CN101937507B - Wavelet feature extraction method for low-contrast vehicle image - Google Patents

Abstract

The invention provides a wavelet feature extraction method for a low-contrast vehicle image, and belongs to the technical field of image feature processing. The method comprises the following steps of: performing pyramidal decomposition on the image by using wavelet transform; then performing gain processing on the wavelet high-frequency coefficient after the pyramidal decomposition by adopting a gain function to improve the contribution of a high-frequency detail part; and finally performing further laying and normalization on all data, and inputting the feature vectors formed by the processed data into a classifier to perform classification and identification. The method solves the problem of poor classification and identification effect after feature extraction of the low-contrast image in the conventional feature extraction mode; and vehicle detection experiments performed on the Twilight image by using the method and a support vector machine classifier show that the method obviously improves the vehicle identification rate and greatly reduces the non-vehicle error identification rate compared with the conventional wavelet feature extraction method.

Description

A kind of wavelet character method for distilling of low contrast vehicle image

Technical field:

The present invention relates to a kind of wavelet character method for distilling of low contrast vehicle image, belong to the characteristics of image processing technology field.

Background technology:

Feature extraction is the key of pattern-recognition (pattern recogniton); At present; In vehicle detection based on vision; Feature extracting method commonly used comprises that wavelet character extracts (wavelet feature extraction), the principal ingredient analytical characteristic extracts (PCA feature extraction; Primary component analysis feature extraction), local direction coding (LOC, local orientation coding), gal cypress feature extraction methods such as (Gabor).Wherein wavelet character is very attractive in vehicle detection is used, and its main cause is following three aspects: the first, and wavelet character can provide the compactness of image border and contour feature to describe; The second, wavelet character can provide the description of characteristics of image under different scale; The 3rd, wavelet character to obtain required amount of computational resources few.

Existing vehicle image pattern-recognition flow process based on wavelet character is as shown in Figure 1:

At first vehicle image is discerned and obtained area-of-interest (ROI, Region of Interest), carry out tower decomposition afterwards; Secondly wavelet coefficient is carried out relevant treatment,, choose bigger wavelet coefficient and simultaneously it is quantified as [1,0,1], the constitutive characteristic vector such as removing the high frequency (highest frequency) of ground floor; Carry out vehicle detection through sorter (like SVMs) at last, obtain the result.

The detailed process of the tower decomposition of aforementioned small echo is as shown in Figure 2:

The first step is carried out low pass (low pass) and high pass (high pass) filtering (filter) to each row in N * N dimension image respectively.And then the image that carries out after low pass, the high-pass filtering carried out 1/2nd samplings downwards respectively, 1/2nd samplings (sample) downwards are equivalent to whenever in N * N dimension image remove row at a distance from row, obtain the image of a N * (N/2) tie up.

In second step, each row of the image that the first step is obtained carry out low pass and high-pass filtering respectively.Carry out 1/2nd samplings downwards then respectively downwards, sampling is removed delegation in every line with the image of the dimension of N after twice filtering * (N/2), and each branch that sets shown in this master drawing has all become the image of a width of cloth (N/2) * (N/2) dimension.If the low pass details to (N/2) * (N/2) dimension is carried out and the identical process of the tower decomposition of original image.Then can produce the subimage of 4 (N/4) * (N/4) dimension: low pass part (LL), and level (LH), vertical (HL) and diagonal angle (HH) part.Analysis can go on always, till the subimage that obtains only comprises a pixel.

When illumination condition, when the vehicle image contrast is relatively good, the wavelet character that above feature extracting method extracts detects in the application at vehicle image has recognition effect preferably.But because distance, illumination condition, weather condition and the Effect on Performance of video camera own make when picture contrast reduction, image blur, edge and details (detail) are unintelligible; Existing wavelet character method for distilling can't extract the edge and the contour feature of vehicle image effectively; Then make image enhancing, not enough occur easily through image pre-processing method such as histogram equalization (histogram equalization) to noise-sensitive and local contrast; The problem that causes detailed information to lose, thus the pattern-recognition effect of vehicle image influenced.

Summary of the invention

Technical matters to be solved of the present invention is to carry out feature extraction to the low contrast vehicle image that exists in the background technology to cause vehicle classification to detect bad problem; Characteristic based on wavelet multiresolution rate (wavelet multi-resolution); Propose a kind of soft image effectively to be carried out feature extracting methods.

