaneously.

Accompanying drawing explanation

Accompanying drawing described herein is used for providing further understanding of the present application, constitutes the part of the application, and the schematic description and description of the application is used for explaining the application, is not intended that the improper restriction to the application.In the accompanying drawings:

Fig. 1 is the flow chart of a kind of one embodiment of characteristics of image coded method of the embodiment of the present application；

Fig. 2 is the flow chart of a kind of characteristics of image another embodiment of coded method of the embodiment of the present application；

Fig. 3 is the flow chart of a kind of characteristics of image another embodiment of coded method of the embodiment of the present application；

Fig. 4 is the structural representation of a kind of one embodiment of characteristics of image code device of the embodiment of the present application；

Fig. 5 is the structural representation of a kind of characteristics of image another embodiment of code device of the embodiment of the present application；

Fig. 6 is the structural representation of a kind of characteristics of image another embodiment of code device of the embodiment of the present application.

Detailed description of the invention

Describe presently filed embodiment in detail below in conjunction with drawings and Examples, thereby the application how application technology means are solved technical problem and reaches the process that realizes of technology effect and can fully understand and implement according to this.

Technical scheme is mainly used in the application scenarios of the image processing operations such as picture search or image classification, by the characteristics of image coding scheme of the application so that provide a kind of degree of accuracy height and the less binary features for representing picture material of resource occupation.

In prior art, it is adopt fixed threshold to quantify that each characteristic in characteristics of image is encoded to a binary data, and the binary features obtained can lose image portion information, thus affecting the accurate of binary features, reduce the ability to express to image, affect image processing operations.

Therefore, inventor has researched and proposed technical scheme through a series of, in the embodiment of the present application, for any one detection image, after the neutral net utilizing training in advance extracts the normalization characteristic of detection image, do not adopt fixed threshold that normalization characteristic is encoded, but utilize described neutral net to extract the normalization characteristic of each sample image；Utilize the normalization characteristic of described detection image and each sample image, calculate the first similarity of described detection image and each sample image；Find optimum code threshold value so that utilize the binary features that described optimum code threshold coding obtains, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；Recycle described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.Due to this optimum code threshold value, make detection image closest with the first similarity of each sample image and the second similarity, namely the binary features obtained is closest with normalization characteristic, now binary features can accurate detection of expression image, improve the degree of accuracy of binary features, improve the fidelity of information of binary features.

Below in conjunction with accompanying drawing, technical scheme is described in detail.

The flow chart of a kind of one embodiment of characteristics of image coded method that Fig. 1 provides for the embodiment of the present application, described method can include following step:

101: utilize the neutral net of training in advance to extract the normalization characteristic of any one detection image.

By neutral net is trained, it is possible to make neutral net can extract the mid-level features of detection image, it is possible to mid-level features to be normalized, obtain the normalization characteristic of correspondence.Wherein, mid-level features is normalized that is to say the data interval that each characteristic in mid-level features is normalized to 0 to 1.

The training of this neutral net can use the image in ImageNet image library exercise supervision formula training, and ImageNet image library is the picture database of an opening, trains picture labeling, is usually utilized to training network or other models.Certainly the image data base of other known image label can also be adopted to be trained.

This neutral net can be such as convolutional neural networks.

It is normalized the characteristics of image extracted before and is the mid-level features that described neutral net is extracted, this neutral net can include multilamellar, one layer of hidden layer was added before output layer, and this hidden layer is trained, for instance study Hash mapping method, and use Sigmoid function as excitation function, the mid-level features of extraction is inputted this hidden layer, can so that the output valve of this hidden layer be determined between 0-1, it is achieved normalization, obtain normalization characteristic.

102: utilize described neutral net to extract the normalization characteristic of each sample image.

103: the normalization characteristic according to described detection image and each sample image, calculate the first similarity of described detection image and each sample image.

104: find optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and described detection image is closest with the first similarity of each sample image.

In the embodiment of the present application, after utilizing the neutral net of training in advance to obtain the normalization characteristic of detection image, then the coding threshold adopted when this normalization characteristic is encoded to binary features, do not adopt fixed threshold 0.5.But utilize sample image to find optimum code threshold value.

Find optimum code threshold value, it is possible to first with detection image and sample image normalization characteristic, calculate the similarity of detection image and sample image, for the differentiation on convenient description, called after the first similarity.

If thus binary features can accurately express normalization characteristic, utilize the binary features of detection image and sample image, calculate the similarity of detection image and the sample image obtained, called after the second similarity, it is necessary to closest with the first similarity of sample image with detection image.

Such that it is able to by finding optimum code threshold value, make the binary features utilizing its coding to obtain, can accurately express normalization characteristic, namely can accurately express picture material so that detection image is closest with the first similarity of each sample image and the second similarity.

105: utilize described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

By finding the optimum code threshold value obtained, namely can utilize this optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features, also be encoded to binary data by each characteristic in normalization characteristic.

For any one detection image, it is possible to find the optimum code threshold value of its correspondence, rather than each detection image all adopts fixed threshold to be encoded, such that it is able to improve binary features to express the ability of image.

