WO2017197618A1 - 一种红外图像中条纹噪声的去除方法及系统 - Google Patents

一种红外图像中条纹噪声的去除方法及系统 Download PDF

Info

Publication number
WO2017197618A1
WO2017197618A1 PCT/CN2016/082581 CN2016082581W WO2017197618A1 WO 2017197618 A1 WO2017197618 A1 WO 2017197618A1 CN 2016082581 W CN2016082581 W CN 2016082581W WO 2017197618 A1 WO2017197618 A1 WO 2017197618A1
Authority
WO
WIPO (PCT)
Prior art keywords
sub
image
noise
filtered
picture
Prior art date
Application number
PCT/CN2016/082581
Other languages
English (en)
French (fr)
Inventor
裴继红
邹咪
谢维信
杨烜
Original Assignee
深圳大学
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 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2016/082581 priority Critical patent/WO2017197618A1/zh
Priority to CN201680000899.4A priority patent/CN106462957B/zh
Publication of WO2017197618A1 publication Critical patent/WO2017197618A1/zh
Priority to US15/838,404 priority patent/US10282824B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/10—Image enhancement or restoration using non-spatial domain filtering
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/80—Camera processing pipelines; Components thereof
    • H04N23/81—Camera processing pipelines; Components thereof for suppressing or minimising disturbance in the image signal generation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N25/00—Circuitry of solid-state image sensors [SSIS]; Control thereof
    • H04N25/60—Noise processing, e.g. detecting, correcting, reducing or removing noise
    • H04N25/67—Noise processing, e.g. detecting, correcting, reducing or removing noise applied to fixed-pattern noise, e.g. non-uniformity of response
    • H04N25/671—Noise processing, e.g. detecting, correcting, reducing or removing noise applied to fixed-pattern noise, e.g. non-uniformity of response for non-uniformity detection or correction
    • H04N25/677—Noise processing, e.g. detecting, correcting, reducing or removing noise applied to fixed-pattern noise, e.g. non-uniformity of response for non-uniformity detection or correction for reducing the column or line fixed pattern noise
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N5/00—Details of television systems
    • H04N5/30—Transforming light or analogous information into electric information
    • H04N5/33—Transforming infrared radiation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10048—Infrared image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20048—Transform domain processing
    • G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]

Definitions

  • the invention belongs to the technical field of image processing, and more particularly to a method and system for removing stripe noise in an infrared image.
  • Infrared technology has been applied to various fields such as biology, medicine, geosciences, and military reconnaissance.
  • infrared images have a lower signal-to-noise ratio and are easily contaminated by various noises.
  • fringe noise is a kind of noise that is often present in infrared images.
  • the causes of streak noise are more complicated.
  • the process variation of multi-sensor, the aging of instruments and components, and the error of internal calibration system are different.
  • the conversion transfer function of each detection unit is different.
  • the external environment such as temperature and surroundings during infrared image acquisition Interference from other devices, etc.
  • the stripe noise removal algorithm is used to first remove the streak noise in the infrared image, which will greatly improve the reliability of subsequent processing and analysis.
  • the fringe noise in an image is generally a kind of periodic noise, which is reflected in the frequency domain where a noise component appears at a fixed frequency point.
  • periodic noise frequency component it is first necessary to find the corresponding position of the noise frequency component in the frequency domain, then use the notch filter to filter out the noise, and finally restore the signal to the time domain or the spatial domain.
  • the key is how to find the exact noise frequency and how to design a suitable notch filter.
  • One of the problems with the prior art image stripe removal method is that the position determination of the noise frequency component is not sufficiently intelligent and accurate.
  • One of the existing methods employs manual detection of the corresponding position of the noise frequency component. This method is very time consuming, computationally inefficient, unable to meet the application requirements of large data volume, and is subject to artificial subjective factors; the existing method 2 locates the frequency point of the stripe noise according to the projection map of the image amplitude spectrum. The method projects the amplitude spectrum of the image in the row direction or the column direction, and obtains the position of the noise frequency component through the projected cumulative distribution function. This method improves efficiency compared to manual methods.
  • Another problem with the prior art image stripe removal method is the periodic alignment problem.
  • the periodic noise in the image is only concentrated when the acquired signal has a strict whole period, and the frequency component of the noise is concentrated to a limited frequency. If the fringe noise in the image does not have a strict integer periodicity, the frequency components of the noise will spread over the entire frequency domain to varying degrees. If the image is not pre-processed with periodic alignment prior to filtering, the notch filter method cannot be used to completely filter out the frequency components of the stripe noise.
  • the object of the present invention is to provide a method and a system for removing streak noise in an infrared image, which aims to solve the problem that the existing stripe noise removal technology cannot automatically and efficiently process infrared images with different intensities. Stripe noise problem.
  • the present invention provides a method for removing streak noise in an infrared image, comprising the following steps:
  • S1 selecting specific row data of the image to be filtered Aim, and obtaining the width N1 of the stripe by using the spectrum information of the specific row data;
  • S2 intercepting a specific line sequence of the image into (N1+1) new sequences of different lengths according to the width N1 of the stripe, and obtaining an optimal intercept length N by analyzing amplitude spectrum information of each new sequence;
  • S5 Perform brightness adjustment and fusion processing on the first sub-picture Bimf and the second sub-picture Cimf to obtain an image Aimf.
  • step S1 the width N1 of the stripe, symbol It represents rounding down, i.e., dropping the fractional part, for the unchanged integer; Nn denotes the length of a line of data; k m denotes a frequency at a peak in the amplitude spectrum corresponding to the coordinates.
  • step S2 includes the following steps:
  • S21 The specific row data of length Nn in the image to be filtered Aim is regarded as a discrete sequence x(n) of Nn points, and the discrete sequence x(n) is intercepted into different lengths in N1+1 according to the width N1 of the stripe. a new sequence and calculate the amplitude spectrum of the new sequence of different lengths;
  • k mN1 are new sequences respectively The first peak of the magnitude spectrum
  • Corresponding amplitude spectrum coordinates k m calculate the maximum value of this set of eigenvalues, and obtain the optimal intercept length N.
  • the new sequence in which the discrete sequence x(n) is truncated to N1+1 different lengths according to the width N1 of the stripe is specifically: starting from the first point of the sequence
  • the Nn point, the first Nn-1 point, the first Nn-2 point, ..., the front Nn-N1 point are taken as a new sequence, and are denoted as x Nn-0 (n), x Nn-1 (n), x Nn- 2 (n), ... x Nn-N1 (n).
  • step S22 the local maximum is found by traversing the local interval [0.5 k m 1.5 k m ], and the first peak of the noise frequency component is obtained.
  • step S4 the parameters of the notch comb filter are determined by detecting the position of the noise frequency component in the optimal to-be-filtered sub-pictures Bim and Cim.
