WO2017117799A1 - 分类结果的融合装置及方法 - Google Patents

分类结果的融合装置及方法 Download PDF

Info

Publication number
WO2017117799A1
WO2017117799A1 PCT/CN2016/070473 CN2016070473W WO2017117799A1 WO 2017117799 A1 WO2017117799 A1 WO 2017117799A1 CN 2016070473 W CN2016070473 W CN 2016070473W WO 2017117799 A1 WO2017117799 A1 WO 2017117799A1
Authority
WO
WIPO (PCT)
Prior art keywords
classification
result
classification result
weight
results
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2016/070473
Other languages
English (en)
French (fr)
Inventor
伍健荣
刘晓青
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fujitsu Ltd
Original Assignee
Fujitsu Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fujitsu Ltd filed Critical Fujitsu Ltd
Priority to PCT/CN2016/070473 priority Critical patent/WO2017117799A1/zh
Publication of WO2017117799A1 publication Critical patent/WO2017117799A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques

Definitions

  • the present invention relates to the field of information technology, and in particular, to a device and method for merging classification results.
  • fusion methods of multiple classification results for example, finding a classifier with the lowest error rate among a plurality of classifiers and outputting the classification result of the classifier, or determining the classification result by the voting result of each classifier.
  • the embodiment of the invention provides a fusion device and method for classifying results, and synthesizing and merging the classification results according to the average degree of difference between each classification result and other classification results, which can effectively improve the accuracy of the classification result.
  • a fusion device for classification results comprising: a first calculation unit, wherein the first calculation unit is configured to calculate any two classification results of at least two classification results a degree of difference between the two calculation units, the second calculation unit is configured to calculate an average difference degree between each classification result and another classification result according to a degree of difference between any two classification results; a fusion unit, Convergence list The element is used to synthesize each classification result according to the average degree of difference between each classification result and other classification results, and fuse the synthesis result to obtain the fusion result of the at least two classification results.
  • an electronic device comprising: a fusion device of classification results according to the first aspect of the embodiments of the present invention.
  • a method for merging classification results comprising: calculating a degree of difference between any two classification results of at least two classification results; The degree of difference, the average difference between each classification result and other classification results is calculated; the classification results are synthesized according to the average difference between each classification result and other classification results, and the synthesis results are fused to obtain the The fusion result of at least two classification results.
  • the invention has the beneficial effects that the classification and fusion of the respective classification results according to the average difference degree between the respective classification results and other classification results can effectively improve the accuracy of the classification result.
  • FIG. 1 is a schematic diagram of a fusion device of a classification result according to Embodiment 1 of the present invention
  • FIG. 2 is a schematic diagram of a fusion unit 103 according to Embodiment 1 of the present invention.
  • FIG. 3 is a schematic diagram of an electronic device according to Embodiment 2 of the present invention.
  • FIG. 4 is a schematic block diagram showing the system configuration of an electronic device according to Embodiment 2 of the present invention.
  • FIG. 5 is a flowchart of a method for merging classification results according to Embodiment 3 of the present invention.
  • Fig. 6 is a flow chart showing a method of fusing the classification result according to the fourth embodiment of the present invention.
  • the device 100 includes:
  • a first calculating unit 101 configured to calculate a degree of difference between any two of the at least two classification results
  • a second calculating unit 102 configured to calculate an average degree of difference between each classification result and other classification results according to a degree of difference between any two classification results
  • the fusion unit 103 is configured to synthesize each classification result according to the average difference degree between each classification result and other classification results, and fuse the synthesis result to obtain a fusion result of at least two classification results.
  • the accuracy and classification of the classification results can be effectively improved by synthesizing and merging the classification results according to the average degree of difference between the classification results and other classification results.
  • At least two classification results can be obtained by using various existing methods.
  • multiple image acquisition devices input the respective acquired image data into different deep convolutional neural networks (dCNN, Deep Convolution Neural). Networks), after calculation, obtain multiple classification results; or, multiple image acquisition devices input the respective acquired image data into the same deep convolutional neural network, and obtain multiple classification results after calculation; or, the depth volume
  • dCNN Deep Convolution Neural
  • Networks After calculation, obtain multiple classification results
  • multiple image acquisition devices input the respective acquired image data into different deep convolutional neural networks
  • multiple image acquisition devices input the respective acquired image data into the same deep convolutional neural network, and obtain multiple classification results after calculation; or, the depth volume
  • the output of the different layers of the neural network is input to the classifier, and the plurality of classification results are obtained after the calculation.
  • the embodiment of the present invention does not limit the acquisition manner of the multiple classification results.
  • the first calculating unit 101 calculates a degree of difference between any two of the at least two classification results, wherein the degree of difference can be characterized using an existing method, for example, by calculating any two classifications
  • the Jousselme distance between the results is used to characterize the difference.
  • the Jousselme distance is taken as an example to characterize the degree of difference between any two classification results.
  • n is an integer greater than or equal to 2
  • n classification results are sorted, respectively, m 1, m 2 , ..., m n , wherein, in any order Sort the n classification results.
  • J distance the Jousselme distance between any two classification results
  • d BPA (i, j) represents the J distance between the i-th classification result and the j-th classification result
  • m i represents the i-th classification result
  • m j represents the j-th classification result
  • a p and A q Representing the p-th category and the q-th category, respectively
  • i, j, p, q, n are positive integers
  • i, j, p, q are all less than or equal to n.
  • the second calculating unit 102 calculates the average degree of difference between each classification result and other classification results according to the degree of difference between any two classification results, for example, each of the following formulas (3) can be calculated.
  • Sum_m(i) represents the average J distance between the i-th classification result and other classification results
  • d BPA (i, j) represents the J distance between the i-th classification result and the j-th classification result.
  • the fusion unit 103 synthesizes the respective classification results according to the average difference degree between the respective classification results and other classification results, and performs the synthesis result. Fusion, obtaining fusion results of at least two classification results.
  • the fusion unit 103 synthesizes the fusion result of each classification result with the previous classification result of the classification result or all the classification results before the classification result according to the average degree of difference between each classification result and other classification results, and The synthesis results are fused to obtain fusion results of at least two classification results.
  • the fusion unit 103 synthesizes the first classification result and the second classification result, and for the third classification result, the fusion unit 103 synthesizes the fusion result of the third classification result and the first two classification results. , for the classification result after the third classification result, and so on.
  • the merging unit 103 may set a weight for performing synthesis according to a comparison result between an average difference degree between the respective classification results and other classification results and a preset threshold value, and perform each classification result according to the weight value. Synthesis, and fusion of the synthesis results to obtain fusion results of at least two classification results. For example, when somebody When the average difference between the classification result and other classification results is greater than the preset threshold, the weight used for the classification result at the time of composition is set to be small.
  • the preset threshold may be set according to actual needs, for example, the threshold is set to a value between 0 and 1, for example, the threshold is 0.5.
  • the following is a description of the configuration of the fusing unit 103 of the present embodiment and the method of synthesizing and fusing according to the average degree of difference.
  • FIG. 2 is a schematic diagram of a merging unit 103 according to Embodiment 1 of the present invention.
  • the fusion unit 103 includes:
  • the first setting unit 201 is configured to: when the first average difference between the first classification result and the other classification result is greater than a preset threshold, and the second average difference between the second classification result and the other classification result is smaller than the preset
  • the threshold value is set, the first weight value of the first classification result is calculated based on the first average difference degree, and the second weight value of the second classification result is calculated based on the first weight value, wherein the first weight value is smaller than the second weight value;