The present invention adopts following technical scheme for realizing the foregoing invention purpose:

A kind of wavelet character method for distilling of low contrast vehicle image comprises and utilizes wavelet transformation that image is carried out tower decomposition step, wavelet coefficient treatment step, sorter detection step; Wherein the wavelet coefficient treatment step comprises small echo high frequency coefficient gain step and layering normalization step:

(1), small echo high frequency coefficient gain step:

Image is divided into low frequency sub-band part and high-frequency sub-band part after the tower decomposition step of wavelet transformation; Adopt gain function that the small echo high-frequency sub-band is partly carried out gain process; Said high-frequency sub-band comprises horizontal subband, vertical subband and diagonal angle subband; Employing formula (1) layering is carried out gain process to the small echo high frequency coefficient

E(m _i，j)＝W(m_max _i，j)×K(m _i，j) (1)

Wherein, W () is the gain weights, and K () is a gain function, i=1 ..., N-1, N are the number of plies of the tower decomposition of small echo, j=1, and 2,3, j=1 representes horizontal subband details, and j=2 representes vertical subband details, and j=3 representes diagonal angle subband details; m _{I, j}Represent i layer j corresponding subband wavelet coefficient, m_max _{I, j}It is the maximal value of i layer j corresponding subband coefficient amplitude; In the formula (1):

$W(m\_{\max }_{i,j})=(\mathrm{\σ}\_{\max }_i/{\mathrm{\σ}}_{i,j})---\left(2\right)$

Wherein, σ _{I, j}Represent the standard variance of i layer j corresponding subband coefficient,

Wherein,

Represent the average of i layer j corresponding subband wavelet coefficient, n represents the number of i layer j corresponding subband wavelet coefficient; σ _ max _iIt is the maximal value of the standard variance of three sub-band coefficients of i layer;

$K\left(m_{i,j}\right)=\frac{S\left[{\mathrm{\α}}_{i,j}(m_{i,j}-{\mathrm{\β}}_{i,j})\right]-S[-{\mathrm{\α}}_{i,j}(m_{i,j}+{\mathrm{\β}}_{i,j})]}{S\left[{\mathrm{\α}}_{i,j}(1-{\mathrm{\β}}_{i,j})\right]-S[-{\mathrm{\α}}_{i,j}(1+{\mathrm{\β}}_{i,j})]}---\left(3\right)$

S [] represents nonlinear function in the formula,

Wherein e is at the bottom of the natural logarithm and e is a constant, α _{I, j}Be the parameter of ride gain function curve shape,

β _{I, j}Be the critical point of S [] function,

(2) layering normalization step:

After above gain process, with each layer wavelet data by formula (4) normalize to [0,1];

e _i，j＝(m′ _i，j-min_value _i，j)/(max_value _i，j-min_value _i，j) (4)

M ' in the formula _{I, j}Be the data of i layer j corresponding subband after above gain function gain; Min_value _{I, j}It is the minimum value of i layer j corresponding subband data after gain process; Max_value _{I, j}It is the maximal value of i layer j corresponding subband data after gain process.

Further, the sorter of the wavelet character method for distilling of aforementioned low contrast vehicle image detects in the step and adopts SVMs, AdaBoost or neuroid to carry out classification and Detection.

The present invention adopts technique scheme compared with prior art; Solved in the prior art soft image has been carried out the relatively poor problem of Classification and Identification effect after the feature extraction; Utilize the method combination supporting vector machine sorter that is proposed that blurred picture (Twilight) is carried out the vehicle detection experiment; Compare with existing wavelet character method for distilling, obviously improved the vehicle identification rate, reduced non-vehicle false recognition rate significantly.

The present invention normalizes to wavelet coefficient values [0,1], compare wavelet coefficient values in the background technology can only get 1,0, some among the 1}, trickleer to the sign of characteristics of image, distance is littler in type of making, robustness is better when being used to discern.

Description of drawings:

Fig. 1 is an existing vehicle image pattern-recognition process flow diagram based on wavelet character in the background technology.

Fig. 2 is about the detailed process synoptic diagram of the tower decomposition of small echo among Fig. 1; Wherein HP representes Hi-pass filter, and LP representes low-pass filter; ↓ 2 expressions 1/2nd sampling downwards.

Fig. 3 is the process flow diagram of the wavelet character method for distilling of low contrast vehicle image of the present invention.