  • step of determining the parameter of the notch comb filter is specifically:
  • the parameters of the notch comb filter are obtained based on the distance and the cutoff frequency.
  • step S5 the brightness adjustment is specifically: taking the first sub-picture Bimf and the second sub-picture Cimf
  • the average of the DC components is taken as the respective DC component.
  • the merging process is specifically: taking an average value in an overlapping area of the first sub-picture Bimf and the second sub-picture Cimf, and taking a non-overlapping area of the first sub-picture Bimf and the second sub-picture Cimf Valid value.
  • the invention also provides a stripe noise removal system in an infrared image, comprising:
  • An estimation module configured to select specific row data of the image to be filtered Aim, and obtain a width N1 of the stripe by using spectrum information of the specific row data;
  • a query module configured to intercept a specific line sequence of the image into (N1+1) new sequences of different lengths according to the width N1 of the stripe, and obtain an optimal intercept length N by analyzing amplitude spectrum information of each new sequence;
  • a segmentation module configured to divide a to-be-filtered image Aim of size M ⁇ N n into a pair of optimal to-be-filtered sub-pictures Bim and Cim of size M ⁇ N according to an optimal intercept length N;
  • a filtering module configured to perform frequency filtering on two optimal to-be-filtered sub-pictures Bim and Cim by using a notch comb filter to obtain a first sub-picture Bimf and a second sub-picture Cimf;
  • a fusion module configured to perform brightness adjustment and fusion processing on the first sub-picture Bimf and the second sub-picture Cimf to obtain an image Aimf.
  • the invention realizes a method and a system for removing infrared image stripe noise, and better solves the problem that the existing stripe noise removal technology cannot automatically and efficiently process infrared light images with different intensity stripe noise.
  • the length of the period alignment is first selected according to the periodicity of the noise, and the image is divided into two sub-graphs with overlapping regions, and then the two sub-pictures are subjected to frequency domain filtering.
  • frequency domain filtering firstly, the noise frequency components in the frequency domain are found according to the characteristics of the discrete Fourier transform of the finite long period sequence, and then the adaptive Butterworth notch filter is filtered for each pair of noise frequency components. Finally, the two filtered subgraphs are brightness adjusted and fused to obtain the final result.
  • the manual participation is small, the batch processing efficiency is high, and the filtering effect is good, and is particularly suitable for processing a large amount of infrared video data including stripe noise.
  • FIG. 1 is a flowchart of a method for removing stray noise of an infrared image according to an embodiment of the present invention
  • FIG. 2(a) is a schematic diagram of a subsequence described in the embodiment of the present invention.
  • Figure 2 (b) is an amplitude spectrum of the sequence shown in Figure 2 (a) in the embodiment of the present invention.
  • FIG. 2(c) is a schematic diagram showing a finite long period sequence obtained by extending a subsequence shown in FIG. 2(a) in the embodiment of the present invention
  • Figure 2 (d) is an amplitude spectrum of the sequence shown in Figure 2 (c) in the embodiment of the present invention.
  • Figure 3 (a) is an infrared image to be processed in the embodiment of the present invention.
  • Figure 3 (b) is an amplitude spectrum of the first line of data of the image shown in Figure 3 (a) in the embodiment of the present invention
  • 5(a)-(m) are amplitude spectrum diagrams of 13 sequences of different lengths in the embodiment of the present invention, and the thick line is the first peak of noise;
  • Figure 5 (n) is a spectral peak diagram of the first peak of the noise in the graphs of Figures 5 (a) - (m) in the embodiment of the present invention, the thick line is the maximum value of these peaks;
  • FIG. 6 is a schematic diagram of dividing an image to be filtered into two optimal sub-filters to be filtered according to an embodiment of the present invention
  • FIG. 8 is a schematic diagram of a first-order Butterworth notch comb filter including five pairs of notch pairs in an embodiment of the present invention
  • FIG. 11 is a detailed flowchart of determining a cutoff frequency of a filter in an embodiment of the present invention.
  • FIG. 12 is a schematic diagram of image fusion in an embodiment of the present invention.
  • FIG. 13(a)-(h) are comparison diagrams of several examples of infrared image filtering before and after the embodiment of the present invention.
  • FIG. 14 is a system structural diagram of a method for removing stray noise of an infrared image provided in an embodiment of the present invention.
  • the stripe noise removal method proposed by the invention fully utilizes the characteristics of the periodic noise in the frequency domain when searching for the position of the noise frequency component, and automatically corrects the parameters of the filter through feedback until the ideal filtering effect can be achieved.
  • the method provided by the invention can adaptively process stripe noises with different strengths, and does not require manual participation, and can automatically, efficiently and accurately filter out different stripe noises in the image.
  • the specific method for stripe noise removal in the infrared image proposed by the present invention is to first select a specific line of the image, and estimate the width of the strip by analyzing the spectral information of the line. Then, using the width of the stripe, the specific line sequence of the image is truncated into new sequences of different lengths. By analyzing the amplitude spectrum information of each new sequence, the optimal truncation length is found. Then, the original image to be filtered is cut into a pair of optimal to-be-filtered sub-pictures according to the optimal intercept length. Then, the position of the noise frequency component in the sub-picture is detected, and the notch comb filter is designed to perform frequency filtering on the optimal sub-picture to be filtered. Finally, the filtered pair of subgraphs are fused to obtain the final result.
  • FIG. 1 shows a flow of a method for removing infrared image stripe noise provided by the present invention, comprising the following steps:
  • S1 Select specific line data of the image to be filtered Aim, analyze the spectrum information of the line, and estimate the width N1 of the stripe.
  • the stripe noise is periodic noise with a period N1 of its stripe.
  • each sequence of lines in the image is superimposed with a finite long period noise sequence with a period N1 of the fringe.
  • the so-called finite-length periodic sequence refers to the process of performing a plurality of periodic extensions of a sub-sequence.
  • Fig. 2(a) a 6-point discrete sequence is shown, and Fig. 2(c) shows a new sequence obtained by performing a 9-cycle extension sequence with the 6-point discrete sequence shown in Fig. 2(a) as a subsequence.
  • the method of obtaining the stripe width N1 is: selecting a specific line data of the original image Aim (assuming that the length of one line of data is Nn) as a discrete sequence x(n) of Nn points, wherein the specific line can be any line, but The middle row data is more stable, so the middle row is taken as a specific row in the present invention.
  • the discrete Fourier transform X(k) of x(n) is obtained from the formula of the discrete Fourier transform, as shown in the formula (1). Then find the absolute value of X(k)
  • the fast Fourier transform is used to find X(k). When all the spectra of one sequence are obtained, including the spectrum of a two-dimensional sequence, the fast Fourier transform is used in the present invention.