In the present embodiment, utilize optimum code threshold value that the normalization characteristic of detection image is encoded, due to this optimum code threshold value, make the first similarity of detection image and each sample image and closest with the second similarity of each sample image, namely the binary features obtained is closest with normalization characteristic, now binary features can accurately express image, improves the degree of accuracy of binary features, improves fidelity of information.

It should be noted that the operation that the optimum code threshold value of step 102～step 104 obtains can first carry out in advance, do not limit and the operation order described in the embodiment of the present application.

Step 105 utilizes described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features, can be that the output layer to described neutral net is trained, utilize the normalization characteristic by described detection image that this output layer realizes to be encoded to binary features.

Owing to, in prior art, being utilize a fixed threshold, each characteristic in normalization characteristic be encoded to a bit binary data.And inventor finds under study for action, in signal quantization process, the number of bits of coding is more high, the expression of initial data is more accurate, therefore, it can each characteristic in normalization characteristic is encoded to many bit binary data, to reduce the information loss in quantizing process further.

In order to realize that each characteristic in normalization characteristic is encoded to many bit binary data, described optimum code threshold value includes multiple, and multiple optimum code threshold values are respectively positioned in the data interval of 0～1, and the numerical value of multiple optimum code threshold value is different.

Therefore as another embodiment, described optimum code threshold value includes M；Utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰, wherein, M is be more than or equal to 1.

As a kind of possible implementation, described utilize described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features and may include that

M optimum code threshold value is arranged from small to large according to numerical value, it is thus achieved that M+1 interval；

According to order from small to large, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

Such as, for M equal to 3, now N is 2, namely utilizes the optimum code threshold value of 3 different numerical value, it is possible to each characteristic in normalization characteristic is encoded to binary data.

Assume 3 optimum code threshold values respectively 0.25,0.5,0.75, then namely obtain four intervals 0～0.25；0.25～0.5；0.5～0.75；0.75～1.In order to facilitate the determination of interval boundary value, it is believed that each interval includes minima, not including maximum, last interval had both included minima and had also included maximum.Can certainly being that each interval does not include minima, including maximum, first interval have both included minima and had also included maximum.

2 bit binary data include: 00,01,10,11.

Then according to order from small to large, the binary data that each interval is corresponding is: 0～0.25 correspondence 00；0.25～0.5 correspondence 01；0.5～0.75 correspondence 10；0.75～1 correspondence 11.

Assuming that in normalization characteristic, a characteristic is 0.4, it is in valued space 0.25～0.5, then namely this characteristic 0.4 is encoded to binary data 01；

Assuming that in normalization characteristic, another characteristic is 0.6, it is in valued space 0.5～0.74, then namely this characteristic 0.6 is encoded to binary data 10.

Preferably, in the present embodiment, M can be equal to 3, N and be 2 so that both can ensure that accuracy, will not increase again encoder complexity and coding difficulty.

Wherein, finding optimum code threshold value can have multiple possible implementation, as a kind of possible implementation, it is possible to obtained by exhaustive mode, first sets initial code threshold value；

Then, present encoding threshold value is utilized to obtain the binary features of described detection image and each sample image；Utilize described binary features, calculate the second similarity obtaining described detection image with each sample image；Judge the first similarity of described detection image and each sample image and whether closest with the second similarity of each sample image；If it is, using described present encoding threshold value as optimum code threshold value；If not, then adjust described present encoding threshold value, continue executing with the step utilizing present encoding threshold value to obtain binary features, until the first similarity of described detection image and each sample image and closest with the second similarity of each sample image, now corresponding present encoding threshold value is optimum code threshold value.

The initial value of present encoding threshold value is initial code threshold value.

In order to avoid endless exhaustive, first similarity of described detection image and each sample image and with the second similarity of each sample image closest to can be when detecting the first similarity of image and each sample image and meeting condition of similarity with the second similarity of each sample image, namely stopping exhaustive, now corresponding coding threshold can as optimum code threshold value.

Wherein, this condition of similarity can have multiple possible implementation, namely whether the first similarity of detection image and each sample image and the second similarity be closest to there being multiple possible judgment mode.

It is likely to implementation, the described binary features obtained according to described optimum code threshold coding as one, calculates the second similarity of the described detection image and each sample image that obtain, and with the first similarity of each sample image closest to may is that

Each sample image carries out the first ranking result of ranking according to the first similarity, carry out in the second ranking result of ranking according to the second similarity with each sample image, whether occurring Y identical sample image before ranking in the sample image of X name, wherein, Y is less than or equal to X.

Namely condition of similarity is each sample image and carries out the first ranking result of ranking according to the first similarity, carry out, in the second ranking result of ranking, whether the sample image of X name occurring before ranking Y identical sample image according to the second similarity with each sample image

It is likely to implementation as another, can be that each sample image carries out the first ranking result of ranking according to the first similarity, carry out in the second ranking result of ranking according to the second similarity with each sample image, whether the meansigma methods of the first ranking result of each sample image and the ratio sum of the second ranking result reaches maximum, or whether more than predetermined threshold etc..

