Embodiment
Here will be described exemplary embodiment in detail, its sample table shows in the accompanying drawings.When description below relates to accompanying drawing, unless otherwise indicated, the same numbers in different accompanying drawing represents same or analogous key element.Embodiment described in following exemplary embodiment does not represent all embodiments consistent with the disclosure.On the contrary, they only with as in appended claims describe in detail, the example of apparatus and method that aspects more of the present disclosure are consistent.
Fig. 1 is the process flow diagram of the detection method of a kind of image content according to an exemplary embodiment, and the present embodiment is configured to illustrate in the pick-up unit of image content with the detection method of this image content.As shown in Figure 1, the detection method of this image content comprises following several step:
In step S101, obtain the linear transformation function of the full articulamentum of picture to be detected and convolutional neural networks.
In embodiment of the present disclosure, picture to be detected is the picture needing Detection of content, and such as, picture to be detected can be the picture that user downloads on webpage, or also can be user to be taken pictures the picture obtained by video camera picture to be detected.
At present in technical field of image processing, user can utilize computing machine to process image, analyze and understand, and with the target of different mode various in recognition image and object, such as, can identify the position of object and the category attribute of object in picture to be detected.
As shown in Figure 2, Fig. 2 is the position of object in a kind of image content and the category attribute schematic diagram of object, and 21 in Fig. 2 is for identifying the position of object, and 22 for identifying the category attribute of object.
Convolutional neural networks is a kind of feedforward neural network, and its artificial neuron can respond the unit in partial coverage.
Alternatively, can detect the position of object and the category attribute of object in picture by the method based on convolutional neural networks, convolutional neural networks comprises convolutional layer, pond layer, full articulamentum, and activation function.
In the disclosed embodiments, process is optimized to the computation process of the full articulamentum in convolutional neural networks, to promote the detection speed of image content.
Alternatively, the linear transformation function of the full articulamentum of convolutional neural networks can show with following formula table:
B＝A×W+b；
Wherein, A is input matrix, and B is output matrix, and W is weight matrix, and b is bias term.
In step s 102, svd process is done to the weight matrix in the linear transformation function of full articulamentum, obtain the weight matrix after processing.
Wherein, svd (SingularValueDecomposition, SVD) be a kind of matrix disassembling method in linear algebra, SVD decomposable process is such as: suppose that M is m × n rank matrixes, element wherein all belongs to territory K, K can be real number field, or complex field, then there is a decomposition and make:
M＝C×Ε×DT；
Wherein, Matrix C is m × r rank matrixes, and matrix Ε is positive semidefinite r × r rank diagonal matrix, and matrix D, be n × r rank matrixes, this decomposition be called the svd of M, the element on matrix Ε diagonal line is the singular value of M.
Such as, SVD resolution process can be carried out to the weight matrix W in the full articulamentum linear transformation function got in step S101, to obtain the weight matrix after resolution process.
Alternatively, the matrix after decomposition comprises left singular vector matrix, diagonal matrix, and right singular vector matrix, and wherein, left singular vector matrix comprises the left singular vector of weight matrix W, and right singular vector matrix comprises the right singular vector of weight matrix W.
Such as, the weight matrix W after decomposition can be expressed as:
W＝U×S×VT；
Wherein, U is the left singular vector matrix on m × r rank, and S is the diagonal matrix on r × r rank, and V is the right singular vector matrix on n × r rank.
In embodiment of the present disclosure, svd process is done to the weight matrix in the linear transformation function of full articulamentum, obtains the weight matrix after processing, comprising: the singular value obtaining weight matrix, and the singular value of weight matrix is sorted, obtain the singular value after sorting; Svd is done to the weight matrix in the linear transformation function of full articulamentum, obtains the matrix after decomposing; And in singular value after sequence, choose the singular value of predetermined number, and obtain the weight matrix after process according to the matrix after the singular value Sum decomposition of predetermined number.
Such as, the weight matrix W ' after process can be expressed as:
W’＝U’×S’×V’T；
Wherein, U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and k is the number of the singular value of predetermined number.