  • Fig. 3(a) an infrared image a to be filtered is shown, wherein the two straight lines in Fig. 3(a) are a complete stripe.
  • the amplitude spectrum of a finite-length periodic sequence is equivalent to inserting a zero value and then amplifying the amplitude spectrum of the sub-sequence.
  • the fringe noise is equivalent to a sub-strip multiple period.
  • the frequency component of the noise will appear at equal intervals of the spectrogram relative to the frequency component of the entire image, and the amplitude of the frequency component of the noise will be large, manifesting as a local peak.
  • the amplitude spectrum of the middle row sequence of Fig. 3(a) is the result of uncentered. The peak appearing at the mid-high frequency is the frequency component representing the noise (the DC component is not considered).
  • the method adopted by the present invention first finds the amplitude spectrum coordinate k m corresponding to the first peak at the mid-high frequency, and the stripe width N1 is determined as symbol Indicates rounding down, ie removing the fractional part, leaving the integer unchanged.
  • the method for obtaining the amplitude spectral coordinate k m corresponding to the first peak of the noise in the present invention is that for the amplitude spectrum
  • ,i -5,-4,...,0,1,...,4,5 ⁇ ,max[S Xk ] is the value of the element having the largest value in the set S Xk , Med[S Xk ] is the value of the element with the median value in the set S Xk , that is, the elements in the set are sorted in descending order, the value of the element in the middle. Then k m is calculated by the following formula:
  • ⁇ in the above formula is a threshold value indicating the difference between the local maximum value and the local median value, and the larger the threshold value, the larger the difference between the local maximum value and the local median value.
  • the optimal intercept length refers to a new sequence of a certain length intercepted from the original sequence, and the noise frequency component in the spectrum is the most concentrated, and the length of the sequence is the optimal intercept length. Since the original point, the last two points, or the last multiple points are removed for the original sequence, the distribution of the noise frequency components in the spectrum changes significantly. To this end, we aim to find a truncation length such that the sequence under this length has the most concentrated noise frequency components in the spectrum for subsequent filtering.
  • step S2 includes the following steps:
  • the specific row also takes the middle row, that is, takes the middle row as a discrete sequence x(n) of Nn points, and intercepts x(n) as a new sequence of different length according to the width N1 of the stripe: refers to the first sequence
  • the point is the starting point, and the first Nn point, the first Nn-1 point, the first Nn-2 point, ..., the front Nn-N1 point are sequentially taken as a new sequence, which is denoted as x Nn-0 (n), x Nn-1 (n ), x Nn-2 (n), ... x Nn-N1 (n).
  • the amplitude spectrum of each sequence is obtained from the discrete Fourier transform
  • Figs. 5(a)-(m) 13 new sequences of different lengths are truncated, and the amplitude spectrum of each sequence is obtained, as shown in Figs. 5(a)-(m). It can be seen from Fig. 5(a)-(m) that if a point is added or decreased on the basis of a sequence, the corresponding amplitude spectrum will have a large difference, especially the frequency component representing the periodic noise will be obvious. Variety.
  • k mN1 are respectively new sequences
  • the amplitude spectrum coordinate k m is used to find the extreme value of this set of eigenvalues, so as to obtain the optimal intercept length N.
  • the first peak representing the noise component refers to the first peak of the DC component from which the noise is removed, that is, the first peak at the middle and high frequencies.
  • the coarse peak pulse line in each amplitude spectrum is the first peak representing the noise frequency component.
  • the present invention employs traversing local intervals to find the local maximum, the first peak of the noise frequency component.
  • step S1 it is given how to find the amplitude spectral coordinate k m corresponding to the first peak.
  • the coordinate corresponding to the first peak representing the noise component in the amplitude spectrum of the new sequence of different lengths must be at the coordinate k.
  • Near m the present invention traverses the local interval [0.5k m , 1.5k m ] to find the local maximum, and extracts all the first peaks as a set of eigenvalues, as shown in FIG. 5(n) is the extraction step S21.
  • the method for finding the extremum is to first find the maximum value of the set of eigenvalues, and then determine whether the maximum value is the first value. If it is not the first value, the maximum value is considered to be the extremum. For a value, remove the first value of the set of eigenvalues and re-find the maximum value of the remaining values in the array until the maximum value is not the first value.
  • the image to be filtered Aim of size M ⁇ N n is divided into a pair of optimal to-be-filtered sub-pictures Bim and Cim of size M ⁇ N.
  • the image Aim is now divided into a pair of optimal sub-pictures Bim and Cim.
  • the method adopted is: taking the first N columns as a sub-picture Bim, and taking the N-th column as another sub-picture Cim.
  • N 477 is obtained, and the first 477 columns are used as one sub-picture b, and the last 477 columns are used as sub-picture c.
  • Fig. (b) represents subgraph b
  • Fig. (c) represents subgraph c.
  • the notch comb filter is selected to filter the optimal sub-picture to be filtered.
  • the notch filter has an ideal type, Butterworth type and Gauss type.
  • the Butterworth notch filter is selected in the present invention.
  • the notch filter is designed by first detecting the coordinate position of all possible frequency components of the noise, and then designing different notch pairs according to the magnitude of each pair of noise frequency components.
  • the second-order Butterworth trapping band rejection filter is used in the present invention. The main difference between different notch pairs is reflected in the difference between the center and the cutoff frequency. Since the centers of the notch pairs appear equally spaced, the filter is comb-like.
  • step S4 includes the following steps:
  • one of the sub-pictures Bim is first selected for spectrum analysis to estimate the total logarithm Np of all frequency components of the sub-picture Bim that may contain noise frequency components, wherein the method of estimating the total logarithm Np is to take out the sub-picture to be filtered Bim
  • the middle row sequence is denoted by x b (n), the length of the sequence is N, and its amplitude spectrum
  • the method used is similar to the traversing local interval [0.5k m , 1.5k m ] described in step S22, where k m is obtained in step S1,
  • the local maximum, that is, the first peak of the noise frequency component, assuming that the abscissa corresponding to the first peak is q
  • the frequency components of the stripe noise of the two are equally spaced.
  • the vertical stripes in the image are concentrated on the horizontal axis of the Fourier frequency domain energy spectrum, and the frequency components are also present at equal intervals of the spacing ⁇ .
  • noise frequency component is obtained according to the formula (3) and the formula (4).
  • the frequency components at (u i1 , v i1 ) and (u i2 , v i2 ) are a pair of symmetrically centered frequency components that may contain noise frequency components, and the subscript i represents different pairs of frequency components.
  • M refers to the number of lines of the image
  • N refers to the number of columns of the image, that is, the size of the processed image is M ⁇ N.
  • the notch comb filter is constructed with the product of a high pass filter whose center has been translated to the center of the notch filter. Its form is shown in formula (5).
