The content of the invention
For a kind of not enough present in " high definition " printed words detection technique, video high definition printed words detection method of the invention, lead to
The mode for crossing the position candidate for precalculating " high definition " printed words determines the interval of characteristics of image to be extracted, to the interval using intersection
The method of checking and " difficult example excavation " calculates the convergent determinant optimized parameter of accuracy for causing SVM result of determination, and then
SVM judgements are carried out to all pictures in video by the parameter set, while effectively improving " high definition " printed words correct judgment, is carried
The fuzzy compatibility ability of high " high definition " printed words produced to different radio station, substantially increases the speed and effect of HD video detection
Rate.
The technical solution adopted for the present invention to solve the technical problems is comprised the following steps:
Position candidate step is calculated, the position candidate of " high definition " printed words in inputted video image is determined using ad hoc approach.
Preferably, the ad hoc approach includes:Slip window sampling and the upper picture position candidate method of memory.
Preferably, if the first two field picture or previous frame image are not detected by " high definition " printed words, then sliding window is used
Method determines position candidate;Otherwise, position candidate is determined jointly using two methods.
Image characteristic step is extracted, the picture region to " high definition " printed words position candidate extracts characteristics of image.
Preferably, described image feature includes space domain characteristic and frequency domain character.
Preferably, the space domain characteristic use direction histogram of gradients(Histogram of Oriented
Gradient , HOG)Feature;The frequency domain character passes through DFT(Discrete Cosine
Transformation , DCT)Extract.
Determinant optimized parameter step is calculated, the picture that known result of determination is input under off-line mode carries out " difficult example to SVM
Excavate " training study, draw determinant optimized parameter.
Preferably, the positive and negative data sample for known result of determination is input into, Gaussian kernel-SVM vector parameter collection is output as, is held
Line mode is cross validation.
Preferably, the positive negative sample is respectively the picture feature containing " high definition " printed words and the figure without " high definition " printed words
Piece feature, and by 1:1 ratio is delivered.
Preferably, after each " the difficult example excavation " training study terminates, the negative sample that will be mistaken for positive sample adds
Enter the negative sample pond of " difficult example excavation " next time, to improve judgement accuracy.
Preferably, the condition that described " difficult example excavation " study terminates is that the parameter set being calculated using the last time is carried out
SVM judges that the result of determination accuracy for obtaining restrains.
Online determination step, using the SVM determinant optimized parameters specified to video in all images carry out " high definition "
Printed words detection judges.
Preferably, the optimal critical parameter for being calculated according to the calculating determinant optimized parameter step, in video
All images carry out decision through the characteristics of image that the extraction image characteristic step is extracted, if in the presence of a position candidate bag
Containing " high definition " printed words, then it is assumed that the image has " high definition " printed words.
Using above-mentioned technical proposal, the present invention has advantages below:
The present invention relates to a kind of video high definition printed words detection method, by way of precalculating the position candidate of " high definition " printed words
Determine the interval of characteristics of image to be extracted, the interval is calculated so that SVM using cross validation and the method for " difficult example excavation "
The convergent determinant optimized parameter of accuracy of result of determination, and then SVM is carried out to all pictures in video by the parameter set
Judge, while effectively increasing the correctness of " high definition " judgement, improve the mould of " high definition " printed words produced to different radio station
Paste compatibility, substantially increases the speed and efficiency of HD video detection.
Specific embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete
Whole description, it is clear that described embodiment is only one embodiment of the present of invention, rather than whole embodiments.Based on this
Embodiment in invention, other realities that those of ordinary skill in the art are obtained on the premise of creative work is not made
Example is applied, the scope of protection of the invention is belonged to.
The embodiment of the invention discloses a kind of video high definition printed words detection method, shown in Figure 1, the method includes:
Step S1:The position candidate of " high definition " printed words is obtained from every two field picture;
Step S2:Extract the characteristics of image of " high definition " printed words position candidate;
Step S3:" difficult example excavation " is carried out according to the setting ratio positive negative sample of input draw determinant optimized parameter;
Step S4:Characteristics of image to being extracted carries out SVM differentiations, determines whether picture contains " high definition " printed words.
In the embodiment of the present invention, video figure is calculated by the method to sliding window and memory previous frame image candidate position
The position candidate that " high definition " printed words occur as in, and then the provincial characteristics and frequency domain character of the area image are extracted as " high definition "
The input that printed words judge, obtains determinant optimized parameter, finally by the ginseng by the method for " difficult example excavation " cross validation
Manifold using Gaussian kernel-SVM to video in every two field picture carry out " high definition " printed words judge.
It can be seen that, in the embodiment of the present invention, calculated by the method for cross validation and " difficult example excavation " so that SVM judges knot
The convergent determinant optimized parameter of accuracy of fruit, and then SVM judgements are carried out to all pictures in video by the parameter set, have
While effect improves the correctness of " high definition " judgement, the fuzzy compatibility energy of " high definition " printed words produced to different radio station is improve
Power, substantially increases the speed and efficiency of HD video detection.
The embodiment of the invention discloses a kind of video high definition printed words detection method, referring to Fig. 2, an embodiment is gone up relatively, this
Embodiment has made further instruction and optimization to technical scheme.Specifically, a kind of video high definition printed words detection in the present embodiment
Method is comprised the steps of:
S1:Calculate position candidate.
Preferably, in the step, the position candidate of " high definition " printed words in inputted video image is determined using ad hoc approach.