Specific embodiments:

Below in conjunction with accompanying drawing the enforcement of technical scheme is done further to describe in detail:

As shown in Figure 1, the vehicle image pattern-recognition flow process based on wavelet character in the prior art is: at first vehicle image is discerned and obtained area-of-interest (ROI, Region of Interest), carry out tower decomposition afterwards; Secondly wavelet coefficient is carried out relevant treatment,, choose bigger wavelet coefficient and simultaneously it is quantified as [1,0,1] such as removing the high frequency (highest frequency) of ground floor; The constitutive characteristic vector carries out vehicle detection through sorter (like SVMs) at last then, obtains the result.

As shown in Figure 2, the detailed process of the tower decomposition of small echo is: the first step, respectively each row in N * N dimension image is carried out low pass (low pass) and high pass (high pass) filtering (filter).And then the image that carries out after low pass, the high-pass filtering carried out 1/2nd samplings downwards respectively, 1/2nd samplings (sample) downwards are equivalent to whenever in N * N dimension image remove row at a distance from row, obtain the image of a N * (N/2) tie up.

In second step, each row of the image that the first step is obtained carry out low pass and high-pass filtering respectively.Carry out 1/2nd samplings downwards then respectively downwards, sampling is removed delegation in every line with the image of the dimension of N after twice filtering * (N/2), and each branch that sets shown in this master drawing has all become the image of a width of cloth (N/2) * (N/2) dimension.If the low pass details to (N/2) * (N/2) dimension is carried out and the identical process of the tower decomposition of original image.Then can produce the subimage of 4 (N/4) * (N/4) dimension: low pass part (LL), and level (LH), vertical (HL) and diagonal angle (HH) part.Analysis can go on always, till the subimage that obtains only comprises a pixel.

As shown in Figure 3, the wavelet character method for distilling of low contrast vehicle image of the present invention comprises that tower decomposition step, wavelet coefficient treatment step, sorter detect step; Wherein the wavelet coefficient treatment step comprises small echo high frequency coefficient gain step and layering normalization step, and is specific as follows:

(1), small echo high frequency coefficient gain step:

Image is divided into low frequency sub-band (low frequency sub-band) and high-frequency sub-band (highfrequency sub-band) two large divisions behind wavelet transformation; Low frequency sub-band comprises the low frequency detailed information of image; Be the approximate of image; Details is fuzzy, and high-frequency sub-band comprises the detail of the high frequency (like level, vertical and diagonal line) and the much noise of image.When the vehicle image contrast is relatively lower, make existing wavelet character method for distilling can't extract the image detail of the high frequency effectively, and anti-noise ability (antinoise ability) is poor.For this reason, utilize gain function that small echo HFS (being LH, HL and three subbands of HH) is carried out gain process.For example, adopt and the small echo high frequency coefficient to be carried out enhancement process suc as formula (1) defined gain function E (x) layering, when improving wavelet character and being used for classification and Detection to the adaptive faculty of noise.

E(m _i，j)＝W(m_max _i，j)×K(m _i，j) (1)

Wherein, W () is gain weights (gain weight), and K () is a gain function, i=1 ..., N-1, N are the number of plies (level) of the tower decomposition of small echo, j=1, and 2,3, j=1 representes level detail, and j=2 representes vertical detail, and j=3 representes diagonal detail; m _{I, j}Represent i layer j corresponding subband wavelet coefficient, m_max _{I, j}It is the maximal value of i layer j corresponding subband coefficient amplitude (coefficient magnitudes).

$W(m\_{\max }_{i,j})=(\mathrm{\σ}\_{\max }_i/{\mathrm{\σ}}_{i,j})---\left(2\right)$

Wherein

Represent the standard variance of i layer j corresponding subband coefficient,

Represent i layer j corresponding subband coefficient average, n represents i layer j corresponding subband wavelet coefficient number, σ _ max _iIt is the maximal value of the standard variance of three sub-band coefficients of i layer.