  • H k (u, v) and H - k (u, v) are high-pass filters whose centers are at (u k , x k ) and (-u k respectively , -v k ).
  • the center is based on the center of the frequency rectangle It is determined, and the size of the frequency rectangle is the same as the size of the image to be filtered. For each filter, the distance is calculated by equation (6) and equation (7):
  • a Butterworth notch comb filter is used, and below is an n-th order Butterworth trap comb filter comprising three notch pairs:
  • D k and D - k are given by equations (6) and (7).
  • the constant D 0k is the same for each notch pair and can be different for different notches.
  • Figure 8 shows a first-order Butterworth notch comb filter containing five pairs of notch pairs.
  • the frequency component of the noise is at most Np pairs, that is, less than or equal to Np pairs.
  • the method for realizing frequency domain filtering is not directly designing a filter containing Np notch pairs, but designing each pair of noise components in the image pair of filter pairs containing only one notch pair multiple times. Filtering is performed in sequence until all noise frequency components are filtered out.
  • the specific filtering step includes the following steps:
  • the noise frequency component found in step S41 is not necessarily a true frequency component containing noise. Therefore, before filtering, it is first determined whether the frequency component at the pair of points actually contains the noise frequency component. information.
  • the amplitude spectrum F A (u, v)
  • after image centering is first obtained. Since the amplitude is symmetric about the center, with (u i1 , v i1 ) as the center, the neighborhood with the size of (2m+1)(2n+1) is selected as S FAi , and (u i1 , v i1 ) is the center. The neighborhood of size (2m+3) (2n+3) is S FAEi .
  • the neighborhoods S FAi and S FAEi are defined as follows:
  • the decision threshold TF for whether noise filtering is required for the frequency point is calculated by:
  • S FAEi -S FAi denotes a set of all elements belonging to the set S FAEi but not belonging to the set S FAi
  • the symbol med ⁇ denotes the median in the set, that is, the elements in the set are arranged in descending order.
  • T>TF it is considered that the frequency components at (u i1 , v i1 ) and (u i2 , v i2 ) contain noise frequency components and need to be filtered; otherwise, if T ⁇ TF, then (u i1 , v The noise frequency components that are not included in the frequency components at i1 ) and (u i2 , v i2 ) do not need to be filtered. Then, examine the next pair of frequency components that may contain noise frequency components.
  • S422 Design a Butterworth notch band rejection filter for all frequency components including the noise frequency component, and filter the noise frequency components to obtain the filtered sub-pictures Bimf and Cimf.
  • the amplitude is generally different, so different notch filters should be designed for different noise frequency components.
  • the idea of designing the notch filter is the same for different noise components.
  • a pair of noise frequency components will be taken as an example to illustrate the process of designing the notch filter.
  • filters are designed for filtering. Further, as shown in FIG. 9, the steps of designing the filter include the following steps:
  • S4221 Select a noise frequency component to be processed, and calculate a distance from each point to a coordinate corresponding to the noise frequency component, that is, determine distances D ki1 (u, v) and D ki2 (u, v).
  • M is the number of rows of the frequency rectangle
  • N is the number of columns of the frequency rectangle
  • ⁇ is the pitch between the frequency components of the noise obtained in step S41.
  • the method for determining the cutoff frequency of the filter is to first initialize a smaller cutoff frequency, obtain a filter function to filter the image, and then examine the filtered result to update the cutoff frequency until the filtered result is reached.
  • the step of determining the cutoff frequency and performing frequency domain filtering on the image includes the following steps:
  • S42221 Initialize the cutoff frequency D 0ki of the filter for the noise frequency component selected for processing.
  • S42222 Design a filter containing a notch pair according to the determined distance and cutoff frequency to filter the noise frequency components at (u i1 , v i1 ) and (u i2 , v i2 ) in the image to obtain filtered image.
  • the expression of the second-order Butterworth trapping rejection filter H NRi (u, v) containing a notch pair is as follows:
  • D ki1 (u, v) and D ki2 (u, v) are the distances determined in step S4221.
  • the process of frequency domain filtering is to multiply the centered spectrum of the image and the filter function to obtain the filtered result.
  • S42223 Inspect the amplitude spectrum of the filtered image to determine whether the amplitudes of the noise frequency components in (u i1 , v i1 ) and (u i2 , v i2 ) and their neighborhoods are reduced to an acceptable range. If reduced to an acceptable range, the next pair of noise frequency components are processed, and if not, the update cutoff frequency is redesigned for filtering until the new filter reduces the noise frequency component to an acceptable range. Specifically:
  • the symbol max ⁇ represents the maximum value in the set.
  • TH and TF are compared, where TF is the TF in step S421. If TH ⁇ TF, it is considered that the pair of noise frequency components have been filtered out, and the cutoff frequency at this time is the final cutoff frequency of the notch filter; if TH>TF, the value of the cutoff frequency is updated in a certain step size.
  • S4223 Obtain a second-order Butterworth trapping rejection filter according to the determined distance and the determined cutoff frequency, and filter the selected noise frequency component to be processed.
  • S5 Perform brightness adjustment and fusion on the filtered two sub-pictures Bimf and Cimf to obtain a final filtered image Aimf.
  • the method for fusing the filtered two sub-pictures Bimf and Cimf is: the overlapping area takes the mean value, and the non-overlapping area takes the effective value. The sub-picture is first adjusted for brightness before fusion.
  • step S3 we know that the original image Aim size is M ⁇ N n , the sub-picture Bim is the first N columns of the original image, and the sub-picture Cim is the last N columns of the original image. Similarly, the subgraph Bimf contains information from the first N columns, and the subgraph Cimf contains information from the last N columns. The size of the final filtered image Aimf is M ⁇ N n .
  • the particular fusion method is: after the final filtering column first to (N n -N) Aimf image values of the first column to the (N n -N) columns for the sub-column of FIG Bimf Corresponding values, the value of the (N n -N+1) th column to the Nth column of the image Aimf is taken from the (N n -N+1) th column of the subgraph Bimf to the Nth column and the first column of the subgraph Cimf to The average value corresponding to the (2N-N n ) th column, the value of the N+1th column to the N nth column of the image Aimf is the value corresponding to the (NN-N n +1) th column to the Nth column of the subgraph Cimf .
  • FIG. 12 is a schematic diagram of image fusion in an embodiment of the present invention. The concrete implementation of the expression is as follows:
  • Figures 13(a)-(h) show comparisons of several examples of infrared image filtering before and after the embodiment of the present invention.
  • the left column image is the original image
  • the right column image is the filtered result.
  • the present invention can process a large amount of video data in addition to filtering a single image.
  • the method of filtering the entire video data is to first edit the infrared video data into a short segment according to the difference in noise intensity. video.
  • the noise present in the images of successive frames is considered to be the same during this period of time. Therefore, for successive frames in this period of time, we believe that The optimal intercept lengths are the same, and the noise frequency components in each subgraph are also the same.