Preferably, when first video image is input into or previous frame image is not detected by " high definition " printed words, step is performed
Rapid S11 calculates the position candidate of " high definition " printed words using active window method.
Preferably, when a upper image is detected in the presence of " high definition " printed words position candidate, step S12 is performed by memory
" high definition " printed words position candidate method of previous frame image simultaneously determines " high definition " printed words candidate bit of this picture with reference to S11 jointly
Put.
S2:Extract characteristics of image.
Preferably, in the step, the picture region to " high definition " printed words position candidate extracts characteristics of image.
Preferably, space domain characteristic is extracted by performing step S21 histograms of oriented gradients.
Preferably, by perform step S22 will in step S1 obtain position candidate image down into 32*32 sizes, enter
And performing step S23 carries out two-dimensional dct transform to gained image, finally perform step S24 and conversion acquired results are extracted
8*8 frequency domain characters.
S3:Calculate determinant optimized parameter.
Preferably, the picture of known result of determination is input into the step, under off-line mode to SVM, by cross validation
Mode carry out " difficult example excavation " training study, draw determinant optimized parameter.
Preferably, by performing step S31 by the area comprising the area image of " high definition " printed words and without " high definition " printed words
Area image is according to 1:1 ratio input SVM carries out " difficult example excavation " and judges training.
Preferably, the result for obtaining is performed for step S31, is performed during step S32 extracts result and is mistaken for positive sample
Negative sample then perform step S33.
Preferably, by performing the Gaussian kernel-SVM vector machine parameters that step S33 is produced come collection step S31, and then hold
Row step S34.
Preferably, the SVM result of determination for being obtained after step S34 judges that step S31 is performed, if the result is not received
Hold back, then the negative sample that will be extracted in step S32 adds the implementation procedure of repeat step S31 to step S34 behind new negative sample pond
Persistently amendment step S31 performs the Gaussian kernel-SVM vector machine parameter sets for obtaining, until the result of determination that step S34 is obtained is received
Hold back.
S4:It is online to judge.
Preferably, in the step, using the SVM determinant optimized parameters specified to video in all images carry out
The detection of " high definition " printed words judges.
Preferably, when the result of determination convergence that step S34 is obtained, then step S4, the Gauss being collected into using S34 are performed
Core-SVM vector machine parameter sets, to all images of target video by after step S1 and S2 to characteristics of image carry out online
SVM " high definition " printed words are detected, for each two field picture, if including " high definition " printed words in the presence of a position candidate, then it is assumed that the frame
Image includes " high definition " printed words.
In sum, " high definition " printed words position candidate is determined by performing step S1, specifically firstly for video figure
First frame or former frame of picture are that the situation for detecting " high definition " printed words performs step S11 using slip window sampling calculating image
The position candidate of " high definition " printed words, otherwise performs step S12 memories previous frame image " high definition " printed words position candidate method tentatively true
After determining position candidate, further perform step S11 to determine the position candidate of " high definition " printed words jointly, for " high definition " word
Sample position candidate image-region performs step S2, extracts candidate region characteristics of image, specific first-selected execution step S21 use sides
Space domain characteristic is extracted to histogram of gradients, then performing step S22 will perform after image-region boil down to 32*32 format sizes
Step S23 carries out two-dimensional dct transform to the compressed images, and finally performing step S24 to transformation results extracts the low of image
Frequency 8*8 coefficients, the result for obtaining are performed to step S2 and perform step S3, and the image of the known result of determination of selection carries out " difficult example digging
Pick " draws critical parameter, specifically first carries out step S31, will be containing the characteristics of image of " high definition " printed words and without " high definition " word
The characteristics of image of sample is respectively as positive and negative sample set according to 1:1 ratio input Gaussian kernel-SVM carries out " difficult example excavation " training, so
The negative sample of positive sample is mistaken in execution step S32 extraction step S31 result of determination afterwards, and negative sample injection is new
Negative sample pond, and then perform the Gaussian kernel-SVM vector machine parameters that step S33 collection steps S31 is produced, finally perform step
S34, judges whether the accuracy of S31 implementing results restrains, if not converged, the new negative sample pond that step S32 is produced is made
For new negative sample input repeats step S31 to S34 untill the implementing result accuracy of S31 restrains, S3 has been performed
Finally perform S4, using the Gaussian kernel-SVM vector machine parameters that S3 is produced, to all images of target video by step S1 and
After S2 to characteristics of image carry out online SVM " high definition " printed words detection, for each two field picture, if in the presence of a position candidate
Comprising " high definition " printed words, then it is assumed that the two field picture includes " high definition " printed words.By the position candidate for precalculating " high definition " printed words
Mode determine the interval of characteristics of image to be extracted, the interval is calculated using cross validation and the method for " difficult example excavation " and is sent as an envoy to
The convergent determinant optimized parameter of accuracy of SVM result of determination is obtained, and then all pictures in video are entered by the parameter set
Row SVM judges, while effectively increasing the correctness that " high definition " judges, improves " high definition " printed words produced to different radio station
Fuzzy compatibility ability, substantially increase HD video detection speed and efficiency.
The foregoing is only illustrative, rather than for restricted.Those skilled in the art can carry out various changing to invention
Dynamic and modification is without departing from the spirit and scope of the present invention.So, if these modifications of the invention and modification belong to the present invention
Within the scope of claim and its equivalent technologies, then the present invention is also intended to including including these changes and modification.