Namely condition of similarity is the meansigma methods of the first ranking result of each sample image and the ratio sum of the second ranking result and whether reaches maximum, or whether more than predetermined threshold.

Wherein, as another embodiment, can according to the detection image the first similarity with each sample image and the second similarity target setting optimizing function, using the first similarity of detection image and each sample image and with the second similarity of each sample image closest to the best optimizing result as this target optimizing function, the coding threshold that best optimizing result is corresponding is optimum code threshold value, best optimizing result namely as above-mentioned condition of similarity.

Therefore, as in figure 2 it is shown, the flow chart of characteristics of image another embodiment of coded method provided for the application, the method can include following step:

201: utilize the neutral net of training in advance to extract the normalization characteristic of any one detection image.

202: utilize described neutral net to extract the normalization characteristic of each sample image.

203: the normalization characteristic according to described detection image and each sample image, calculate the first similarity of described detection image and each sample image.

Step 201～step 203 is identical with the step 101～step 103 of above-described embodiment, does not repeat them here.

204: the binary features of described detection image and each sample image will be utilized, calculate the second similarity of described detection image and each sample image obtained as unknown data.

This binary features obtains according to coding threshold coding immediately, coding threshold namely as unknown data.

205: utilize the first similarity of described detection image and each sample image and the second similarity that target optimizing function is set.

206: from initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

From initial code threshold value, coding obtains the binary features of described detection image and each sample image successively, and calculate the described detection image of acquisition and the second similarity of each sample image, check the second similarity that this initial code threshold value is corresponding, target optimizing function whether is made to obtain best optimizing result, if not, continue coding binary feature after then adjusting initial code threshold value, calculate the second similarity, until target optimizing function obtains best optimizing result, now corresponding coding threshold is optimum code threshold value.

207: utilize described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

This target optimizing function, as a kind of possible implementation can be:

$map=\frac{1}{n}\mathrm{\Σ}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small with detection image_iRepresent the second similarity ranking result that i-th sample image obtains, i=1,2,3 ... n according to the second similarity ranking from big to small with detection image.This ranking result namely refer to ranking ranking, therefore rp_iWith the numerical value that rn is in 1～n.

Described best optimizing result can be such as that described map reaches maximum or more than predetermined threshold.Namely map reaches maximum or coding threshold corresponding during more than predetermined threshold may act as optimum code threshold value.

As alternatively possible implementation, this target optimizing function can be:

Each sample image carries out the first ranking result of ranking according to the first similarity, carries out in the second ranking result of ranking according to the second similarity with each sample image, the number of sample image identical in the sample image of X name before ranking.

Optimum optimizing result, for instance can being occur Y identical sample image before ranking in the sample image of X name, now corresponding coding threshold be optimum code threshold value.

From initial code threshold value, finding optimum code threshold value, it is possible to be according to certain regulation rule, adjust the initial value of coding threshold, making target optimizing function obtain best optimizing result until adjusting the coding threshold obtained.It is sufficiently small that this initial code threshold value can be arranged, then can pass through to incrementally increase the mode of coding threshold, find optimum code threshold value.

Wherein, in order to find optimum code threshold value easily and fast, it is possible to use global optimizing algorithm realizes, and this global optimizing algorithm can be such as simulated annealing, ant group algorithm or genetic algorithm etc..

Can quickly be obtained so that target optimizing function obtains the optimum code threshold value of best optimizing result by global optimizing algorithm.

Accordingly, as another embodiment, described in, from initial code threshold value, find optimum code threshold value so that described target optimizing function obtains best optimizing result and may include that

Utilize simulated annealing, set initial annealing temperature and initial code threshold value；

Obtain new explanation coding threshold；

Judge the target optimizing result whether target optimizing result that described new explanation coding threshold is corresponding is corresponding more than present encoding threshold value；

If it is, accept described new explanation coding threshold and as present encoding threshold value, the step returning described acquisition new explanation coding threshold continues executing with；

If it does not, accept described new explanation coding threshold and as present encoding threshold value using certain probability, the step returning described acquisition new explanation coding threshold continues executing with；

When satisfied annealing end condition, using present encoding threshold value as optimum code threshold value.

Wherein, accepting described new explanation coding threshold using certain probability as present encoding threshold value can be, difference according to the target optimizing result that new explanation coding threshold the is corresponding target optimizing result corresponding with present encoding threshold value, calculate acceptance probability, obtain the random number between 0 to 1, if random number is less than acceptance probability, then accept described new explanation coding threshold, and as present encoding threshold value；If random number is more than acceptance probability, then refusal accepts described new explanation coding threshold.

Wherein, the initial value of present encoding threshold value is initial code threshold value.

Wherein, if new explanation coding threshold is not accepted, present encoding threshold value is constant, returns the step of described acquisition new explanation coding threshold, regains a new explanation coding threshold and continues executing with,

Wherein, annealing end condition can be that new explanation coding threshold is rejected continuously and accepts number of times more than pre-determined number, or annealing temperature is reduced to minimum.

Wherein, the acquisition of new explanation coding threshold can by obtaining present encoding threshold value plus the mode of random number.