In step s 103, obtain target linear transformation function according to the weight matrix after process, and according to target linear transformation function, picture to be detected is detected, to obtain the content in picture to be detected.
Wherein, the content in picture to be detected comprises the position of object and the category attribute of object.As shown in Figure 2.Obtain the content in picture to be detected, namely obtain position and the category attribute of the object in picture to be detected.
Alternatively, process can be optimized, to detect picture to be detected according to the full articulamentum linear transformation function after optimization process according to the linear transformation function of weight matrix to the full articulamentum of convolutional neural networks after process.
Such as, the weight matrix W ' after process can be expressed as: W '=U ' × S ' × V 't, then the full articulamentum linear transformation function after optimization process can be expressed as:
B＝A×U’×(S’×V’T)+b；
Wherein, A is input matrix, and B is output matrix, and U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and b is bias term.
In the present embodiment, by doing svd process to the weight matrix of articulamentum linear transformation function complete in convolutional neural networks, and obtain target linear transformation function according to the weight matrix after process, picture to be detected is detected, to obtain the content in picture to be detected, the detection speed of image content effectively can be promoted.
Fig. 3 is the process flow diagram of the detection method of a kind of image content according to another exemplary embodiment, and the detection method of this image content comprises:
In step S301, obtain the linear transformation function of the full articulamentum of picture to be detected and convolutional neural networks.
In embodiment of the present disclosure, picture to be detected is the picture needing Detection of content, and such as, picture to be detected can be the picture that user downloads on webpage, or also can be user to be taken pictures the picture obtained by video camera picture to be detected.
At present in technical field of image processing, user can utilize computing machine to process image, analyze and understand, and with the target of different mode various in recognition image and object, such as, can identify the position of object and the category attribute of object in picture to be detected.
As shown in Figure 2, Fig. 2 is the position of object in a kind of image content and the category attribute schematic diagram of object, and 21 in Fig. 2 is for identifying the position of object, and 22 for identifying the category attribute of object.
Convolutional neural networks is a kind of feedforward neural network, and its artificial neuron can respond the unit in partial coverage.
Alternatively, can detect the position of object and the category attribute of object in picture by the method based on convolutional neural networks, convolutional neural networks comprises convolutional layer, pond layer, full articulamentum, and activation function.
In the disclosed embodiments, process is optimized to the computation process of the full articulamentum in convolutional neural networks, to promote the detection speed of image content.
Alternatively, the linear transformation function of the full articulamentum of convolutional neural networks can show with following formula table:
B＝A×W+b；
Wherein, A is input matrix, and B is output matrix, and W is weight matrix, and b is bias term.
In step s 302, obtain the singular value of weight matrix, and the singular value of weight matrix is sorted, obtain the singular value after sorting.
Wherein, suppose that M is m × n rank matrixes, M*the non-negative square root of n the eigenwert of × M is called the singular value of matrix M, M*for the associate matrix of M.
In embodiment of the present disclosure, first can obtain the associate matrix W of weight matrix W*, and then by W*be multiplied with W, obtain matrix W*× W, compute matrix W*the non-negative square root of the eigenwert of × W, to obtain r the singular value of weight matrix W, finally carries out descending order sequence to r the singular value got, obtains r singular value after sorting.
By obtaining r singular value after sequence, the singular value of the predetermined number that subsequent step retention is larger can be made, at utmost can keep the degree of accuracy of the image detect based on convolutional neural networks.
In step S303, svd is done to the weight matrix in the linear transformation function of full articulamentum, obtain the matrix after decomposing.
In embodiment of the present disclosure, SVD decomposition can be carried out according to r the singular value sorted through descending order obtained in step S302 to weight matrix W, obtain the matrix after decomposing.
Alternatively, the matrix after decomposition comprises left singular vector matrix, diagonal matrix, and right singular vector matrix, and wherein, left singular vector matrix comprises the left singular vector of weight matrix W, and right singular vector matrix comprises the right singular vector of weight matrix W.
Such as, the weight matrix W after decomposition can be expressed as:
W＝U×S×VT；
Wherein, U is the left singular vector matrix on m × r rank, and S is the diagonal matrix on r × r rank, and V is the right singular vector matrix on n × r rank.