  • the present invention analyzes one of the frames by using the method described in steps S1-S5, obtains the optimal intercept length and the corresponding notch comb filter, and then directly processes the video. Other frame images in . In this way, the amount of calculation will be reduced without degrading the quality of the filtering.
  • FIG. 14 is a system structural diagram of a method for removing stray noise of an infrared image according to an embodiment of the present invention. For the convenience of description, only parts related to the present invention are shown.
  • the stripe noise removal system in the infrared image comprises: an estimation module 1, a query module 2, a segmentation module 3, a filtering module 4, and a fusion module 5; wherein the estimation module 1 is configured to select a specific row data of the image Aim to be filtered, And obtaining the width N1 of the stripe by using the spectrum information of the specific line data; the query module 2 is configured to intercept the specific line sequence of the image into (N1+1) new sequences of different lengths according to the width N1 of the stripe, and analyze each The amplitude spectrum information of the new sequence obtains an optimal intercept length N; the segmentation module 3 is configured to slice the image to be filtered Aim of size N ⁇ N n into a pair of sizes M ⁇ N according to the optimal intercept length N.
  • the preferred filter sub-pictures Bim and Cim; the filtering module 4 is used for frequency filtering the two best to-be-filtered sub-pictures Bim and Cim by using a notch comb filter to obtain a first sub-picture Bimf and a second sub-picture Cimf
  • the fusion module 5 is configured to perform fusion processing on the first sub-picture Bimf and the second sub-picture Cimf to obtain an image Aimf.
  • the specific method for stripe noise removal in the infrared image proposed by the present invention is to first divide the image into two sub-graphs with overlapping regions according to the periodicity of the periodicity of the noise, and then perform frequency domain filtering on the two sub-pictures.
  • frequency domain filtering the noise frequency components in the frequency domain are first found according to the characteristics of the discrete Fourier transform of the finite long period sequence. Then, for each pair of noise frequency components, an adaptive Butterworth notch filter is designed. Filtering. Finally, the two filtered subgraphs are fused to obtain the final result.
  • the manual participation is small, the batch processing efficiency is high, and the filtering effect is good, and is particularly suitable for processing a large amount of infrared video data including stripe noise.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

一种红外图像中条纹噪声的去除方法，包括下述步骤：S1：选取待滤波图像Ain的特定行数据，并通过所述特定行数据的频谱信息获得条纹的宽度N1；S2：根据条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；S3：根据最优的截取长度N将大小为M×N n的待滤波图像Ain切分为一对大小为M×N的最佳待滤波子图Bim和Cim；S4：采用陷波梳状滤波器分别对两个最佳待滤波子图Bim和Cim进行滤波后获得第一子图Bimf和第二子图Cimf；S5：将第一子图Bimf和第二子图Cimf进行亮度调整及融合处理，获得图像Aimf。相对于现有的条纹噪声去除方法，上述方案人工参与少，批量处理效率高，滤波效果好，特别适合于处理大量包含条纹噪声的红外视频数据。

Description

一种红外图像中条纹噪声的去除方法及系统 技术领域
本发明属于图像处理技术领域，更具体地，涉及一种红外图像中条纹噪声的去除方法及系统。
背景技术