Simulated annealing method is used to carry out the global optimizing of dynamic threshold, it is ensured that overall situation probability converges to globally optimal solution when being 1.

Below in conjunction with a practical application, find the process of optimum code threshold value for simulated annealing introduction.

Assuming willAs target optimizing function.N is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small_iRepresent the second similarity ranking result that i-th sample image obtains, i=1,2,3 ... n according to the second similarity ranking from big to small.

Find optimum code threshold value, it is possible to include following step:

A, arranging a sufficiently large numerical value as initial annealing temperature T, arrange initial code threshold value, wherein, initial code threshold value includes three, it is assumed that respectively 0.25,0.5,0.75.Each characteristic in characteristics of image can be encoded to the binary data of 2 figure places by three coding threshold.

New explanation is set and produces function, for instance can be T_m-next=T_m+ rand (m-next), rand (m-next) is randomizer, and random number range can be determined by last iteration result.

Wherein, m=1,2,3, T_mRepresent m-th coding threshold, T_m-nextRepresent the coding threshold after m-th coding threshold adjustment.

Wherein, T_mMeet following constraints once:

0<T₁<T₂<T₃<1

B, will currently solve S input new explanation produce function, produce a new explanation S ' meeting constraints.

Wherein, the current S that solves includes three present encoding threshold values respectively, three present encoding threshold values inputs this new explanation respectively and produces function, three new explanation coding threshold after obtaining three present encoding adjusting thresholds and S '.

The current initial value solving S is three initial code threshold values.

C, utilize described new explanation S ', calculate the target optimizing result map ' of target optimizing function.

D, judge map ' whether more than utilize the current S of solution to calculate to obtain as target optimizing result map

If E is map ' >=map, accept S ' as currently solving S, and return step B and continue executing with；

If map ' is < map, accepting S ' as currently solving S using certain probability, and return step B and continue executing with, otherwise refusal accepts S ' as currently solving S.

Wherein, accept to refer to and determine whether to accept map ' as currently solving S using probability exp (-Δ t '/kT) with certain probability.Wherein ,-Δ t ' is negative value, represents the difference of map ' and map, and k is constant, and along with the reduction of annealing temperature T, exp (-Δ t '/kT) can be gradually lowered.

Obtaining a random number between (0,1), compared with exp (-Δ t '/kT) by random number, if less than exp (-Δ t '/kT), then accept S ' as currently solving S, otherwise namely refusal accepts S ' as currently solving S.

Wherein, annealing temperature T often calculates to obtain a new explanation, namely reduces once, or when calculating the new explanation obtaining predetermined number, reduces once, for instance calculates and obtain 10 new explanations, carries out a temperature and reduces.

F, judge whether meet annealing end condition time, three coding threshold that current solution comprises are optimum code threshold value.

Wherein, when annealing end condition may refer to new explanation refusal number of times more than preset times, three coding threshold comprised by current solution are as optimum code threshold value；Or, when annealing temperature T is reduced to minimum temperature, three coding threshold current solution comprised are as optimum code threshold value.

In the embodiment of the present application, for any one detection image, after the neutral net utilizing training in advance extracts the normalization characteristic of detection image, do not adopt fixed threshold that normalization characteristic is encoded, but utilize described neutral net to extract the normalization characteristic of each sample image；Utilize the normalization characteristic of described detection image and each sample image, calculate the first similarity of described detection image and each sample image；Find optimum code threshold value so that utilize the binary features that described optimum code threshold coding obtains, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；Recycle described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

Due to this optimum code threshold value, make the first similarity of detection image and each sample image and closest with the second similarity of each sample image, namely the binary features obtained is closest with normalization characteristic, now binary features can accurately express image, improve the degree of accuracy of binary features, improve fidelity of information.

And do not adopt the fixed threshold to encode, it is to avoid the shortcoming that algorithm data adaptability that fixed threshold brings is strong not.

And optimum code coding threshold can include multiple, such that it is able to each characteristic is quantified as many bit binary data, to improve the degree of accuracy that binary features is expressed further.

Find optimum code threshold value to be obtained by global optimizing algorithm, it is hereby achieved that globally optimal solution.

Owing in actual applications, there is the situation needing that an image data base is compressed, according to the technical scheme that the application provides, it is possible to be binary features by each picture coding in image data base, thus realizing being effectively compressed of image data base.

Accordingly, as another embodiment, as it is shown on figure 3, the characteristics of image coded method described in the present embodiment can include following step:

301: utilize the neutral net of training in advance to extract the normalization characteristic of each detection image obtained from image data base.

302: utilize described neutral net to extract the normalization characteristic of each sample image.

Wherein, each sample image can obtain from image data base.

As another embodiment, each sample image can carry out sampling obtaining from each detection image, therefore, and the normalization characteristic of each sample image, it is possible to be directly carry out sampling obtaining from the normalization characteristic of each detection image described.Therefore, step 302 can be obtain the normalization characteristic of each sample image from the normalization characteristic of each detection image described.

303: the normalization characteristic according to each detection image and each sample image, calculate first similarity of each detection image and each sample image.

304: find optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image.