In step s 304, in the singular value after sequence, choose the singular value of predetermined number, and obtain the weight matrix after process according to the matrix after the singular value Sum decomposition of predetermined number.
Alternatively, the number of the singular value of predetermined number is preset by user.
By being preset the number of the singular value of predetermined number by user, user can be made to select the singular value of suitable predetermined number according to demand, effectively promote the applicability of image detect method.
In embodiment of the present disclosure, predetermined number can represent with k, when the value of k is much smaller than the number of r the singular value of weight matrix W, effectively can promote the detection speed of image content.
In step S305, obtain target linear transformation function according to the weight matrix after process, and according to target linear transformation function, picture to be detected is detected, to obtain the content in picture to be detected.
Wherein, the content in picture to be detected comprises the position of object and the category attribute of object.As shown in Figure 2.Obtain the content in picture to be detected, namely obtain position and the category attribute of the object in picture to be detected.
Alternatively, process can be optimized, to detect picture to be detected according to the full articulamentum linear transformation function after optimization process according to the linear transformation function of weight matrix to the full articulamentum of convolutional neural networks after process.
Such as, the weight matrix W ' after process can be expressed as: W '=U ' × S ' × V 't, then the full articulamentum linear transformation function after optimization process can be expressed as:
B＝A×U’×(S’×V’T)+b；
Wherein, A is input matrix, and B is output matrix, and U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and b is bias term.
Such as, can use the full articulamentum linear transformation function B=A × U ' after optimization process × (S ' × V 't)+b, as the target linear transformation function of the full articulamentum based on convolutional neural networks, picture to be detected is detected.
In the present embodiment, by obtaining r singular value after sequence, and in singular value after sequence, choose the singular value of predetermined number, and can the singular value of the larger predetermined number of retention, at utmost can keep the degree of accuracy of the image detect based on convolutional neural networks.By being preset the number of the singular value of predetermined number by user, user can be made to select the singular value of suitable predetermined number as required, effectively promote the applicability of image detect method.By doing svd process to the weight matrix of articulamentum linear transformation function complete in convolutional neural networks, and obtain target linear transformation function according to the weight matrix after process, picture to be detected is detected, to obtain the content in picture to be detected, the detection speed of image content effectively can be promoted.
Fig. 4 is the process flow diagram of the detection method of a kind of image content according to another exemplary embodiment, and the detection method of this image content comprises:
In step S401, obtain picture to be detected.
In embodiment of the present disclosure, picture to be detected is the picture needing Detection of content, and such as, picture to be detected can be the picture that user downloads on webpage, or also can be user to be taken pictures the picture obtained by video camera picture to be detected.
At present in technical field of image processing, user can utilize computing machine to process image, analyze and understand, and with the target of different mode various in recognition image and object, such as, can identify the position of object and the category attribute of object in picture to be detected.
As shown in Figure 2, Fig. 2 is the position of object in a kind of image content and the category attribute schematic diagram of object, and 21 in Fig. 2 is for identifying the position of object, and 22 for identifying the category attribute of object.
In step S402, obtain the linear transformation function of the full articulamentum of convolutional neural networks.
Convolutional neural networks is a kind of feedforward neural network, and its artificial neuron can respond the unit in partial coverage.
Alternatively, can detect the position of object and the category attribute of object in picture by the method based on convolutional neural networks, convolutional neural networks comprises convolutional layer, pond layer, full articulamentum, and activation function.
In the disclosed embodiments, process is optimized to the computation process of the full articulamentum in convolutional neural networks, to promote the detection speed of image content.
Alternatively, the linear transformation function of the full articulamentum of convolutional neural networks can show with following formula table:
B＝A×W+b；
Wherein, A is input matrix, and B is output matrix, and W is weight matrix, and b is bias term.
In step S403, obtain the singular value of weight matrix, and the singular value of weight matrix is sorted, obtain the singular value after sorting.
Wherein, suppose that M is m × n rank matrixes, M*the non-negative square root of n the eigenwert of × M is called the singular value of matrix M, M*for the associate matrix of M.