  • the second setting unit 202 is configured to: when the first average difference between the first classification result and the other classification result is less than the preset threshold, and the second average difference between the second classification result and the other classification result is greater than
  • the threshold value is preset
  • the second weight value of the second classification result is calculated based on the second average difference degree
  • the first weight value of the first classification result is calculated according to the second weight value, wherein the second weight value is smaller than the first weight value ;
  • the third setting unit 203 is configured to: when the first average difference between the first classification result and the other classification result and the second average difference between the second classification result and the other classification result are greater than or equal to the preset When the threshold is less than the preset threshold, the first weight of the first classification result and the second weight of the second classification result are set to be equal;
  • the first combining unit 204 is configured to synthesize the first classification result and the second classification result according to the first weight value and the second weight value;
  • the first fusion unit 205 is configured to fuse the combined result of the first classification result and the second classification result to obtain a fusion result of the first classification result and the second classification result.
  • the first setting unit 201 calculates the first weight of the first classification result based on the first average difference degree.
  • the first weight value of the first classification result can be calculated using the following formula (4):
  • the first weight is represented, Sum_m(1) represents the first average degree of difference of the first classification result, and ⁇ represents the weight adjustment factor.
  • ⁇ can be set according to actual needs, for example, ⁇ is a value of 2 to 5.
  • the first setting unit 201 calculates the second weight of the second classification result based on the first weight, for example, As the second weight. among them, which is
  • the second setting unit 202 calculates the second weight of the second classification result based on the second average difference degree, for example, the second weight of the second classification result can be calculated using the formula (5) similar to the above formula (4). value:
  • the second weight is represented, Sum_m(2) represents the second average degree of difference of the second classification result, and ⁇ represents the weight adjustment factor, and the value is the same as before.
  • the second setting unit 202 calculates the first weight of the first classification result based on the second weight, for example, As the first weight. among them, which is
  • the third setting unit 203 sets the first weight of the first classification result and the second weight of the second classification result to be equal, for example, the first weight and the second weight. Set to 0.5.
  • the first synthesizing unit 204 synthesizes the first classification result and the second classification result based on the calculated first weight value and the second weight value, for example,
  • the first weight and the second weight set by the first setting unit 201 can be synthesized according to the following formula (6):
  • a synthesis result indicating the first classification result and the second classification result m 1 represents the first classification result, and m 2 represents the second classification result, Indicates the first weight, Indicates the second weight.
  • the first weight and the second weight set by the second setting unit 202 can be synthesized according to the following formula (7):
  • the first weight and the second weight set by the third setting unit 203 can be synthesized according to the following formula (8):
  • m 1 represents the first classification result
  • m 2 represents the second classification result
  • the first fusion unit 205 fuses the synthesis result of the first classification result and the second classification result to obtain the first classification result and the first classification result.
  • a fusion result indicating the result of the first classification and the result of the second classification
  • a synthesis result indicating the result of the first classification and the result of the second classification.
  • the fusion of the synthesis results may use existing fusion methods, for example, fusion based on D-S (Dempster/Shafer) evidence theory.
  • the fusion unit 103 may further include:
  • the fourth setting unit 206 is configured to calculate an ith weight of the i-th classification result according to the ith average difference degree when the ith average difference degree between the i-th classification result and the other classification result is greater than the preset threshold value, Calculating an i-1th weight of the fusion result of the i-1 classification results before the i-th classification result according to the i-th weight, wherein the i-th weight is smaller than the i-th weight, and i is greater than or equal to 3. Integer, and i ⁇ n;
  • the fifth setting unit 207 is configured to set an ith weight of the i-th classification result according to the preset parameter when the ith average difference degree between the ith classification result and the other classification result is less than or equal to the preset threshold value And an i-1th weight of the fusion result of the i-1 classification results before the i-th classification result;
  • the fourth setting unit 206 calculates the i-th of the i-th classification result according to the ith average difference degree.
  • the weight for example, the i-th weight can be calculated using the formula (10) similar to the above formula (4):
  • the fourth setting unit 206 calculates the i-1th weight of the fusion result of the i-1 classification results before the i-th classification result according to the i-th weight value, for example, As the i-1th weight. among them, which is
  • the fifth setting unit 207 sets the first i-category result according to the preset parameter.
  • the i-th weight and the i-1th weight of the fusion result of the i-1 classification results before the i-th classification result for example, the i-th weight is set to ⁇ , then the i-th weight is (1- ⁇ ).
  • ⁇ is a value of 0 to 1, which can be set according to actual needs.
  • the first synthesizing unit 204 is further configured to synthesize the fusion result of the i-th classification result and the i-1 classification result before the i-th classification result according to the i-th weight value and the i-th weight value, for example,
  • the i-th weight and the i-th weight set by the fourth setting unit 206 can be synthesized according to the following formula (11):
  • a synthesis result indicating a fusion result of the i-th classification result and the i-1 classification result before the i-th classification result and mi represents the i-th classification result
  • the fusion result of the i-1 classification results before the i-th classification result Indicates the i-th weight, Indicates the i-1th weight.
  • the i-th weight and the i-th weight set by the fifth setting unit 207 can be synthesized according to the following formula (12):
  • mi represents the i-th classification result
  • the fusion result of i-1 classification results before the i-th classification result is expressed, ⁇ represents the i-th weight, and (1- ⁇ ) represents the i-th weight.
  • the first fusion unit 205 is further used for the i-th classification result and the i-th classification.
  • the synthesis results of the fusion results of the previous i-1 classification results were fused, and the fusion results of the i classification results were obtained.
  • the fusion method is the same as the previous description, and will not be described here.
  • the fusion unit 103 when n ⁇ 3, the fusion unit 103 performs the above iterative calculation, and finally obtains n The result of the fusion of the classification results.
  • the accuracy and classification of the classification results can be effectively improved by synthesizing and merging the respective classification results according to the average degree of difference between the respective classification results and other classification results.
  • FIG. 3 is a schematic diagram of an electronic device according to Embodiment 2 of the present invention.
  • the electronic device 300 includes the fusion device 301 of the classification result.
  • the structure and function of the fusion device 301 of the classification result are the same as those in the first embodiment, and details are not described herein again.
  • FIG. 4 is a schematic block diagram showing the system configuration of an electronic device according to Embodiment 2 of the present invention.
  • electronic device 400 can include central processor 401 and memory 402; memory 402 is coupled to central processor 401.
  • the figure is exemplary; other types of structures may be used in addition to or in place of the structure to implement telecommunications functions or other functions.
  • the electronic device 400 may further include an input unit 403, a display 404, and a power source 405.
  • the functionality of the fusion device of the classification results described in Embodiment 1 can be integrated into the central processor 401.
  • the central processing unit 401 may be configured to: calculate a degree of difference between any two of the at least two classification results; calculate the classification result and other classification results according to the degree of difference between the two classification results.
  • the average degree of difference between the two classification results is synthesized according to the average degree of difference between each classification result and other classification results, and the fusion result is fused to obtain the fusion result of the at least two classification results.
  • the combining, according to the average degree of difference between each classification result and other classification results, synthesizing the classification results, and combining the synthesis results to obtain the fusion result of the at least two classification results including: according to each classification result
  • the comparison result of the average difference degree and the preset threshold value with other classification results sets a weight for performing synthesis, synthesizes each classification result according to the weight value, and fuses the synthesis result to obtain the at least The fusion result of the two classification results.
  • the at least two classification results include a first classification result and a second classification result, and the classification results are synthesized according to the average difference degree between each classification result and other classification results, and the synthesis result is merged.