305: utilize described optimum code threshold value, the normalization characteristic of each detection image described is encoded to binary features.

Being encoded to binary features by each being detected the normalization characteristic of image, namely can realize the compression of image data base.

In the present embodiment, optimum code threshold value is it is required that the second similarity of each detection image and each sample image, and any one detection image described is closest with the first similarity of each sample image.

As a kind of possible implementation, it is possible to by utilizing the binary features of each detection image and each sample image, calculate the second similarity of each detection image and each sample image obtained as unknown data；

The first similarity and the second similarity that utilize each detection image and each sample image arrange target optimizing function；

From initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of each detection image and each sample image and closest with the second similarity of each sample image.

This target optimizing function can be:

$map=\frac{1}{m}{\mathrm{\&Sigma;map}}_j$

${\mathrm{map}}_j=\frac{1}{n}\mathrm{\&Sigma;}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, and m is detection image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small detecting image with jth_iRepresent the second similarity ranking result that i-th sample image obtains according to the first similarity ranking from big to small detecting image with jth；Ranking result namely refer to ranking ranking, therefore rp_iAnd rn_iIt is the numerical value in 1～n.

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

Wherein, optimum code threshold value can include M and utilize described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, and wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰.Concrete coded system may refer to, shown in above-described embodiment, not repeat them here.

Embodiment illustrated in fig. 3 part unlike the embodiments above is in that, optimum code threshold value acquisition pattern is different, in above-described embodiment, optimum code threshold value makes a detection image closest with the first similarity of each sample image and the second similarity, embodiment illustrated in fig. 3 needs ensure that each detection image is closest with the first similarity of each sample image and the second similarity, when making image data base is compressed, calculate and obtain an optimum code threshold value, improve compression efficiency.

The structural representation of a kind of one embodiment of characteristics of image code device that Fig. 4 the embodiment of the present application provides, this device may include that

First extraction module 401, for utilizing the neutral net of training in advance to extract the normalization characteristic of any one detection image.

By neutral net is trained, it is possible to make neutral net can extract the mid-level features of detection image, it is possible to mid-level features to be normalized, obtain the normalization characteristic of correspondence.

Wherein, mid-level features is normalized that is to say the data interval that each characteristic in mid-level features is normalized to 0 to 1.

Second extraction module 402, for utilizing described neutral net to extract the normalization characteristic of each sample image.

First computing module 403, for the normalization characteristic according to described detection image and each sample image, calculates the first similarity of described detection image and each sample image.

First finds module 404, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image.

First coding module 405, is used for utilizing described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

In the present embodiment, utilize optimum code threshold value that the normalization characteristic of detection image is encoded, due to this optimum code threshold value, make the first similarity of detection image and each sample image and closest with the second similarity of each sample image, namely the binary features obtained is closest with normalization characteristic, now binary features can accurately express image, improves the degree of accuracy of binary features, improves the accuracy of image procossing.

In order to realize that each characteristic in normalization characteristic is encoded to many bit binary data, described optimum code threshold value also includes multiple, and multiple optimum code threshold values are respectively positioned in the data interval of 0～1, and the numerical value of multiple optimum code threshold value is different.

Therefore as another embodiment, described optimum code threshold value includes M；Then utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰, wherein, M is be more than or equal to 1.

Now, as a kind of possible implementation, described first coding module can be specifically for:

M optimum code threshold value is arranged according to numerical values recited, it is thus achieved that M+1 interval；

According to numerical value order from big to small, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

Wherein, find module searching optimum code threshold value and can have multiple possible implementation, for instance, it is possible to obtained by exhaustive mode, first set initial code threshold value；

Then, present encoding threshold value is utilized to obtain the binary features of described detection image and each sample image；Utilize described binary features, calculate the second similarity obtaining described detection image with each sample image；Judge the first similarity of described detection image and each sample image and whether closest with the second similarity of each sample image；If it is, using described present encoding threshold value as optimum code threshold value；If not, then adjust described present encoding threshold value, continue executing with the step utilizing present encoding threshold value to obtain binary features, until the first similarity of described detection image and each sample image and closest with the second similarity of described sample image, now corresponding present encoding threshold value is optimum code threshold value.

The initial value of present encoding threshold value is initial code threshold value.

In order to avoid endless exhaustive, first similarity of described detection image and each sample image and with the second similarity of each sample image closest to can be when detecting the first similarity of image and each sample image and meeting condition of similarity with the second similarity of each sample image, namely stopping exhaustive, now corresponding coding threshold can as optimum code threshold value.

And whether detect the image the first similarity with each sample image with the second similarity closest to there being multiple possible detection mode.

It is likely to implementation, the described binary features obtained according to described optimum code threshold coding as one, calculates the second similarity of the described detection image and each sample image that obtain, and with the first similarity of each sample image closest to may is that

Each sample image carries out the first ranking result of ranking according to the first similarity, carry out in the second ranking result of ranking according to the second similarity with each sample image, whether occurring Y identical sample image before ranking in the sample image of X name, wherein, Y is less than or equal to X.