In embodiment of the present disclosure, first can obtain the associate matrix W of weight matrix W*, and then by W*be multiplied with W, obtain matrix W*× W, compute matrix W*the non-negative square root of the eigenwert of × W, to obtain r the singular value of weight matrix W, finally carries out descending order sequence to r the singular value got, obtains r singular value after sorting.
By obtaining r singular value after sequence, the singular value of the predetermined number that subsequent step retention is larger can be made, at utmost can keep the degree of accuracy of the image detect based on convolutional neural networks.
In step s 404, svd is done to the weight matrix in the linear transformation function of full articulamentum, obtain the matrix after decomposing.
In embodiment of the present disclosure, SVD decomposition can be carried out according to r the singular value sorted through descending order obtained in step S403 to weight matrix W, obtain the matrix after decomposing.
Alternatively, the matrix after decomposition comprises left singular vector matrix, diagonal matrix, and right singular vector matrix, and wherein, left singular vector matrix comprises the left singular vector of weight matrix W, and right singular vector matrix comprises the right singular vector of weight matrix W.
Such as, the weight matrix W after decomposition can be expressed as:
W＝U×S×VT；
Wherein, U is the left singular vector matrix on m × r rank, and S is the diagonal matrix on r × r rank, and V is the right singular vector matrix on n × r rank.
In step S405, according to the singular value of predetermined number to left singular vector matrix, diagonal matrix, and right singular vector matrix deals with, and obtains left singular vector matrix, diagonal matrix after processing, and right singular vector matrix.
Alternatively, the number of the singular value of predetermined number is preset by user.
By being preset the number of the singular value of predetermined number by user, user can be made to select the singular value of suitable predetermined number according to demand, effectively promote the applicability of image detect method.
In embodiment of the present disclosure, predetermined number can represent with k, when the value of k is much smaller than the number of r the singular value of weight matrix W, effectively can promote the detection speed of image content.
In embodiment of the present disclosure, can to the left singular vector matrix in the matrix after decomposition and right singular vector matrix, choose front k left singular vector in left singular vector matrix respectively, choose front k right singular vector in right singular vector matrix, to obtain the left singular vector matrix U after process ', U ' is the matrix on m × k rank, right singular vector matrix V ', V ' is the matrix on n × k rank, a front k singular value can be chosen to the diagonal matrix after decomposition, to obtain the diagonal matrix S ' after process, S ' is the matrix on k × k rank.
By choosing the left singular vector of predetermined number before in left singular vector matrix respectively, choose the right singular vector of the front predetermined number in right singular vector matrix, choose the singular value of predetermined number before in diagonal matrix, operational parameter is decreased on largely, reduce the algorithm complex based on convolutional neural networks detection calculations, effectively promote the detection efficiency of image content.
In step S406, according to left singular vector matrix, diagonal matrix after process, and right singular vector matrix obtains the weight matrix after processing.
Such as, can by the left singular vector matrix U ', right singular vector matrix V ' after process, and diagonal matrix S ', substitute into the expression formula W=U × S × V of the weight matrix W after decomposingt; In, obtain the weight matrix W ' after processing.
Such as, the weight matrix W ' after process can be expressed as:
W’＝U’×S’×V’T；
Wherein, U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and k is the number of the singular value of predetermined number.
In step S 407, target linear transformation function is obtained according to the weight matrix after process.
Such as, the weight matrix W ' after process can be expressed as: W '=U ' × S ' × V 't, then the full articulamentum linear transformation function after optimization process can be expressed as:
B＝A×U’×(S’×V’T)+b；
Wherein, A is input matrix, and B is output matrix, and U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and b is bias term.
In step S408, according to target linear transformation function, picture to be detected is detected, to obtain the content in picture to be detected.
Wherein, the content in picture to be detected comprises the position of object and the category attribute of object.As shown in Figure 2.Obtain the content in picture to be detected, namely obtain position and the category attribute of the object in picture to be detected.
Alternatively, process can be optimized, to detect picture to be detected according to the full articulamentum linear transformation function after optimization process according to the linear transformation function of weight matrix to the full articulamentum of convolutional neural networks after process.