  • Obtaining a fusion result of the at least two classification results comprising: a first average difference between the first classification result and the other classification result is greater than a preset threshold and a second average between the second classification result and other classification results
  • the difference degree is less than the predetermined threshold
  • the threshold value is greater than the predetermined threshold
  • the second weight value of the second classification result is calculated according to the second average difference degree
  • the first weight value of the first classification result is
  • the at least two classification results include n classification results including the first classification result and the second classification result, and n is an integer greater than or equal to 3, and the average is based on the average of the classification results and other classification results.
  • the degree of difference is synthesized for each classification result, and the synthesis result is fused to obtain the fusion result of the at least two classification results, and further includes: when the ith average difference between the i-th classification result and other classification results is greater than
  • the threshold value is preset
  • the i-th weight of the i-th classification result is calculated according to the ith average difference degree
  • the fusion result of the i-1 classification results before the i-th classification result is calculated according to the i-th weight value
  • An i-1th weight wherein the i-th weight is smaller than the i-th weight, i is an integer greater than or equal to 3, and i ⁇ n; when the i-th classification result and other classification results When the ith average difference is less than or equal to the preset threshold, the
  • the i-1th weight according to the i-th weight and An i-1 weight, synthesizing the fusion result of the i-th classification result and the i-1 classification result before the i-th classification result; i-1 before the i-th classification result and the i-th classification result
  • the synthesis results of the fusion results of the classification results are fused, and the fusion results of the i classification results are obtained.
  • the preset parameter is set such that the i-th weight and the i-th 1 The weights are equal.
  • the fusion device of the classification result described in Embodiment 1 may be configured separately from the central processing unit 401.
  • the fusion device of the classification result may be configured as a chip connected to the central processing unit 401 through the central processing unit.
  • the control of 401 is to implement the function of the fusion device of the classification result.
  • the electronic device 400 also does not have to include all of the components shown in FIG. 4 in this embodiment.
  • central processor 401 may include a microprocessor or other processor device and/or logic device that receives input and controls electronic device 400. The operation of each component.
  • Memory 402 can be one or more of a buffer, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable device.
  • the central processing unit 401 can execute the program stored in the memory 402 to implement information storage or processing and the like.
  • the functions of other components are similar to those of the existing ones and will not be described here.
  • the various components of electronic device 400 may be implemented by special purpose hardware, firmware, software, or a combination thereof without departing from the scope of the invention.
  • the accuracy and classification of the classification results can be effectively improved by synthesizing and merging the respective classification results according to the average degree of difference between the respective classification results and other classification results.
  • the embodiment of the present invention further provides a fusion method of classification results, which corresponds to the fusion device of the classification result of Embodiment 1.
  • Fig. 5 is a flow chart showing a method of fusing the classification result according to the third embodiment of the present invention. As shown in FIG. 5, the method includes:
  • Step 501 Calculate a degree of difference between any two of the at least two classification results
  • Step 502 Calculate an average difference between each classification result and other classification results according to the degree of difference between any two classification results
  • Step 503 Synthesize each classification result according to the average difference degree between each classification result and other classification results, and fuse the synthesis result to obtain a fusion result of at least two classification results.
  • the method of calculating the difference degree and the average degree of difference and performing the synthesis and fusion according to the average degree of difference is the same as that described in Embodiment 1, and details are not described herein again.
  • the accuracy and classification of the classification results can be effectively improved by synthesizing and merging the classification results according to the average degree of difference between the classification results and other classification results.
  • the embodiment of the present invention further provides a fusion method of classification results, which corresponds to the fusion device of the classification result of Embodiment 1.
  • Fig. 6 is a flow chart showing a method of fusing the classification result according to the fourth embodiment of the present invention. As shown in FIG. 6, the method includes:
  • Step 601 Calculate the degree of difference between any two of the n classification results, where n is an integer greater than or equal to 2;
  • Step 602 Calculate an average difference between each classification result and other classification results according to the degree of difference between any two classification results
  • Step 603 determining the n-th classification results of the first classification result m a first average degree of difference Sum_m (1) 1 and the second classification result m the second average difference of Sum_m 2 (2) satisfies the following condition: Sum_m (1 >th_d and Sum_m(2) ⁇ th_d, th_d are preset thresholds.
  • Step 604 Calculate a first weight of the first classification result according to the first average difference degree, and calculate a second weight value of the second classification result according to the first weight value;
  • Step 605 determining a first classification result m a first average degree of difference Sum_m (1) 1 and the second classification result m the second average difference of Sum_m 2 (2) satisfies the following condition: Sum_m (1) ⁇ th_d and Sum_m (2)>th_d, th_d is the preset threshold.
  • Step 606 Calculate a second weight of the second classification result according to the second average difference degree, and calculate a first weight value of the first classification result according to the second weight value;
  • Step 607 Set the first weight of the first classification result and the second weight of the second classification result to be equal;
  • Step 608 Synthesize the first classification result and the second classification result according to the first weight value and the second weight value
  • Step 609 Fusing the synthesis result of the first classification result and the second classification result to obtain a fusion result of the first classification result and the second classification result;
  • Step 610 Determine whether n is greater than or equal to 3. When the determination result is "Yes”, proceed to step 611, and when the determination result is "No", the process ends.
  • Step 611 Determine whether the ith average difference degree Sum_m(i) of the i-th classification result m i is greater than the preset threshold value th_d.
  • the process proceeds to step 612, and when the determination result is “No”. , then proceeds to step 613, i is an integer greater than or equal to 3, and i ⁇ n;
  • Step 612 Calculate an i-th weight of the i-th classification result according to the ith average difference degree, and calculate an i-th weight value of the fusion result of the i-1 classification results before the i-th classification result according to the i-th weight value;
  • Step 613 Set an i-th weight of the i-th classification result and an i-th weight value of the fusion result of the i-1 classification results before the i-th classification result according to the preset parameter;
  • Step 614 synthesize the fusion result of the i-th classification result and the i-1 classification result before the i-th classification result according to the i-th weight value and the i-th weight value;
  • Step 615 merging the combined results of the fusion result of the i-th classification result and the i-1 classification result before the i-th classification result, and obtaining the fusion result of the i classification results;
  • Step 616 determining whether i is less than n, when the determination result is "Yes”, proceeding to step 617, and when the determination result is "No", ending the process;
  • Step 617 Add i to 1.
  • the accuracy and classification of the classification results can be effectively improved by synthesizing and merging the classification results according to the average degree of difference between the classification results and other classification results.
  • Embodiments of the present invention also provide a computer readable program, wherein when the program is executed in a fusion device or an electronic device of a classification result, the program causes a computer to perform implementation in a fusion device or an electronic device of the classification result
  • An embodiment of the present invention further provides a storage medium storing a computer readable program, wherein the computer readable program causes a computer to perform the classification result described in Embodiment 3 or Embodiment 4 in a fusion device or an electronic device of a classification result. Fusion method.
  • the above apparatus and method of the present invention may be implemented by hardware or by hardware in combination with software.
  • the present invention relates to a computer readable program that, when executed by a logic component, enables the logic component to implement the apparatus or components described above, or to cause the logic component to implement the various methods described above Or steps.
  • the present invention also relates to a storage medium for storing the above program, such as a hard disk, a magnetic disk, an optical disk, a DVD, a flash memory, or the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种分类结果的融合装置及方法。该装置包括：第一计算单元，用于计算至少两个分类结果中任意两个分类结果之间的差异度；第二计算单元，用于根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；融合单元，用于根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。