It is likely to implementation as another, can be that each sample image carries out the first ranking result of ranking according to the first similarity, carry out in the second ranking result of ranking according to the second similarity with each sample image, whether the first ranking result of each sample image is closest with the meansigma methods of the ratio sum of the second ranking result is 1, or whether reach maximum, or whether more than predetermined threshold etc..

Wherein, as another embodiment, can according to the detection image the first similarity with each sample image and the second similarity target setting optimizing function, using the first similarity of detection image and each sample image and with the second similarity of each sample image closest to the best optimizing result as this target optimizing function, the coding threshold that best optimizing result is corresponding is optimum code threshold value.

Therefore, as it is shown in figure 5, be different in that with embodiment illustrated in fig. 4, described first finds module 404 may include that

First arranges unit 501, for the binary features that will utilize described detection image and each sample image, calculates the second similarity of described detection image and each sample image obtained as unknown data；

Second arranges unit 502, arranges target optimizing function for the first similarity and the second similarity utilizing described detection image and each sample image；

Find unit 503, for from initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

From initial code threshold value, coding obtains the binary features of described detection image and each sample image successively, and calculate the described detection image of acquisition and the second similarity of each sample image, check the second similarity that this initial code threshold value is corresponding, target optimizing function whether is made to obtain best optimizing result, if not, continue coding binary feature after then adjusting initial code threshold value, calculate the second similarity, until target optimizing function obtains best optimizing result, now corresponding coding threshold is optimum code threshold value.

Wherein, this target optimizing function, as a kind of possible implementation can be:

$map=\frac{1}{n}\mathrm{\&Sigma;}\frac{rpi}{rni};$

Wherein, n is sample image number, rpi represents the first similarity ranking result that i-th sample image obtains according to the first similarity ranking from big to small, and rni represents the second similarity ranking result that i-th sample image obtains, i=1,2,3 ... n according to the second similarity ranking from big to small.

Described best optimizing result can be that described map reaches maximum or more than predetermined threshold.Namely map reaches maximum or coding threshold corresponding during more than predetermined threshold may act as optimum code threshold value.

As alternatively possible implementation, this target optimizing function can be:

Each sample image carries out the first ranking result of ranking according to the first similarity, carries out, in the second ranking result of ranking, whether occurring the number of identical sample image before ranking in the sample image of X name according to the second similarity with each sample image.

Optimum optimizing result, for instance can being occur Y identical sample image before ranking in the sample image of X name, now corresponding coding threshold be optimum code threshold value.

From initial code threshold value, finding optimum code threshold value, it is possible to be according to certain regulation rule, adjust the initial value of coding threshold, making target optimizing function obtain best optimizing result until adjusting the coding threshold obtained.It is sufficiently small that this initial code threshold value can be arranged, then can pass through to incrementally increase the mode of coding threshold, find optimum code threshold value.

Wherein, in order to find optimum code threshold value easily and fast, it is possible to use global optimizing algorithm realizes, and this global optimizing algorithm can be such as simulated annealing, ant group algorithm or genetic algorithm etc..

Can quickly be obtained so that target optimizing function obtains the optimum code threshold value of best optimizing result by global optimizing algorithm.

Accordingly, as another embodiment, described searching unit 403 can be specifically for:

Utilize simulated annealing, set initial annealing temperature and initial code threshold value；

Obtain new explanation coding threshold；

Judge the target optimizing result whether target optimizing result that described new explanation coding threshold is corresponding is corresponding more than present encoding threshold value；

If it is, accept described new explanation coding threshold and as present encoding threshold value, the step returning described acquisition new explanation coding threshold continues executing with；

If it does not, accept described new explanation coding threshold and as present encoding threshold value using certain probability, the step returning described acquisition new explanation coding threshold continues executing with；

When satisfied annealing end condition, using present encoding threshold value as optimum code threshold value.

Wherein, accepting described new explanation coding threshold using certain probability as present encoding threshold value can be, difference according to the target optimizing result that new explanation coding threshold the is corresponding target optimizing result corresponding with present encoding threshold value, calculate acceptance probability, obtain the random number between 0 to 1, if random number is less than acceptance probability, then accept described new explanation coding threshold, and as present encoding threshold value；If random number is more than acceptance probability, then refusal accepts described new explanation coding threshold.

Wherein, the initial value of present encoding threshold value is initial code threshold value.

Wherein, if new explanation coding threshold is not accepted, present encoding threshold value is constant, returns the step of described acquisition new explanation coding threshold, regains a new explanation coding threshold and continues executing with,

Wherein, annealing end condition can be that new explanation coding threshold is rejected continuously and accepts number of times more than pre-determined number, or annealing temperature is reduced to minimum.

Wherein, the acquisition of new explanation coding threshold can by obtaining present encoding threshold value plus the mode of random number.

The embodiment of the present application, by finding optimum code threshold value, rather than employing fixed threshold, make the binary features utilizing described optimum code threshold coding to obtain, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；So that the binary features obtained is closest with normalization characteristic, now binary features can accurately express image, improves the degree of accuracy of binary features, improves the accuracy of image procossing.And optimum code coding threshold can include multiple, such that it is able to each characteristic is quantified as many bit binary data, to improve the degree of accuracy that binary features is expressed further.