红外技术已被应用于生物、医学、地学、军事侦察等各个方面。相比于可见光图像，红外图像具有较低的信噪比，容易被各种噪声污染，其中，条纹噪声就是经常存在于红外图像中的一种噪声。条纹噪声产生的原因比较复杂，例如，多传感器的工艺差异、仪器及元件的老化和内部校准系统的误差等导致的各个探测单元的转换传递函数不同；红外图像采集过程中外在环境如温度、周围其他设备的干扰等。条纹噪声的存在干扰了红外图像中有用信息的检测和提取，尤其使一些基于红外图像的识别和目标跟踪性能受到影响。因此，使用条纹噪声去除算法首先将红外图像中的条纹噪声去除，将大大改善后续对其进行处理和分析的可靠性。
图像中的条纹噪声一般是一种周期性噪声，表现在频域上是在固定频点上出现噪声分量。为了滤除掉这种周期噪声频率分量，首先需要找到噪声频率分量在频率域上的对应位置，然后使用陷波滤波器滤除噪声，最后再将信号恢复到时域或空间域。其中的关键是如何寻找到准确的噪声频点，以及如何设计合适的陷波滤波器。
现有技术的图像条纹去除方法存在的问题之一是噪声频率分量的位置确定不够智能与准确。现有方法之一采用人工检测噪声频率分量对应位置。该方法十分耗时，计算效率低，不能满足大数据量的应用需求，且受人工主观因素的影响；现有方法之二根据图像幅度谱的投影图来定位条纹噪声的频率点。该方法将图像的幅度谱分别按行方向或列方向进行投影，再通过投影后的累计分布函数得到噪声频率分量的位置。该方法相比于人工方法，提高了效率。
现有技术的图像条纹去除方法存在的另一个问题周期对齐问题。图像中的周期噪声只有在采集的信号具有严格整周期时，噪声的频率分量才会集中到有限的频点上。若图像中的条纹噪声不具有严格的整周期性，噪声的频率分量会不同程度地扩散在整个频率域上。如果在滤波之前不对图像进行周期对齐的预处理，则使用陷波滤波器方法不能完全滤除条纹噪声的频率分量。
发明内容
针对现有技术的缺陷，本发明的目的在于提供一种红外图像中条纹噪声的去除方法及系统，旨在解决现有的条纹噪声去除的技术中不能自动、高效率处理红外图像中含有不同强度条纹噪声的问题。
为实现上述目的，本发明提供了一种红外图像中条纹噪声的去除方法，包括下述步骤：
S1：选取待滤波图像Aim的特定行数据，并通过所述特定行数据的频谱信息获得条纹的宽度N1；
S2：根据所述条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；
S3：根据最优的截取长度N将大小为M×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim；
S4：采用陷波梳状滤波器分别对两个最佳待滤波子图Bim和Cim进行频率滤波后获得第一子图Bimf和第二子图Cimf；
S5：将所述第一子图Bimf和所述第二子图Cimf进行亮度调整及融合处理，获得图像Aimf。
更进一步地，在步骤S1中，所述条纹的宽度N1，
Figure PCTCN2016082581-appb-000001
符号
Figure PCTCN2016082581-appb-000002
表示向下取整，即去掉小数部分，对于整数保持不变；Nn表示一行数据的长度；km表示中高频处的第一个峰值所对应的幅度谱坐标。
更进一步地，步骤S2包括以下步骤：
S21：将待滤波图像Aim中长度为Nn的特定行数据作为一个Nn点的离散序列x(n)，根据条纹的宽度N1将所述离散序列x(n)截取为N1+1中不同长度的新序列，并计算不同长度的新序列的幅度谱；
S22：提取步骤S21中各个不同长度的新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中代表噪声分量的第一个峰值作为一组特征值即{|XNn-0(k＝km0)|,|XNn-1(k＝km1)|,|XNn-2(k＝km2)|,......,|XNn-N1(k＝kmN1)|}，km0,km1,...,kmN1为分别为新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中第一峰值所对应的幅度谱坐标km，计算这一组特征值的极大值，并获得最优的截取长度N。
更进一步地，在步骤S21中，所述根据条纹的宽度N1将所述离散序列x(n)截取为N1+1种不同长度的新序列具体为：以序列的第一个点为起点，依次截取前Nn点、前Nn-1点、前Nn-2点、……、前Nn-N1点作为新序列，记为xNn-0(n)、xNn-1(n)、xNn-2(n)、……xNn-N1(n)。
更进一步地，在步骤S22中，采用遍历局部区间[0.5km 1.5km]找出局部最大值，获得噪声频率分量的第一个峰值。
更进一步地，在步骤S4中，通过检测最佳待滤波子图Bim和Cim中的噪声频率分量所在位置来确定陷波梳状滤波器的参数。
更进一步地，所述确定陷波梳状滤波器的参数步骤具体为：
选择待处理的噪声频率分量，并计算各点到该噪声频率分量对应的坐标的距离Dki1(u,v)和Dki2(u,v)；
根据所选择待处理的噪声频率分量来获得截止频率D0ki；
根据所述距离和所述截止频率获得陷波梳状滤波器的参数。
更进一步地，在步骤S5中，所述亮度调整具体为：取所述第一子图Bimf和所述第二子图Cimf的 直流分量的均值作为各自的直流分量。所述融合处理具体为：在所述第一子图Bimf和所述第二子图Cimf的重叠区域取均值，在所述第一子图Bimf和所述第二子图Cimf的非重叠区域取有效值。
本发明还提供了一种红外图像中条纹噪声的去除系统，包括：
估计模块，用于选取待滤波图像Aim的特定行数据，并通过所述特定行数据的频谱信息获得条纹的宽度N1；
查询模块，用于根据所述条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；
分割模块，用于根据最优的截取长度N将大小为M×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim；
滤波模块，用于采用陷波梳状滤波器分别对两个最佳待滤波子图Bim和Cim进行频率滤波后获得第一子图Bimf和第二子图Cimf；以及
融合模块，用于将所述第一子图Bimf和所述第二子图Cimf进行亮度调整及融合处理，获得图像Aimf。
本发明实现了一种红外图像条纹噪声的去除方法及系统，较好地解决了现有的条纹噪声去除的技术中，不能自动、高效率处理红外图像中含有不同强度条纹噪声的问题。具体地，先根据噪声的周期性选择周期对齐的长度，将图像分为两个有重叠区域的子图，之后对两个子图进行频域滤波。对于频域滤波，首先根据有限长周期序列的离散傅里叶变换的特点找到频域中的噪声频率分量，然后对于每对噪声频率分量，设计自适应巴特沃斯陷波帯阻滤波器进行滤波；最后将两个滤波后的子图进行亮度调整及融合，得到最后的结果。相对于现有的条纹噪声去除方法，人工参与少，批量处理效率高，滤波效果好，特别适合于处理大量包含条纹噪声的红外视频数据。
附图说明
图1是本发明实施例提供的红外图像条纹噪声的去除方法的流程图；
图2(a)是本发明实施例中所描述的一个子序列的示意图；
图2(b)是本发明实施例中图2(a)所示序列的幅度谱图；
图2(c)是本发明实施例中所描述的由图2(a)所示的一个子序列周期延拓后得到的有限长周期序列的示意图；
图2(d)是本发明实施例中图2(c)所示序列的幅度谱图；
图3(a)是本发明实施例中一个待处理的红外图像；
图3(b)是本发明实施例中图3(a)所示图像的第一行数据的幅度谱图；
图4是本发明实施例中寻找最优截取长度的详细流程图；
图5(a)-(m)是本发明实施例中13个不同长度的序列的幅度谱图，粗线为噪声的第一个谱峰；
图5(n)是本发明实施例中图5(a)-(m)13个图中噪声的第一个谱峰组成的谱峰图，粗线为这些谱峰中的极大值；