Description

分类结果的融合装置及方法 技术领域
本发明涉及信息技术领域，尤其涉及一种分类结果的融合装置及方法。
背景技术
随着信息技术的不断发展，计算机视觉以及智能交通系统的应用逐渐广泛。基于这些应用的要求，需要对获得的视频图像中的目标物体进行识别和分类。当视频图像由于例如光照、烟雾、雨雪等成像环境以及运动模糊等因素的影响而导致成像模糊、扭曲或残缺等，导致对于目标物体的单个分类结果的精度大大降低。因此，获得目标物体的多个分类结果并对这些分类结果进行融合，相对于单个分类结果，能够提高目标物体的识别和分类的精度。
目前，存在多个分类结果的融合方法，例如，在多个分类器中寻找错误率最低的分类器并输出该分类器的分类结果，或者，通过各个分类器的投票结果来决定分类结果。
应该注意，上面对技术背景的介绍只是为了方便对本发明的技术方案进行清楚、完整的说明，并方便本领域技术人员的理解而阐述的。不能仅仅因为这些方案在本发明的背景技术部分进行了阐述而认为上述技术方案为本领域技术人员所公知。
发明内容
当利用上述现有的方法进行多个分类结果的融合时，获得的分类结果准确度仍然较低。
本发明实施例提供一种分类结果的融合装置及方法，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
根据本发明实施例的第一方面，提供一种分类结果的融合装置，所述装置包括：第一计算单元，所述第一计算单元用于计算至少两个分类结果中任意两个分类结果之间的差异度；第二计算单元，所述第二计算单元用于根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；融合单元，所述融合单 元用于根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
根据本发明实施例的第二方面，提供一种电子设备，包括：根据本发明实施例的第一方面所述的分类结果的融合装置。
根据本发明实施例的第三方面，提供一种分类结果的融合方法，所述方法包括：计算至少两个分类结果中任意两个分类结果之间的差异度；根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
本发明的有益效果在于：根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
参照后文的说明和附图，详细公开了本发明的特定实施方式，指明了本发明的原理可以被采用的方式。应该理解，本发明的实施方式在范围上并不因而受到限制。在所附权利要求的精神和条款的范围内，本发明的实施方式包括许多改变、修改和等同。
针对一种实施方式描述和/或示出的特征可以以相同或类似的方式在一个或更多个其它实施方式中使用，与其它实施方式中的特征相组合，或替代其它实施方式中的特征。
应该强调，术语“包括/包含”在本文使用时指特征、整件、步骤或组件的存在，但并不排除一个或更多个其它特征、整件、步骤或组件的存在或附加。
附图说明
所包括的附图用来提供对本发明实施例的进一步的理解，其构成了说明书的一部分，用于例示本发明的实施方式，并与文字描述一起来阐释本发明的原理。显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。在附图中：
图1是本发明实施例1的分类结果的融合装置的示意图；
图2是本发明实施例1的融合单元103的示意图；
图3是本发明实施例2的电子设备的示意图；
图4是本发明实施例2的电子设备的系统构成的一示意框图；
图5是本发明实施例3的分类结果的融合方法的流程图；
图6是本发明实施例4的分类结果的融合方法的流程图。
具体实施方式
参照附图，通过下面的说明书，本发明的前述以及其它特征将变得明显。在说明书和附图中，具体公开了本发明的特定实施方式，其表明了其中可以采用本发明的原则的部分实施方式，应了解的是，本发明不限于所描述的实施方式，相反，本发明包括落入所附权利要求的范围内的全部修改、变型以及等同物。
实施例1
图1是本发明实施例1的分类结果的融合装置的示意图。如图1所示，装置100包括：
第一计算单元101，用于计算至少两个分类结果中任意两个分类结果之间的差异度；
第二计算单元102，用于根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；
融合单元103，用于根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得至少两个分类结果的融合结果。
由上述实施例可知，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
在本实施例中，至少两个分类结果可通过多种现有方式获得，例如，多个图像采集装置将各自采集到的图像数据分别输入到不同的深度卷积神经网络(dCNN,Deep Convolution Neural Networks)，经过计算后获得多个分类结果；或者，多个图像采集装置将各自采集到的图像数据输入到同一个深度卷积神经网络，经过计算后获得多个分类结果；或者，将深度卷积神经网络不同层的输出结果输入到分类器中，经过计算后获得多个分类结果，本发明实施例不对多个分类结果的获取方式进行限制。
在本实例中，第一计算单元101计算至少两个分类结果中任意两个分类结果之间的差异度，其中，该差异度可使用现有方法来表征，例如，可以通过计算任意两个分类结果之间的Jousselme距离来表征该差异度。在本实施例中，以Jousselme距离为例，来表征任意两个分类结果之间的差异度。
在本实施例中，假设共有n个分类结果，n为大于等于2的整数，对n个分类结果进行排序，分别为m1,m2,……,mn，其中，可以按照任意的顺序对n个分类结果进行排序。
在本实例中，例如，可根据以下的公式(1)和(2)计算任意两个分类结果之间的Jousselme距离(以下简称为J距离)：
Figure PCTCN2016070473-appb-000001
Figure PCTCN2016070473-appb-000002
其中，dBPA(i,j)表示第i个分类结果与第j个分类结果之间的J距离，mi表示第i个分类结果，mj表示第j个分类结果，Ap和Aq分别表示第p个类别和第q个类别，i,j,p,q,n均为正整数，且i,j,p,q均小于或等于n。
在本实施例中，第二计算单元102根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度，例如，可根据以下的公式(3)计算各个分类结果：
Figure PCTCN2016070473-appb-000003
其中，Sum_m(i)表示第i个分类结果与其他分类结果之间的平均J距离，dBPA(i,j)表示第i个分类结果与第j个分类结果之间的J距离。
在本实施例中，在获得各个分类结果与其他分类结果的平均差异度之后，融合单元103根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得至少两个分类结果的融合结果。
例如，融合单元103根据各个分类结果与其他分类结果之间的平均差异度，对各个分类结果与该分类结果的前一个分类结果或者该分类结果之前的所有分类结果的融合结果进行合成，并对合成结果进行融合，获得至少两个分类结果的融合结果。
例如，对于前两个分类结果，融合单元103对第1分类结果和第2分类结果进行合成，对于第3分类结果，融合单元103对第3分类结果和前两个分类结果的融合结果进行合成，对于第3分类结果之后的分类结果，以此类推。
在本实施例中，融合单元103可根据各个分类结果与其他分类结果之间的平均差异度和预设阈值的比较结果设定用于进行合成的权值，根据该权值对各个分类结果进行合成，并对合成结果进行融合，获得至少两个分类结果的融合结果。例如，当某个 分类结果与其他分类结果之间的平均差异度大于该预设阈值时，将该分类结果在合成时使用的权值设置为较小。
在本实施例中，该预设阈值可根据实际需要而设置，例如，将该阈值设置为0～1之间的数值，例如该阈值为0.5。
以下对本实施例的融合单元103的构成以及根据平均差异度进行合成和融合的方法进行示例性的说明。
图2是本发明实施例1的融合单元103的示意图。当需要对n个分类结果中的前两个分类结果，即第1分类结果m1和第2分类结果m2进行合成和融合时，如图2所示，融合单元103包括：
第一设定单元201，用于当第1分类结果与其他分类结果之间的第1平均差异度大于预设阈值且第2分类结果与其他分类结果之间的第2平均差异度小于该预设阈值时，根据第1平均差异度计算第1分类结果的第1权值，根据第1权值计算第2分类结果的第2权值，其中，第1权值小于第2权值；
第二设定单元202，用于当第1分类结果与其他分类结果之间的第1平均差异度小于所述预设阈值且第2分类结果与其他分类结果之间的第2平均差异度大于该预设阈值时，根据第2平均差异度计算第2分类结果的第2权值，根据第2权值计算第1分类结果的第1权值，其中，第2权值小于第1权值；
第三设定单元203，用于当第1分类结果与其他分类结果之间的第1平均差异度和第2分类结果与其他分类结果之间的第2平均差异度均大于或等于该预设阈值或者均小于该预设阈值时，将第1分类结果的第1权值与第2分类结果的第2权值设置为相等；
第一合成单元204，用于根据第1权值和所述第2权值，对第1分类结果和第2分类结果进行合成；
第一融合单元205，用于对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果。
在本实施例中，当第1分类结果与其他分类结果之间的第1平均差异度大于预设阈值且第2分类结果与其他分类结果之间的第2平均差异度小于该预设阈值时，第一设定单元201根据第1平均差异度计算第1分类结果的第1权值，例如，可使用以下的公式(4)计算第1分类结果的第1权值：