$K\left(m_{i,j}\right)=\frac{S\left[{\mathrm{\α}}_{i,j}(m_{i,j}-{\mathrm{\β}}_{i,j})\right]-S[-{\mathrm{\α}}_{i,j}(m_{i,j}+{\mathrm{\β}}_{i,j})]}{S\left[{\mathrm{\α}}_{i,j}(1-{\mathrm{\β}}_{i,j})\right]-S[-{\mathrm{\α}}_{i,j}(1+{\mathrm{\β}}_{i,j})]}---\left(3\right)$

In the formula

α _{I, j}Be the parameter of ride gain function curve shape, β _{I, j}Be the critical point of S (x) function, here order

β _{I, j}=σ _{I, j}/ 2;

(2) layering normalization:

After above gain process, wavelet data is carried out layering normalization handle.The reason that layering normalization is handled is: first is based on wavelet multiresolution rate characteristic, has better comparability with layer wavelet coefficient; The secondth, in order to improve of the contribution of the smaller attribute of data value to classification; The 3rd is to reduce calculated amount, practices thrift computational resource.For example, with each layer wavelet data by formula (4) normalize to [0,1].

e _i，j＝(m′ _i，j-min_value _i，j)/(max_value _i，j-min_value _i，j) (4)

M ' in the formula _{I, j}Be the data of i layer j corresponding subband after above gain function gain.Min_value _{I, j}Be the minimum value of i layer j corresponding subband data after gain process, max_value _{I, j}It is the maximal value of i layer j corresponding subband data after gain process.

Claims (2)

1. the wavelet character method for distilling of a low contrast vehicle image comprises and utilizes wavelet transformation that image is carried out tower decomposition step, wavelet coefficient treatment step, sorter detection step; It is characterized in that: said wavelet coefficient treatment step comprises small echo high frequency coefficient gain step and layering normalization step, wherein:

(1), small echo high frequency coefficient gain step:

Image is divided into low frequency sub-band part and high-frequency sub-band part after the tower decomposition step of wavelet transformation; Adopt gain function that the small echo high-frequency sub-band is partly carried out gain process; Said high-frequency sub-band comprises horizontal subband, vertical subband and diagonal angle subband; Employing formula (1) layering is carried out gain process to the small echo high frequency coefficient

E(m _i，j)＝W(m_max _i，j)×K(m _i，j) (1)

Wherein, W () is the gain weights, and K () is a gain function, i=1 ..., N-1, N are the number of plies of the tower decomposition of small echo, j=1, and 2,3, j=1 representes horizontal subband details, and j=2 representes vertical subband details, and j=3 representes diagonal angle subband details; m _{I, j}Represent i layer j corresponding subband wavelet coefficient, m_max _{I, j}It is the maximal value of i layer j corresponding subband coefficient amplitude; In the formula (1):

W(m_max _i，j)＝(σ_max _i/σ _i，j) (2)

Wherein, σ _{I, j}Represent the standard variance of i layer j corresponding subband coefficient,

Wherein,

Represent the average of i layer j corresponding subband wavelet coefficient, n represents the number of i layer j corresponding subband wavelet coefficient; σ _ max _iIt is the maximal value of the standard variance of three sub-band coefficients of i layer;

$K\left(m_{i,j}\right)=\frac{S\left[{\mathrm{\α}}_{i,j}(m_{i,j}-{\mathrm{\β}}_{i,j})\right]-S[-{\mathrm{\α}}_{i,j}(m_{i,j}+{\mathrm{\β}}_{i,j})]}{S[{\mathrm{\α}}_{i,j}(1-{\mathrm{\β}}_{i,j})]-S[-{\mathrm{\α}}_{i,j}(1+{\mathrm{\β}}_{i,j})]}---\left(3\right)$

S [] represents nonlinear function in the formula,

At the bottom of wherein e is natural logarithm, α _{I, j}Be the parameter of ride gain function curve shape,

β _{I, j}Be the critical point of S [] function, β _{I, j}=σ _{I, j}/ 2;

(2) layering normalization step:

After above gain process, with each layer wavelet data by formula (4) normalize to [0,1];

e _i，j＝(m′ _i，j-min_value _i，j)/(max_value _i，j)-min_value _i，j) (4)

M ' in the formula _{I, j}Be the data of i layer j corresponding subband after above gain function gain; Min_value _{I, j}It is the minimum value of i layer j corresponding subband data after gain process; Max_value _{I, j}It is the maximal value of i layer j corresponding subband data after gain process.

2. the wavelet character method for distilling of low contrast vehicle image according to claim 1 is characterized in that: said sorter detects in the step and adopts SVMs, AdaBoost or neuroid to carry out classification and Detection.

Images