图6是本发明实施例中将待滤波图像切分为两个最佳待滤波子图的示意图；
图7是本发明实施例中对两个子图进行滤波处理的详细流程图；
图8是本发明实施例中一个包含5对陷波对的一阶巴特沃斯陷波梳状滤波器示意图；
图9是本发明实施例中对其中一对噪声频率分量进行滤除的详细流程图；
图10是本发明实施例中设计二阶巴特沃斯陷波帯阻滤波器的详细流程图；
图11是本发明实施例中确定滤波器的截止频率的详细流程图；
图12是本发明实施例中图像融合的示意图；
图13(a)-(h)是本发明实施例中几例红外图像滤波前后的对比图；
图14是本发明实施例中提供的红外图像条纹噪声的去除方法的系统结构图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
本发明提出的条纹噪声去除方法在寻找噪声频率分量位置时，充分利用了周期噪声在频域上的特性，通过反馈来自动修改滤波器的参数，直到能达到理想的滤波效果。发明提供的方法能够自适应的处理各种强度不同的条纹噪声，不需要人工参与，能够自动、高效、准确的滤除掉图像中的强弱不同的条纹噪声。
现以图像中存在的条纹为竖直条纹为例来说明本发明提出的红外图像中条纹噪声的去除方法，对于横条纹，只需在处理图像之前将图像旋转90度，即可当做竖直条纹来处理。
针对竖直条纹，本发明提出的红外图像中条纹噪声去除的具体方法是先选取图像特定行，通过分析该行的频谱信息估计条纹的宽度。然后利用条纹的宽度将图像的特定行序列截取为不同长度的新序列，通过分析各个新序列的幅度谱信息，找出最优的截取长度。然后根据最优的截取长度将待滤波的原始图像切一对最佳待滤波子图。之后检测子图中噪声频率分量所在位置，设计陷波梳状滤波器对该对最佳待滤波子图进行频率滤波。最后，将滤波后的一对子图进行融合，得到最后的结果。
图1示出了本发明提供的红外图像条纹噪声去除方法的流程，包括以下步骤：
S1：选取待滤波图像Aim的特定行数据，分析该行的频谱信息，估计条纹的宽度N1。
对于含有条纹噪声的图像，假设其条纹噪声是以其条纹的宽度N1为周期的周期噪声。对于含有竖直条纹噪声的图像，相当于图像中的每一行序列都叠加了一个以条纹的宽度N1为周期的有限长周期噪声序列。所谓有限长周期序列是指将一个子序列进行多次周期延拓得到的。如图2(a)所示为一个6点离散序列，图2(c)所示为以图2(a)中所示的6点离散序列作为子序列进行9次周期延拓得到的新序列(即含有10个子序列)，我们称这个新序列为有限长周期序列。如图2(b)所示为图2(a)中所示序列的幅度谱，图2(d)所示为图2(c)所示的有限长周期序列的幅度谱。从图2(b)和图 2(d)可以看出，子序列经过9次周期延拓之后，有限长周期序列的幅度谱与子序列的幅度谱以及延拓的次数有很大的关系，有限长周期序列的幅度谱相当于是先在子序列的幅度谱的所有相邻两个值之间插入9个零值，然后将所有幅值乘以10得到。这一现象并不是偶然，对于所有有限长周期序列，若该序列由一个子序列周期延拓m次得到，则该序列的幅度谱可以通过对子序列的幅度谱进行插入零值以及放大处理得到，具体的做法是先在子序列幅度谱中的所有相邻两个值中插入m个零值，然后将所有值放大(m+1)倍。有限长周期序列的这点性质可以通过离散傅里叶变换进行证明，此处不详细解释。
本发明中，获得条纹宽度N1的方法是：选取原图像Aim的特定行数据(假设一行数据的长度为Nn)作为一个Nn点的离散序列x(n)，其中特定行可以为任意行，但中间行数据更稳定，所以本发明中取中间行作为特定行。根据离散傅里叶变换的公式求出x(n)的离散傅里叶变换X(k)，如公式(1)所示。再求X(k)的绝对值|X(k)|得到x(n)的幅度谱。本发明中是采用快速傅里叶变换求X(k)，在接下来所有求一个序列的频谱时，包括求一个二维序列的频谱，本发明中都采用快速傅里叶变换。
Figure PCTCN2016082581-appb-000003
如图3(a)所示为一幅待滤波的红外图像a，其中图3(a)中两条直线所框出的即为一个完整的条纹。如前所述，有限长周期序列的幅度谱相当于是在其子序列的幅度谱上插入零值再放大的结果，对于含有条纹噪声的图像，其中的条纹噪声相当于是由一个子条纹多次周期延拓的结果，因此，噪声的频率分量相对于整幅图像的频率分量将会出现在频谱图的等间隔位置，并且噪声的频率分量的幅值会较大，表现为局部峰值。如图3(b)所示为图3(a)的中间行序列的幅度谱，该幅度谱是未中心化的结果。其中在中高频处出现的峰值就是代表噪声的频率分量(直流分量不考虑)。
为了估计条纹的宽度，本发明所采用的方法是首先找到中高频处的第一个峰值所对应的幅度谱坐标km，则条纹宽度N1确定为
Figure PCTCN2016082581-appb-000004
符号
Figure PCTCN2016082581-appb-000005
表示向下取整，即去掉小数部分，对于整数保持不变。
本发明中获取噪声的第一峰值所对应的幅度谱坐标km的方法是，对于幅度谱|X(k)|，令SXk为以|X(k)|为中心的11个元素的集合{|X(k+i)|,i＝-5,-4,…,0,1,…,4,5}，max[SXk]为集合SXk中具有最大值的那个元素的值，med[SXk]为集合SXk中具有中值的那个元素的值，即将集合中的元素按照由大到小的顺序排序，排在最中间的那个元素的值。则km由下面的公式计算得到：
Figure PCTCN2016082581-appb-000006
且
Figure PCTCN2016082581-appb-000007
上式表示，km是一个满足|X(k)|为局部最大值，且该局部最大值大于局部中值的那些k值中的最小值。其中上式中的θ是一个阈值，表示局部最大值与局部中值的差异，当阈值越大表示局部最大值与局部 中值差异越大。本发明中对于图像中所含噪声相对较强时选取阈值θ＝5，相对较弱时选取阈值θ＝2。阈值的选取应根据实际情况进行调整。
例如，对图3(a)所示的图像a，一行数据的长度Nn＝485，求得的中高频处的第一个峰值所对应的幅度谱km＝45，求得其中的条纹宽度N1＝12。
S2：利用条纹的宽度N1，将图像的特定行序列截取为(N1+1)种不同长度的新序列，分析各个新序列的幅度谱信息，找出最优的截取长度N。
其中最优的截取长度是指：从原始序列中截取的某一种长度的新序列，其频谱中的噪声频率分量最集中，则该序列的长度即为最优的截取长度。由于对于原始序列，去掉最后一个点，最后两个点，或者最后多个点，其频谱中的噪声频率分量的分布会发生明显的变化。为此，我们旨在找到一个截取长度，使得该长度下的序列，其频谱中的噪声频率分量最集中，以便于后续的滤波。
进一步地，如图4所示，步骤S2包括以下步骤：
S21:将待滤波图像Aim的特定行数据(数据的长度为Nn)作为一个Nn点的离散序列x(n)，根据条纹的宽度N1将x(n)截取为不同长度的新序列，计算不同长度的新序列的幅度谱。
其中特定行同样取中间行，即取中间行作为一个Nn点的离散序列x(n)，根据条纹的宽度N1将x(n)截取为不同长度的新序列是指：以序列的第一个点为起点，依次截取前Nn点、前Nn-1点、前Nn-2点、……、前Nn-N1点作为新序列，记为xNn-0(n)、xNn-1(n)、xNn-2(n)、……xNn-N1(n)。同样，根据离散傅里叶变换求出各序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|。以图3(a)所示的图像a的中间行数据作为原始数据，截取得到13个不同长度的新序列，求出各序列的幅度谱，如图5(a)-(m)所示。从图5(a)-(m)可以看出，在一个序列的基础上增加或者减少一个点，其相应的幅度谱会出现较大的差异，尤其是代表周期噪声的频率分量会有明显的变化。