Such as, can use the full articulamentum linear transformation function B=A × U ' after optimization process × (S ' × V 't)+b, as the target linear transformation function of the full articulamentum based on convolutional neural networks, picture to be detected is detected.
In the present embodiment, by obtaining r singular value after sequence, and in singular value after sequence, choose the singular value of predetermined number, and can the singular value of the larger predetermined number of retention, at utmost can keep the degree of accuracy of the image detect based on convolutional neural networks.By being preset the number of the singular value of predetermined number by user, user can be made to select the singular value of suitable predetermined number as required, effectively promote the applicability of image detect method.By choosing the left singular vector of predetermined number before in left singular vector matrix respectively, choose the right singular vector of the front predetermined number in right singular vector matrix, choose the singular value of predetermined number before in diagonal matrix, operational parameter is decreased on largely, reduce the algorithm complex based on convolutional neural networks detection calculations, effectively promote the detection efficiency of image content.By doing svd process to the weight matrix of articulamentum linear transformation function complete in convolutional neural networks, and obtain target linear transformation function according to the weight matrix after process, picture to be detected is detected, to obtain the content in picture to be detected, the detection speed of image content effectively can be promoted.
Fig. 5 is the block diagram of the pick-up unit of a kind of image content according to an exemplary embodiment, the pick-up unit of this image content can pass through software, hardware or both combinations and realize, the pick-up unit 50 of this image content comprises acquisition module 501, is configured to the linear transformation function of the full articulamentum obtaining picture to be detected and convolutional neural networks; Svd processing module 502, is configured to do svd process to the weight matrix in the linear transformation function of the full articulamentum that acquisition module 501 gets, and obtains the weight matrix after processing; And detection module 503, be configured to the weight matrix acquisition target linear transformation function after processing according to svd processing module 502, and according to target linear transformation function, picture to be detected detected, to obtain the content in picture to be detected.
Acquisition module 501, is configured to the linear transformation function of the full articulamentum obtaining picture to be detected and convolutional neural networks.
In embodiment of the present disclosure, picture to be detected is the picture needing Detection of content, and such as, picture to be detected can be the picture that user downloads on webpage, or also can be user to be taken pictures the picture obtained by video camera picture to be detected.
At present in technical field of image processing, user can utilize computing machine to process image, analyze and understand, and with the target of different mode various in recognition image and object, such as, can identify the position of object and the category attribute of object in picture to be detected.
As shown in Figure 2, Fig. 2 is the position of object in a kind of image content and the category attribute schematic diagram of object, and 21 in Fig. 2 is for identifying the position of object, and 22 for identifying the category attribute of object.
Convolutional neural networks is a kind of feedforward neural network, and its artificial neuron can respond the unit in partial coverage.
Alternatively, can detect the position of object and the category attribute of object in picture by the method based on convolutional neural networks, convolutional neural networks comprises convolutional layer, pond layer, full articulamentum, and activation function.
In the disclosed embodiments, process is optimized to the computation process of the full articulamentum in convolutional neural networks, to promote the detection speed of image content.
Alternatively, the linear transformation function of the full articulamentum of convolutional neural networks can show with following formula table:
B＝A×W+b；
Wherein, A is input matrix, and B is output matrix, and W is weight matrix, and b is bias term.
Svd processing module 502, is configured to do svd process to the weight matrix in the linear transformation function of the full articulamentum that acquisition module 501 gets, and obtains the weight matrix after processing.
Wherein, svd (SingularValueDecomposition, SVD) be a kind of matrix disassembling method in linear algebra, SVD decomposable process is such as: suppose that M is m × n rank matrixes, element wherein all belongs to territory K, K can be real number field, or complex field, then there is a decomposition and make:
M＝C×Ε×DT；
Wherein, Matrix C is m × r rank matrixes, and matrix Ε is positive semidefinite r × r rank diagonal matrix, and matrix D, be n × r rank matrixes, this decomposition be called the svd of M, the element on matrix Ε diagonal line is the singular value of M.
Alternatively, the matrix after decomposition comprises left singular vector matrix, diagonal matrix, and right singular vector matrix, and wherein, left singular vector matrix comprises the left singular vector of weight matrix W, and right singular vector matrix comprises the right singular vector of weight matrix W.