Owing in actual applications, there is the situation needing that an image data base is compressed, according to the technical scheme that the application provides, it is possible to be binary features by each picture coding in image data base, thus realizing being effectively compressed of image data base.

As another embodiment, as shown in Figure 6, the characteristics of image code device that the present embodiment provides may include that

3rd extraction module 601, for utilizing the neutral net of training in advance to extract the normalization characteristic of each detection image obtained from image data base；

4th extraction module 602, for utilizing described neutral net to extract the normalization characteristic of each sample image.

Wherein, each sample image can also obtain from image data base.

As another embodiment, each sample image can be carry out sampling obtaining from each detection image that image data base obtains, therefore, the normalization characteristic of each sample image, it is possible to be directly carry out sampling obtaining from the normalization characteristic of each detection image described.

Second computing module 603, for the normalization characteristic according to each detection image and each sample image, calculates first similarity of each detection image and each sample image；

Second finds module 604, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image；

Second coding module 605, is used for utilizing described optimum code threshold value, and the normalization characteristic of each detection image described is encoded to binary features.

Being encoded to binary features by each being detected the normalization characteristic of image, namely can realize the compression of image data base.

In the present embodiment, optimum code threshold value is it is required that the second similarity of each detection image and each sample image, and any one detection image described is closest with the first similarity of each sample image.

As a kind of possible implementation, it is possible to by utilizing the binary features of each detection image and each sample image, calculate the second similarity of each detection image and each sample image obtained as unknown data；

The first similarity and the second similarity that utilize each detection image and each sample image arrange target optimizing function；

From initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of each detection image and each sample image and closest with the second similarity of each sample image.

This target optimizing function can be:

$map=\frac{1}{m}{\mathrm{\&Sigma;map}}_j;$

${\mathrm{map}}_j=\frac{1}{n}\mathrm{\&Sigma;}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, and m is detection image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small detecting image with jth_iRepresent the second similarity ranking result that i-th sample image obtains according to the first similarity ranking from big to small detecting image with jth；Ranking result namely refer to ranking ranking, therefore rp_iWith the numerical value that rn is in 1～n.

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

Wherein, optimum code threshold value can include M and utilize described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, and wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰.Concrete coded system may refer to, shown in above-described embodiment, not repeat them here.

In a typical configuration, computing equipment includes one or more processor (CPU), input/output interface, network interface and internal memory.

Internal memory potentially includes the forms such as the volatile memory in computer-readable medium, random access memory (RAM) and/or Nonvolatile memory, such as read only memory (ROM) or flash memory (flashRAM).Internal memory is the example of computer-readable medium.

Computer-readable medium includes permanent and impermanency, removable and non-removable media can by any method or technology to realize information storage.Information can be computer-readable instruction, data structure, the module of program or other data.The example of the storage medium of computer includes, but it is not limited to phase transition internal memory (PRAM), static RAM (SRAM), dynamic random access memory (DRAM), other kinds of random access memory (RAM), read only memory (ROM), Electrically Erasable Read Only Memory (EEPROM), fast flash memory bank or other memory techniques, read-only optical disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tape, the storage of tape magnetic rigid disk or other magnetic storage apparatus or any other non-transmission medium, can be used for the information that storage can be accessed by a computing device.According to defining herein, computer-readable medium does not include non-temporary computer readable media (transitorymedia), such as data signal and the carrier wave of modulation.

As employed some vocabulary in the middle of description and claim to censure specific components.Those skilled in the art are it is to be appreciated that hardware manufacturer may call same assembly with different nouns.This specification and claims are not used as distinguishing in the way of assembly by the difference of title, but are used as the criterion distinguished with assembly difference functionally." comprising " as mentioned in the middle of description and claim in the whole text is an open language, therefore should be construed to " comprise but be not limited to "." substantially " referring in receivable range of error, those skilled in the art can solve described technical problem within the scope of certain error, basically reaches described technique effect.Additionally, " coupling " word comprises any directly and indirectly electric property coupling means at this.Therefore, if a first device described in literary composition is coupled to one second device, then represents described first device and can directly be electrically coupled to described second device, or be indirectly electrically coupled to described second device by other devices or the means that couple.Description subsequent descriptions is implement the better embodiment of the application, and right described description is for the purpose of the rule so that the application to be described, is not limited to scope of the present application.The protection domain of the application is when being as the criterion depending on the defined person of claims.

It can further be stated that, term " includes ", " comprising " or its any other variant are intended to comprising of nonexcludability, so that include the commodity of a series of key element or system not only includes those key elements, but also include other key elements being not expressly set out, or also include the key element intrinsic for this commodity or system.When there is no more restriction, statement " including ... " key element limited, it is not excluded that there is also other identical element in the commodity including described key element or system.

Described above illustrate and describes some preferred embodiments of the application, but as previously mentioned, it is to be understood that the application is not limited to form disclosed herein, it is not to be taken as the eliminating to other embodiments, and can be used for other combinations various, amendment and environment, and in application contemplated scope described herein, can be modified by the technology of above-mentioned instruction or association area or knowledge.And the change that those skilled in the art carry out and change are without departing from spirit and scope, then all should in the protection domain of the application claims.