Figure PCTCN2016070473-appb-000004
其中，
Figure PCTCN2016070473-appb-000005
表示第1权值，Sum_m(1)表示第1分类结果的第1平均差异度，α表示权值调节因子。其中，α可根据实际需要而设置，例如，α为2～5的数值。
在计算出第1权值之后，第一设定单元201根据第1权值计算第2分类结果的第2权值，例如，将
Figure PCTCN2016070473-appb-000006
作为第2权值。其中，
Figure PCTCN2016070473-appb-000007
即
Figure PCTCN2016070473-appb-000008
在本实施例中，当第1分类结果与其他分类结果之间的第1平均差异度小于该预设阈值且第2分类结果与其他分类结果之间的第2平均差异度大于该预设阈值时，第二设定单元202根据第2平均差异度计算第2分类结果的第2权值，例如，可使用与上述公式(4)类似的公式(5)计算第2分类结果的第2权值：
Figure PCTCN2016070473-appb-000009
其中，
Figure PCTCN2016070473-appb-000010
表示第2权值，Sum_m(2)表示第2分类结果的第2平均差异度，α表示权值调节因子，其取值同前。
在计算出第2权值之后，第二设定单元202根据第2权值计算第1分类结果的第1权值，例如，将
Figure PCTCN2016070473-appb-000011
作为第1权值。其中，
Figure PCTCN2016070473-appb-000012
即
Figure PCTCN2016070473-appb-000013
在本实施例中，当第1分类结果与其他分类结果之间的第1平均差异度和第2分类结果与其他分类结果之间的第2平均差异度均大于或等于该预设阈值或者均小于该预设阈值时，第三设定单元203将第1分类结果的第1权值与第2分类结果的第2权值设置为相等，例如，将第1权值和第2权值均设为0.5。
在本实施例中，第一合成单元204根据计算出的第1权值和所述第2权值，对第1分类结果和第2分类结果进行合成，例如，
对于第一设定单元201设定的第1权值和第2权值，可根据以下的公式(6)进行合成：
Figure PCTCN2016070473-appb-000014
其中，
Figure PCTCN2016070473-appb-000015
表示第1分类结果和第2分类结果的合成结果，m1表示第1分类结果，m2表示第2分类结果，
Figure PCTCN2016070473-appb-000016
表示第1权值，
Figure PCTCN2016070473-appb-000017
表示第2权值。
对于第二设定单元202设定的第1权值和第2权值，可根据以下的公式(7)进行合成：
Figure PCTCN2016070473-appb-000018
其中，
Figure PCTCN2016070473-appb-000019
表示第1分类结果和第2分类结果的合成结果，m1表示第1分类结 果，m2表示第2分类结果，
Figure PCTCN2016070473-appb-000020
表示第1权值，
Figure PCTCN2016070473-appb-000021
表示第2权值。
对于第三设定单元203设定的第1权值和第2权值，可根据以下的公式(8)进行合成：
Figure PCTCN2016070473-appb-000022
其中，
Figure PCTCN2016070473-appb-000023
表示第1分类结果和第2分类结果的合成结果，m1表示第1分类结果，m2表示第2分类结果。
在本实施例中，在获得第1分类结果和第2分类结果的合成结果之后，第一融合单元205对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果，其中，可使用现有的融合方法，将合成结果与其自身进行融合，例如，
Figure PCTCN2016070473-appb-000024
其中，
Figure PCTCN2016070473-appb-000025
表示第1分类结果和第2分类结果的融合结果，
Figure PCTCN2016070473-appb-000026
表示第1分类结果和第2分类结果的合成结果。
在本实施例中，对合成结果进行融合可使用现有的融合方法，例如，可基于D-S(Dempster/Shafer)证据理论进行融合。
以上是对前两个分类结果进行合成和融合的示例性说明，当n≥3时，融合单元103还可以包括：
第四设定单元206，用于当第i分类结果与其他分类结果之间的第i平均差异度大于该预设阈值时，根据第i平均差异度计算第i分类结果的第i权值，根据第i权值计算第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，其中，第i权值小于第i-1权值，i为大于或等于3的整数，且i≤n；
第五设定单元207，用于当第i分类结果与其他分类结果之间的第i平均差异度小于或等于该预设阈值时，根据预设参数来设置第i分类结果的第i权值以及第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；
在本实施例中，当第i分类结果与其他分类结果之间的第i平均差异度大于该预设阈值时，第四设定单元206根据第i平均差异度计算第i分类结果的第i权值，例如，可使用与上述公式(4)类似的公式(10)计算第i权值：
Figure PCTCN2016070473-appb-000027
其中，
Figure PCTCN2016070473-appb-000028
表示第i分类结果的第i权值，Sum_m(i)表示第i分类结果的第i平均 差异度，α表示权值调节因子，其取值同前。
在计算出第i权值之后，第四设定单元206根据第i权值计算第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，例如，将
Figure PCTCN2016070473-appb-000029
作为第i-1权值。其中，
Figure PCTCN2016070473-appb-000030
即
Figure PCTCN2016070473-appb-000031
在本实施例中，当第i分类结果与其他分类结果之间的第i平均差异度小于或等于该预设阈值时，第五设定单元207根据预设参数来设置第i分类结果的第i权值以及第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，例如，将第i权值设为β，那么第i-1权值则为(1-β)。其中，β为0～1的数值，可根据实际需要而设置。
第一合成单元204还用于根据第i权值和第i-1权值，对第i分类结果和第i分类结果之前的i-1个分类结果的融合结果进行合成，例如，
对于第四设定单元206设定的第i权值和第i-1权值，可根据以下的公式(11)进行合成：
Figure PCTCN2016070473-appb-000032
其中，
Figure PCTCN2016070473-appb-000033
表示第i分类结果和第i分类结果之前的i-1个分类结果的融合结果的合成结果，mi表示第i分类结果，
Figure PCTCN2016070473-appb-000034
表示第i分类结果之前的i-1个分类结果的融合结果，
Figure PCTCN2016070473-appb-000035
表示第i权值，
Figure PCTCN2016070473-appb-000036
表示第i-1权值。
例如，当i＝3时，
Figure PCTCN2016070473-appb-000037
表示第3分类结果m3和第1分类结果、第2分类结果的融合结果
Figure PCTCN2016070473-appb-000038
的合成结果。
对于第五设定单元207设定的第i权值和第i-1权值，可根据以下的公式(12)进行合成：
Figure PCTCN2016070473-appb-000039
其中，
Figure PCTCN2016070473-appb-000040
表示第i分类结果和第i分类结果之前的i-1个分类结果的融合结果的合成结果，mi表示第i分类结果，
Figure PCTCN2016070473-appb-000041
表示第i分类结果之前的i-1个分类结果的融合结果，β表示第i权值，(1-β)表示第i-1权值。
在本实施例中，在获得第i分类结果和第i分类结果之前的i-1个分类结果的融合结果的合成结果之后，第一融合单元205还用于对第i分类结果和第i分类结果之前的i-1个分类结果的融合结果的合成结果进行融合，获得i个分类结果的融合结果。其中，融合方法与前面的记载相同，此处不再赘述。
在本实施例中，当n≥3时，融合单元103进行以上的迭代计算，最终获得n个 分类结果的融合结果。
在本实施例中，由上述实施例可知，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
实施例2
本发明实施例还提供了一种电子设备，图3是本发明实施例2的电子设备的示意图。如图3所示，电子设备300包括分类结果的融合装置301，其中，分类结果的融合装置301的结构和功能与实施例1中的记载相同，此处不再赘述。
图4是本发明实施例2的电子设备的系统构成的一示意框图。如图4所示，电子设备400可以包括中央处理器401和存储器402；存储器402耦合到中央处理器401。该图是示例性的；还可以使用其它类型的结构，来补充或代替该结构，以实现电信功能或其它功能。
如图4所示，该电子设备400还可以包括：输入单元403、显示器404、电源405。