Such as, the weight matrix W after decomposition can be expressed as:
W＝U×S×VT；
Wherein, U is the left singular vector matrix on m × r rank, and S is the diagonal matrix on r × r rank, and V is the right singular vector matrix on n × r rank.
Alternatively, as shown in Figure 6, Fig. 6 is the block diagram of the pick-up unit of a kind of image content according to another exemplary embodiment, and svd processing module 502, comprising:
Sorting sub-module 5021, is configured to the singular value obtaining weight matrix, and sorts to the singular value of weight matrix, obtains the singular value after sorting.
Wherein, suppose that M is m × n rank matrixes, M*the non-negative square root of n the eigenwert of × M is called the singular value of matrix M, M*for the associate matrix of M.
In embodiment of the present disclosure, first can obtain the associate matrix W of weight matrix W*, and then by W*be multiplied with W, obtain matrix W*× W, compute matrix W*the non-negative square root of the eigenwert of × W, to obtain r the singular value of weight matrix W, finally carries out descending order sequence to r the singular value got, obtains r singular value after sorting.
By obtaining r singular value after sequence, the singular value of the predetermined number that subsequent step retention is larger can be made, at utmost can keep the degree of accuracy of the image detect based on convolutional neural networks.
Svd submodule 5022, is configured to make svd to the weight matrix in the linear transformation function of full articulamentum, obtains the matrix after decomposing.
In embodiment of the present disclosure, r the singular value sorted through descending order that can obtain according to sorting sub-module 5021 carries out SVD decomposition to weight matrix W, obtains the matrix after decomposition.
Alternatively, the matrix after decomposition comprises left singular vector matrix, diagonal matrix, and right singular vector matrix, and wherein, left singular vector matrix comprises the left singular vector of weight matrix W, and right singular vector matrix comprises the right singular vector of weight matrix W.
Such as, the weight matrix W after decomposition can be expressed as:
W＝U×S×VT；
Wherein, U is the left singular vector matrix on m × r rank, and S is the diagonal matrix on r × r rank, and V is the right singular vector matrix on n × r rank.
Weight matrix obtains submodule 5023, be configured to the singular value choosing predetermined number in the singular value after the sequence got at sorting sub-module 5021, and the matrix after decomposing according to the singular value of predetermined number and svd submodule 5022 obtains the weight matrix after process.
Alternatively, the number of the singular value of predetermined number is preset by user.
By being preset the number of the singular value of predetermined number by user, user can be made to select the singular value of suitable predetermined number according to demand, effectively promote the applicability of image detect method.
In embodiment of the present disclosure, predetermined number can represent with k, when the value of k is much smaller than the number of r the singular value of weight matrix W, effectively can promote the detection speed of image content.
Alternatively, weight matrix obtains submodule 5023, is configured to: according to the singular value of predetermined number to left singular vector matrix, diagonal matrix, and right singular vector matrix deals with, obtain left singular vector matrix, diagonal matrix after processing, and right singular vector matrix; And according to left singular vector matrix, the diagonal matrix after process, and right singular vector matrix obtains the weight matrix after processing.
In embodiment of the present disclosure, can to the left singular vector matrix in the matrix after decomposition and right singular vector matrix, choose front k left singular vector in left singular vector matrix respectively, choose front k right singular vector in right singular vector matrix, to obtain the left singular vector matrix U after process ', U ' is the matrix on m × k rank, right singular vector matrix V ', V ' is the matrix on n × k rank, a front k singular value can be chosen to the diagonal matrix after decomposition, to obtain the diagonal matrix S ' after process, S ' is the matrix on k × k rank.
By choosing the left singular vector of predetermined number before in left singular vector matrix respectively, choose the right singular vector of the front predetermined number in right singular vector matrix, choose the singular value of predetermined number before in diagonal matrix, operational parameter is decreased on largely, reduce the algorithm complex based on convolutional neural networks detection calculations, effectively promote the detection efficiency of image content.
Such as, can by the left singular vector matrix U ', right singular vector matrix V ' after process, and diagonal matrix S ', substitute into the expression formula W=U × S × V of the weight matrix W after decomposingt; In, obtain the weight matrix W ' after processing.