Claims (10)

1. a characteristics of image coded method, it is characterised in that including:

The neutral net utilizing training in advance extracts the normalization characteristic of any one detection image；

Described neutral net is utilized to extract the normalization characteristic of each sample image；

Normalization characteristic according to described detection image and each sample image, calculates the first similarity of described detection image and each sample image；

Find optimum code threshold value so that according to the binary features that described optimum code threshold coding obtains, calculate the second similarity of described detection image and each sample image obtained, and described detection image is closest with the first similarity of each sample image；

Utilize described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

2. method according to claim 1, it is characterized in that, described searching optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and the first similarity of described detection image and each sample image be closest to including:

The binary features of described detection image and each sample image will be utilized, calculate the second similarity of described detection image and each sample image obtained as unknown data；

The first similarity and the second similarity that utilize described detection image and each sample image arrange target optimizing function；

From initial code threshold value, find optimum code threshold value, making described target optimizing function obtain best optimizing result, described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

3. method according to claim 2, it is characterised in that described target optimizing function is:

$map=\frac{1}{n}\mathrm{\&Sigma;}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small_iRepresent the second similarity ranking result that i-th sample image obtains according to the second similarity ranking from big to small；

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

4. the method according to any one of claims 1 to 3, it is characterised in that described optimum code threshold value includes M；Utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰；

Described utilize described optimum code threshold value, the normalization characteristic of described detection image be encoded to binary features and include:

M optimum code threshold value is arranged according to numerical values recited, it is thus achieved that M+1 interval；

According to numerical value order from big to small, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

5. a characteristics of image coded method, it is characterised in that including:

The neutral net utilizing training in advance extracts the normalization characteristic of each detection image obtained from image data base；

Described neutral net is utilized to extract the normalization characteristic of each sample image；

Normalization characteristic according to each detection image and each sample image, calculates first similarity of each detection image and each sample image；

Find optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image；

Utilize described optimum code threshold value, the normalization characteristic of each detection image described is encoded to binary features.

6. a characteristics of image code device, it is characterised in that including:

First extraction module, for utilizing the neutral net of training in advance to extract the normalization characteristic of any one detection image；

Second extraction module, for utilizing described neutral net to extract the normalization characteristic of each sample image；

First computing module, for the normalization characteristic according to described detection image and each sample image, calculates the first similarity of described detection image and each sample image；

First finds module, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of described detection image and each sample image obtained, and closest with the first similarity of each sample image；

First coding module, is used for utilizing described optimum code threshold value, the normalization characteristic of described detection image is encoded to binary features.

7. device according to claim 6, it is characterised in that first finds module includes:

First arranges unit, for the binary features that will utilize described detection image and each sample image, calculates the second similarity of described detection image and each sample image obtained as unknown data；

Second arranges unit, arranges target optimizing function for the first similarity and the second similarity utilizing described detection image and each sample image；

Find unit, for from initial code threshold value, find optimum code threshold value so that described target optimizing function obtains best optimizing result, and described best optimizing result makes the first similarity of described detection image and each sample image and closest with the second similarity of each sample image.

8. device according to claim 7, it is characterised in that described target optimizing function is:

$map=\frac{1}{n}\mathrm{\&Sigma;}\frac{{\mathrm{rp}}_i}{{\mathrm{rn}}_i};$

Wherein, n is sample image number, rp_iRepresent the first similarity ranking result that i-th sample image obtains, rn according to the first similarity ranking from big to small_iRepresent the second similarity ranking result that i-th sample image obtains according to the second similarity ranking from big to small；

Described best optimizing result is that described map reaches maximum or more than predetermined threshold.

9. device according to claim 6, it is characterised in that described optimum code threshold value includes M；Utilizing described optimum code threshold value that each characteristic in described normalization characteristic is encoded to N bit binary data, wherein, described N is be more than or equal to 2；M=2^N-1+2^N-2+…+2⁰；

Described first coding module specifically for:

M optimum code threshold value is arranged according to numerical values recited, it is thus achieved that M+1 interval；

According to numerical value order from big to small, the corresponding N bit binary data of each interval is set；

Each characteristic in the normalization characteristic of described detection image is encoded to the binary data that described characteristic place interval is corresponding.

10. a characteristics of image code device, it is characterised in that including:

3rd extraction module, for utilizing the neutral net of training in advance to extract the normalization characteristic of each detection image obtained from image data base；

4th extraction module, for utilizing described neutral net to extract the normalization characteristic of each sample image；

Second computing module, for the normalization characteristic according to each detection image and each sample image, calculates first similarity of each detection image and each sample image；

Second finds module, for finding optimum code threshold value, make the binary features obtained according to described optimum code threshold coding, calculate the second similarity of each detection image and each sample image obtained, and any one detection image described is closest with the first similarity of each sample image；

Second coding module, is used for utilizing described optimum code threshold value, and the normalization characteristic of each detection image described is encoded to binary features.