在一个实施方式中，实施例1所述的分类结果的融合装置的功能可以被集成到中央处理器401中。其中，中央处理器401可以被配置为：计算至少两个分类结果中任意两个分类结果之间的差异度；根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
其中，所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，包括：根据各个分类结果与其他分类结果之间的平均差异度和预设阈值的比较结果设定用于进行合成的权值，根据所述权值对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
其中，所述至少两个分类结果包括第1分类结果和第2分类结果，所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，包括：当第1分类结果与其他分类结果之间的第1平均差异度大于预设阈值且第2分类结果与其他分类结果之间的第2平均差异度小于所述预设阈值时，根据所述第1平均差异度计算所述第1分类结果的第1权值，根据所述第1权值计算所述第2分类结果的第2权值，其中，所述第1 权值小于所述第2权值；当第1分类结果与其他分类结果之间的第1平均差异度小于所述预设阈值且第2分类结果与其他分类结果之间的第2平均差异度大于所述预设阈值时，根据所述第2平均差异度计算所述第2分类结果的第2权值，根据所述第2权值计算所述第1分类结果的第1权值，其中，所述第2权值小于所述第1权值；当第1分类结果与其他分类结果之间的第1平均差异度和第2分类结果与其他分类结果之间的第2平均差异度均大于或等于所述预设阈值或者均小于所述预设阈值时，将第1分类结果的第1权值与第2分类结果的第2权值设置为相等；根据所述第1权值和所述第2权值，对第1分类结果和第2分类结果进行合成；对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果。
其中，所述至少两个分类结果包括第1分类结果和第2分类结果在内的n个分类结果，n为大于或等于3的整数，所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，还包括：当第i分类结果与其他分类结果之间的第i平均差异度大于所述预设阈值时，根据所述第i平均差异度计算第i分类结果的第i权值，根据所述第i权值计算所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，其中，所述第i权值小于所述第i-1权值，i为大于或等于3的整数，且i≤n；当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，根据预设参数来设置第i分类结果的第i权值以及所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；根据所述第i权值和第i-1权值，对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果进行合成；对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果的合成结果进行融合，获得i个分类结果的融合结果。
其中，当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，将所述预设参数设置为使得所述第i权值与所述第i-1权值相等。
在另一个实施方式中，实施例1所述的分类结果的融合装置可以与中央处理器401分开配置，例如可以将分类结果的融合装置配置为与中央处理器401连接的芯片，通过中央处理器401的控制来实现分类结果的融合装置的功能。
在本实施例中电子设备400也并不是必须要包括图4中所示的所有部件。
如图4所示，中央处理器401有时也称为控制器或操作控件，可以包括微处理器或其它处理器装置和/或逻辑装置，中央处理器401接收输入并控制电子设备400的 各个部件的操作。
存储器402，例如可以是缓存器、闪存、硬驱、可移动介质、易失性存储器、非易失性存储器或其它合适装置中的一种或更多种。并且中央处理器401可执行该存储器402存储的该程序，以实现信息存储或处理等。其它部件的功能与现有类似，此处不再赘述。电子设备400的各部件可以通过专用硬件、固件、软件或其结合来实现，而不偏离本发明的范围。
在本实施例中，由上述实施例可知，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
实施例3
本发明实施例还提供一种分类结果的融合方法，其对应于实施例1的分类结果的融合装置。图5是本发明实施例3的分类结果的融合方法的流程图。如图5所示，该方法包括：
步骤501：计算至少两个分类结果中任意两个分类结果之间的差异度；
步骤502：根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；
步骤503：根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得至少两个分类结果的融合结果。
在本实施例中，计算该差异度以及平均差异度、根据平均差异度进行合成和融合的方法与实施例1中的记载相同，此处不再赘述。
由上述实施例可知，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
实施例4
本发明实施例还提供一种分类结果的融合方法，其对应于实施例1的分类结果的融合装置。图6是本发明实施例4的分类结果的融合方法的流程图。如图6所示，该方法包括：
步骤601：计算n个分类结果中任意两个分类结果之间的差异度，n为大于或等于2的整数；
步骤602：根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；
步骤603：判断n个分类结果中第1分类结果m1的第1平均差异度Sum_m(1)和第2分类结果m2的第2平均差异度Sum_m(2)是否满足以下条件：Sum_m(1)>th_d且Sum_m(2)<th_d，th_d为预设阈值。当判断结果为“是”时，进入步骤604，当判断结果为“否”时，进入步骤605；
步骤604：根据第1平均差异度计算第1分类结果的第1权值，根据第1权值计算第2分类结果的第2权值；
步骤605：判断第1分类结果m1的第1平均差异度Sum_m(1)和第2分类结果m2的第2平均差异度Sum_m(2)是否满足以下条件：Sum_m(1)<th_d且Sum_m(2)>th_d，th_d为预设阈值。当判断结果为“是”时，进入步骤606，当判断结果为“否”时，进入步骤607；
步骤606：根据第2平均差异度计算第2分类结果的第2权值，根据第2权值计算第1分类结果的第1权值；
步骤607：，将第1分类结果的第1权值与第2分类结果的第2权值设置为相等；
步骤608：根据第1权值和第2权值，对第1分类结果和第2分类结果进行合成；
步骤609：对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果；
步骤610：判断n是否大于或等于3，当判断结果为“是”时，进入步骤611，当判断结果为“否”时，则结束进程；
步骤611：判断第i分类结果mi的第i平均差异度Sum_m(i)是否大于该预设阈值th_d，当判断结果为“是”时，则进入步骤612，当判断结果为“否”时，则进入步骤613，i为大于或等于3的整数，且i≤n；
步骤612：根据第i平均差异度计算第i分类结果的第i权值，根据第i权值计算第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；
步骤613：根据预设参数来设置第i分类结果的第i权值以及第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；
步骤614：根据第i权值和第i-1权值，对第i分类结果和第i分类结果之前的i-1个分类结果的融合结果进行合成；
步骤615：对第i分类结果和第i分类结果之前的i-1个分类结果的融合结果的合成结果进行融合，获得i个分类结果的融合结果；
步骤616：判断i是否小于n，当判断结果为“是”时，进入步骤617，当判断结果为“否”时，结束进程；
步骤617：将i加上1。
在本实施例中，在上述各个步骤中使用的计算和设定方法与实施例1中的记载相同，此处不再赘述。
由上述实施例可知，根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成和融合，能够有效提高分类结果的准确度。
本发明实施例还提供一种计算机可读程序，其中当在分类结果的融合装置或电子设备中执行所述程序时，所述程序使得计算机在所述分类结果的融合装置或电子设备中执行实施例3或实施例4所述的分类结果的融合方法。
本发明实施例还提供一种存储有计算机可读程序的存储介质，其中所述计算机可读程序使得计算机在分类结果的融合装置或电子设备中执行实施例3或实施例4所述的分类结果的融合方法。
本发明以上的装置和方法可以由硬件实现，也可以由硬件结合软件实现。本发明涉及这样的计算机可读程序，当该程序被逻辑部件所执行时，能够使该逻辑部件实现上文所述的装置或构成部件，或使该逻辑部件实现上文所述的各种方法或步骤。本发明还涉及用于存储以上程序的存储介质，如硬盘、磁盘、光盘、DVD、flash存储器等。
以上结合具体的实施方式对本发明进行了描述，但本领域技术人员应该清楚，这些描述都是示例性的，并不是对本发明保护范围的限制。本领域技术人员可以根据本发明的精神和原理对本发明做出各种变型和修改，这些变型和修改也在本发明的范围内。