Such as, the weight matrix W ' after process can be expressed as:
W’＝U’×S’×V’T；
Wherein, U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and k is the number of the singular value of predetermined number.
Detection module 503, is configured to the weight matrix acquisition target linear transformation function after processing according to svd processing module 502, and detects picture to be detected according to target linear transformation function, to obtain the content in picture to be detected.
Wherein, the content in picture to be detected comprises the position of object and the category attribute of object.As shown in Figure 2.Obtain the content in picture to be detected, namely obtain position and the category attribute of the object in picture to be detected.
Alternatively, process can be optimized, to detect picture to be detected according to the full articulamentum linear transformation function after optimization process according to the linear transformation function of weight matrix to the full articulamentum of convolutional neural networks after process.
Such as, the weight matrix W ' after process can be expressed as: W '=U ' × S ' × V 't, then the full articulamentum linear transformation function after optimization process can be expressed as:
B＝A×U’×(S’×V’T)+b；
Wherein, A is input matrix, and B is output matrix, and U ' is the left singular vector matrix on m × k rank, and S ' is the diagonal matrix on k × k rank, and V ' is the right singular vector matrix on n × k rank, and b is bias term.
Such as, can use the full articulamentum linear transformation function B=A × U ' after optimization process × (S ' × V 't)+b, as the target linear transformation function of the full articulamentum based on convolutional neural networks, picture to be detected is detected.
In the present embodiment, by doing svd process to the weight matrix of articulamentum linear transformation function complete in convolutional neural networks, and obtain target linear transformation function according to the weight matrix after process, picture to be detected is detected, to obtain the content in picture to be detected, the detection speed of image content effectively can be promoted.
Fig. 7 is the block diagram of the pick-up unit 700 of a kind of image content according to an exemplary embodiment.Such as, device 700 can be mobile phone, computing machine, digital broadcast terminal, messaging devices, game console, tablet device, Medical Devices, body-building equipment, personal digital assistant etc.
With reference to Fig. 7, device 700 can comprise following one or more assembly: processing components 702, storer 704, power supply module 706, multimedia groupware 708, audio-frequency assembly 710, the interface 712 of I/O (I/O), sensor module 714, and communications component 716.
The integrated operation of the usual control device 700 of processing components 702, such as with display, call, data communication, camera operation and record operate the operation be associated.Processing components 702 can comprise one or more processor 720 to perform instruction, to complete all or part of step of above-mentioned method.In addition, processing components 702 can comprise one or more module, and what be convenient between processing components 702 and other assemblies is mutual.Such as, processing components 702 can comprise multi-media module, mutual with what facilitate between multimedia groupware 708 and processing components 702.
Storer 704 is configured to store various types of data to be supported in the operation of device 700.The example of these data comprises the instruction being configured to any application program or the method operated on device 700, contact data, telephone book data, message, picture, video etc.Storer 704 can be realized by the volatibility of any type or non-volatile memory device or their combination, as static RAM (SRAM), Electrically Erasable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory EPROM (EPROM), programmable read only memory (PROM), ROM (read-only memory) (ROM), magnetic store, flash memory, disk or CD.
The various assemblies that electric power assembly 706 is device 700 provide electric power.Electric power assembly 706 can comprise power-supply management system, one or more power supply, and other and the assembly generating, manage and distribute electric power for device 700 and be associated.
Multimedia groupware 708 is included in the touching display screen providing an output interface between device 700 and user.In certain embodiments, touching display screen can comprise liquid crystal display (LCD) and touch panel (TP).Touch panel comprises one or more touch sensor with the gesture on sensing touch, slip and touch panel.Touch sensor can the border of not only sensing touch or sliding action, but also detects the duration relevant with touch or slide and pressure.In certain embodiments, multimedia groupware 708 comprises a front-facing camera and/or post-positioned pick-up head.When device 700 is in operator scheme, during as screening-mode or video mode, front-facing camera and/or post-positioned pick-up head can receive outside multi-medium data.Each front-facing camera and post-positioned pick-up head can be fixing optical lens systems or have focal length and optical zoom ability.
Audio-frequency assembly 710 is configured to export and/or input audio signal.Such as, audio-frequency assembly 710 comprises a microphone (MIC), and when device 700 is in operator scheme, during as call model, logging mode and speech recognition mode, microphone is configured to receive external audio signal.The sound signal received can be stored in storer 704 further or be sent via communications component 716.In certain embodiments, audio-frequency assembly 710 also comprises a loudspeaker, is configured to output audio signal.
I/O interface 712 is for providing interface between processing components 702 and peripheral interface module, and above-mentioned peripheral interface module can be keyboard, some striking wheel, button etc.These buttons can include but not limited to: home button, volume button, start button and locking press button.
Sensor module 714 comprises one or more sensor, is configured to as device 700 provides the state estimation of various aspects.Such as, sensor module 714 can detect the opening/closing state of device 700, the relative positioning of assembly, such as assembly is display and the keypad of device 700, the position of all right pick-up unit 700 of sensor module 714 or device 700 1 assemblies changes, the presence or absence that user contacts with device 700, the temperature variation of device 700 orientation or acceleration/deceleration and device 700.Sensor module 714 can comprise proximity transducer, be configured to without any physical contact time detect near the existence of object.Sensor module 714 can also comprise optical sensor, as CMOS or ccd image sensor, is configured to use in imaging applications.In certain embodiments, this sensor module 714 can also comprise acceleration transducer, gyro sensor, Magnetic Sensor, pressure transducer or temperature sensor.
Communications component 716 is configured to the communication being convenient to wired or wireless mode between device 700 and other equipment.Device 700 can access the wireless network based on communication standard, as WiFi, 2G or 3G, or their combination.In one exemplary embodiment, communications component 716 receives from the broadcast singal of external broadcasting management system or broadcast related information via broadcast channel.In one exemplary embodiment, communications component 716 also comprises near-field communication (NFC) module, to promote junction service.Such as, can based on radio-frequency (RF) identification (RFID) technology in NFC module, Infrared Data Association (IrDA) technology, ultra broadband (UWB) technology, bluetooth (BT) technology and other technologies realize.
In the exemplary embodiment, device 700 can be realized by one or more application specific integrated circuit (ASIC), digital signal processor (DSP), digital signal processing appts (DSPD), programmable logic device (PLD) (PLD), field programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic components, is configured to the detection method performing above-mentioned image content.
In the exemplary embodiment, additionally provide a kind of non-transitory computer-readable recording medium comprising instruction, such as, comprise the storer 704 of instruction, above-mentioned instruction can perform said method by the processor 720 of device 700.Such as, non-transitory computer-readable recording medium can be ROM, random access memory (RAM), CD-ROM, tape, floppy disk and optical data storage devices etc.
A kind of non-transitory computer-readable recording medium, when the instruction in storage medium is performed by the processor of mobile terminal, make mobile terminal can perform a kind of detection method of image content, the method comprises:
Obtain the linear transformation function of the full articulamentum of picture to be detected and convolutional neural networks;
Svd process is done to the weight matrix in the linear transformation function of full articulamentum, obtains the weight matrix after processing; And
Obtain target linear transformation function according to the weight matrix after process, and according to target linear transformation function, picture to be detected is detected, to obtain the content in picture to be detected.
It should be noted that, the explanation of the aforementioned detection method embodiment to image content illustrates the pick-up unit being also applicable to the image content of this embodiment, and it is similar that it realizes principle, repeats no more herein.
Those skilled in the art, at consideration instructions and after putting into practice invention disclosed herein, will easily expect other embodiment of the present disclosure.The application is intended to contain any modification of the present disclosure, purposes or adaptations, and these modification, purposes or adaptations are followed general principle of the present disclosure and comprised the undocumented common practise in the art of the disclosure or conventional techniques means.Instructions and embodiment are only regarded as exemplary, and true scope of the present disclosure and spirit are pointed out by claim below.
Should be understood that, the disclosure is not limited to precision architecture described above and illustrated in the accompanying drawings, and can carry out various amendment and change not departing from its scope.The scope of the present disclosure is only limited by appended claim.