Claims (11)

  1. 一种分类结果的融合装置，所述装置包括：
    第一计算单元，所述第一计算单元用于计算至少两个分类结果中任意两个分类结果之间的差异度；
    第二计算单元，所述第二计算单元用于根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；
    融合单元，所述融合单元用于根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
  2. 根据权利要求1所述的装置，其中，
    所述融合单元用于根据各个分类结果与其他分类结果之间的平均差异度和预设阈值的比较结果设定用于进行合成的权值，根据所述权值对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
  3. 根据权利要求1或2所述的装置，其中，所述至少两个分类结果包括第1分类结果和第2分类结果，
    所述融合单元包括：
    第一设定单元，所述第一设定单元用于当第1分类结果与其他分类结果之间的第1平均差异度大于预设阈值且第2分类结果与其他分类结果之间的第2平均差异度小于所述预设阈值时，根据所述第1平均差异度计算所述第1分类结果的第1权值，根据所述第1权值计算所述第2分类结果的第2权值，其中，所述第1权值小于所述第2权值；
    第二设定单元，所述第二设定单元用于当第1分类结果与其他分类结果之间的第1平均差异度小于所述预设阈值且第2分类结果与其他分类结果之间的第2平均差异度大于所述预设阈值时，根据所述第2平均差异度计算所述第2分类结果的第2权值，根据所述第2权值计算所述第1分类结果的第1权值，其中，所述第2权值小于所述第1权值；
    第三设定单元，所述第三设定单元用于当第1分类结果与其他分类结果之间的第1平均差异度和第2分类结果与其他分类结果之间的第2平均差异度均大于或等于所 述预设阈值或者均小于所述预设阈值时，将第1分类结果的第1权值与第2分类结果的第2权值设置为相等；
    第一合成单元，所述第一合成单元用于根据所述第1权值和所述第2权值，对第1分类结果和第2分类结果进行合成；
    第一融合单元，所述第一融合单元用于对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果。
  4. 根据权利要求3所述的装置，其中，所述至少两个分类结果包括第1分类结果和第2分类结果在内的n个分类结果，n为大于或等于3的整数，
    所述融合单元还包括：
    第四设定单元，所述第四设定单元用于当第i分类结果与其他分类结果之间的第i平均差异度大于所述预设阈值时，根据所述第i平均差异度计算第i分类结果的第i权值，根据所述第i权值计算所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，其中，所述第i权值小于所述第i-1权值，i为大于或等于3的整数，且i≤n；
    第五设定单元，所述第五设定单元用于当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，根据预设参数来设置第i分类结果的第i权值以及所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；
    并且，所述第一合成单元还用于根据所述第i权值和第i-1权值，对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果进行合成；
    所述第一融合单元还用于对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果的合成结果进行融合，获得i个分类结果的融合结果。
  5. 根据权利要求4所述的装置，其中，所述第五设定单元用于当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，将所述预设参数设置为使得所述第i权值与所述第i-1权值相等。
  6. 一种电子设备，包括根据权利要求1-5的任一项所述的分类结果的融合装置。
  7. 一种分类结果的融合方法，所述方法包括：
    计算至少两个分类结果中任意两个分类结果之间的差异度；
    根据任意两个分类结果之间的差异度，计算各个分类结果与其他分类结果之间的平均差异度；
    根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
  8. 根据权利要求7所述的方法，其中，所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，包括：
    根据各个分类结果与其他分类结果之间的平均差异度和预设阈值的比较结果设定用于进行合成的权值，根据所述权值对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果。
  9. 根据权利要求7或8所述的方法，其中，所述至少两个分类结果包括第1分类结果和第2分类结果，
    所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，包括：
    当第1分类结果与其他分类结果之间的第1平均差异度大于预设阈值且第2分类结果与其他分类结果之间的第2平均差异度小于所述预设阈值时，根据所述第1平均差异度计算所述第1分类结果的第1权值，根据所述第1权值计算所述第2分类结果的第2权值，其中，所述第1权值小于所述第2权值；
    当第1分类结果与其他分类结果之间的第1平均差异度小于所述预设阈值且第2分类结果与其他分类结果之间的第2平均差异度大于所述预设阈值时，根据所述第2平均差异度计算所述第2分类结果的第2权值，根据所述第2权值计算所述第1分类结果的第1权值，其中，所述第2权值小于所述第1权值；
    当第1分类结果与其他分类结果之间的第1平均差异度和第2分类结果与其他分类结果之间的第2平均差异度均大于或等于所述预设阈值或者均小于所述预设阈值时，将第1分类结果的第1权值与第2分类结果的第2权值设置为相等；
    根据所述第1权值和所述第2权值，对第1分类结果和第2分类结果进行合成；
    对第1分类结果和第2分类结果的合成结果进行融合，获得第1分类结果和第2分类结果的融合结果。
  10. 根据权利要求9所述的方法，其中，所述至少两个分类结果包括第1分类结果和第2分类结果在内的n个分类结果，n为大于或等于3的整数，
    所述根据各个分类结果与其他分类结果之间的平均差异度对各个分类结果进行 合成，并对合成结果进行融合，获得所述至少两个分类结果的融合结果，还包括：
    当第i分类结果与其他分类结果之间的第i平均差异度大于所述预设阈值时，根据所述第i平均差异度计算第i分类结果的第i权值，根据所述第i权值计算所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值，其中，所述第i权值小于所述第i-1权值，i为大于或等于3的整数，且i≤n；
    当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，根据预设参数来设置第i分类结果的第i权值以及所述第i分类结果之前的i-1个分类结果的融合结果的第i-1权值；
    根据所述第i权值和第i-1权值，对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果进行合成；
    对第i分类结果和所述第i分类结果之前的i-1个分类结果的融合结果的合成结果进行融合，获得i个分类结果的融合结果。
  11. 根据权利要求10所述的方法，其中，当第i分类结果与其他分类结果之间的第i平均差异度小于或等于所述预设阈值时，将所述预设参数设置为使得所述第i权值与所述第i-1权值相等。
PCT/CN2016/070473 2016-01-08 2016-01-08 分类结果的融合装置及方法 Ceased WO2017117799A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/070473 WO2017117799A1 (zh) 2016-01-08 2016-01-08 分类结果的融合装置及方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/070473 WO2017117799A1 (zh) 2016-01-08 2016-01-08 分类结果的融合装置及方法

Publications (1)

Publication Number Publication Date
WO2017117799A1 true WO2017117799A1 (zh) 2017-07-13

Family

ID=59273355

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/070473 Ceased WO2017117799A1 (zh) 2016-01-08 2016-01-08 分类结果的融合装置及方法

Country Status (1)

Country Link
WO (1) WO2017117799A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050144147A1 (en) * 2003-12-26 2005-06-30 Lee Shih-Jong J. Feature regulation for hierarchical decision learning
CN103605990A (zh) * 2013-10-23 2014-02-26 江苏大学 基于图聚类标签传播的集成多分类器融合分类方法和系统
CN103810482A (zh) * 2014-03-12 2014-05-21 中国矿业大学(北京) 一种多信息融合分类及识别方法
CN105160249A (zh) * 2015-07-02 2015-12-16 哈尔滨工程大学 一种基于改进的神经网络集成的病毒检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050144147A1 (en) * 2003-12-26 2005-06-30 Lee Shih-Jong J. Feature regulation for hierarchical decision learning
CN103605990A (zh) * 2013-10-23 2014-02-26 江苏大学 基于图聚类标签传播的集成多分类器融合分类方法和系统
CN103810482A (zh) * 2014-03-12 2014-05-21 中国矿业大学(北京) 一种多信息融合分类及识别方法
CN105160249A (zh) * 2015-07-02 2015-12-16 哈尔滨工程大学 一种基于改进的神经网络集成的病毒检测方法

Similar Documents

Publication Publication Date Title
CN115115552B (zh) 图像矫正模型训练及图像矫正方法、装置和计算机设备
US10019823B2 (en) Combined composition and change-based models for image cropping
CN110633610B (zh) 一种基于yolo的学员状态检测方法
US9299004B2 (en) Image foreground detection
US9152926B2 (en) Systems, methods, and media for updating a classifier
WO2019091402A1 (zh) 年龄预估方法和装置
CN111667501A (zh) 目标跟踪方法、装置、计算设备及存储介质
CN110826558B (zh) 图像分类方法、计算机设备和存储介质
CN112927279A (zh) 一种图像深度信息生成方法、设备及存储介质
CN110858286A (zh) 一种用于目标识别的图像处理方法及装置
CN114463603B (zh) 图像检测模型的训练方法、装置、电子设备及存储介质
CN113129335B (zh) 一种基于孪生网络的视觉跟踪算法及多模板更新策略
KR20170125720A (ko) 심층 신경망 기반의 인식 장치, 트레이닝 장치, 및 이들의 방법
CN116883781B (zh) 预训练模型的微调方法、装置、电子设备及存储介质
US11645579B2 (en) Automated machine learning tagging and optimization of review procedures
CN110895802B (zh) 一种图像处理方法及装置
CN112597995B (zh) 车牌检测模型训练方法、装置、设备及介质
CN109558901A (zh) 一种语义分割训练方法及装置、电子设备、存储介质
WO2020118588A1 (zh) 图像级jnd阈值的预测方法、装置、设备及存储介质
CN113706428B (zh) 一种图像生成方法及装置
CN112149756A (zh) 模型训练方法、图像识别方法、装置、设备及存储介质
CN112861870B (zh) 指针式仪表图像矫正方法、系统及存储介质
CN111832349A (zh) 遗留物错误检测的识别方法、装置及图像处理设备
CN114723646A (zh) 带标注的图像数据生成方法、装置、存储介质及电子设备
WO2024021504A1 (zh) 人脸识别模型训练方法、识别方法、装置、设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16882944

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 16882944

Country of ref document: EP

Kind code of ref document: A1