S22：提取步骤S21中各个不同长度的新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中代表噪声分量的第一个峰值作为一组特征值即{|XNn-0(k＝km0)|,|XNn-1(k＝km1)|,|XNn-2(k＝km2)|,......,|XNn-N1(k＝kmN1)|}，km0,km1,...,kmN1分别为新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中第一峰值所对应的幅度谱坐标km，求出这一组特征值的极值，从而得到最优的截取长度N。
其中，代表噪声分量的第一个峰值是指去掉噪声的直流分量的第一个峰值，也就是中高频处的第一个峰值。如图5(a)-(m)所示，各个幅度谱图中的粗的峰值脉冲谱线即为代表噪声频率分量的第一个峰值。
本发明采用的是遍历局部区间，找出局部最大值，即噪声频率分量的第一个峰值。在步骤S1中已经给出了如何求第一峰值所对应的幅度谱坐标km，此处，各个不同长度的新序列的幅度谱中代表噪声分量的第一个峰值对应的坐标一定在坐标km附近。因此，本发明遍历局部区间[0.5km,1.5km]的找到局部最大值，并提取出所有的第一个峰值作为一组特征值，如图5(n)所示为提取步骤S21中 |XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|代表噪声分量的第一个峰值，N1＝12，因此共有13个值，找到其中的极大值，即图5(n)中粗线代表的值。则该值对应的序列的长度即为最优的截取长度。以实施例来说，求出图5(n)中极大值的对应坐标p，此实施例中p＝9，Nn＝485，则对应序列的长度为Nn-(p-1)＝477，则最优的截取长度N＝477。即最优的截取长度N＝Nn-(p-1)，其中Nn为序列的长度，(p-1)为去掉的点数。
其中寻找极值的方法是先找出这一组特征值的最大值，然后判断最大值是否是第一个值，若不是第一个值，则认为最大值即为所求极值，若是第一个值，则去掉该组特征值第一个值重新找数组中剩下的值的最大值，直到最大值不是第一个值为止。
S3：根据最优的截取长度N，将大小为M×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim。
本发明中，现将图像Aim切分为一对最佳待滤波子图Bim和Cim所采取的方法是：取前N列为一个子图Bim，取后N列为另一个子图Cim。现结合实施例来说明这个过程。对于图3(a)所示的图像a，求得的N＝477，则前477列作为一个子图b，后477列作为子图c。如图6所示，图(b)代表子图b，图(c)代表子图c。
S4：检测各子图中的噪声频率分量所在位置，确定滤波器的各种参数，设计陷波梳状滤波器，分别对两个最佳待滤波子图Bim和Cim进行频率滤波，得到滤波后的子图Bimf和Cimf。
本发明中，根据条纹噪声的频率特性，选择陷波梳状滤波器对最佳待滤波子图进行滤波。其中，陷波滤波器有理想型，巴特沃斯型以及高斯型。本发明中选择巴特沃斯陷波滤波器。设计陷波滤波器的方法是先检测出噪声的所有可能的频率分量所在的坐标位置，然后根据每对噪声频率分量幅值的大小设计不同的陷波对。本发明中采用的是二阶巴特沃斯陷波带阻滤波器，不同的陷波对之间其主要的差异体现在中心和截止频率的不同。由于陷波对的中心等间隔出现，滤波器呈梳状。
进一步地，如图7所示，步骤S4包括以下步骤：
S41：选择其中一个子图Bim进行频谱分析，估计子图Bim中所有可能含有噪声频率分量的总对数Np，及对应的坐标(ui1,vi1)和(ui2,vi2)，其中i＝1...Np，同时坐标(ui1,vi1)和(ui2,vi2)也是陷波器的中心。
本发明中，首先选择其中一个子图Bim进行频谱分析，估计子图Bim中所有可能含有噪声频率分量的频率分量的总对数Np，其中估计总对数Np的方法是取出待滤波子图Bim的中间行序列记作xb(n)，该序列的长度为N，求出其幅度谱|Xb(k)|。找到噪声的各频率分量之间的间距，本发明中，采用的方法类似于步骤S22中所述的遍历局部区间[0.5km,1.5km]，其中km为步骤S1中所求，找出局部最大值，即噪声频率分量的第一个峰值，假设第一个峰值对应的横坐标为q，则噪声的各频率分量之间的间距δ＝q-1。
通过噪声各频率分量之间的间距δ，可以知道所有可能包含噪声频率分量的点。根据离散傅里叶变换的对称性可知，噪声的频率分量也是成对出现的。首先根据间距δ求出噪声频率分量最多存在的 对数Np，其求解公式如公式(2)所示，符号
Figure PCTCN2016082581-appb-000008
表示向下取整，即去掉小数部分，对于整数保持不变。
Figure PCTCN2016082581-appb-000009
对于存在竖直条纹图像，取其一行做一维离散傅里叶变换得到的幅度谱|Xb(k)|与对整个图像做二维离散傅里叶变换得到的幅度谱|F(u,v)|之间存在着一定的关系，两者的条纹噪声的频率分量间距相等。图像中的竖直条纹集中在傅里叶频域能量谱的水平轴上，其频率分量也是以间距δ等间隔出现。将图像的频谱F(u,v)中心化，得到中心化后的频谱F(u,v)及幅度谱FA(u,v)＝|F(u,v)|。根据公式(3)和公式(4)得到Np对噪声频率分量所在的位置。其中(ui1,vi1)和(ui2,vi2)处的频率分量是可能包含噪声频率分量的一对关于中心对称的频率分量，下标i表示不同的频率分量对。M指图像的行数，N指图像的列数，即处理的图像大小为M×N。
Figure PCTCN2016082581-appb-000010
Figure PCTCN2016082581-appb-000011
S42：针对子图中的噪声频率分量，设计巴特沃斯陷波梳状帯阻滤波器，并对两个子图Bim和Cim进行频率滤波，得到滤波后的子图Bimf和Cimf。
本发明中，陷波梳状滤波器是用中心已被平移到陷波滤波器中心的高通滤波器的乘积来构造。其形式如公式(5)所示。
Figure PCTCN2016082581-appb-000012
其中，Q表示含有的陷波对的数量，Hk(u,v)和H-k(u,v)是高通滤波器，它们的中心分别在(uk,xk)和(-uk,-vk)处。其中，中心是根据频率矩形的中心
Figure PCTCN2016082581-appb-000013
来确定的，并且频率矩形的大小与待滤波图像的大小要相同。对于每个滤波器，距离由公式(6)和公式(7)计算：
Figure PCTCN2016082581-appb-000014
Figure PCTCN2016082581-appb-000015
本发明中采用巴特沃斯陷波梳状滤波器，下面是一个n阶巴特沃斯陷波梳状滤波器，它包含三个陷波对：
Figure PCTCN2016082581-appb-000016
其中，Dk和D-k由式(6)和式(7)给出。常数D0k对每一个陷波对都是相同的，对于不同的陷波对它可以不同。如图8所示为一个包含5对陷波对的一阶巴特沃斯陷波梳状滤波器。
从步骤S41中得知，噪声的频率分量最多为Np对，也就是说少于或等于Np对。本发明实施例中，实现频域滤波的方法并不是直接设计一个含有Np个陷波对的滤波器，而是多次设计只含有一个陷波对的滤波器对图像中的每一对噪声分量依次进行滤波，直到所有的噪声频率分量都被滤除。
现以实施例来说明整个滤波过程。
设f(x,y)是大小为M×N的数字图像，则其二维傅里叶变换F(u,v)由公式(8)得到：
Figure PCTCN2016082581-appb-000017
在计算F(u,v)之前，由(-1)x+y乘以f(x,y)，然后进行傅里叶变换可以得到图像的中心化后的频谱。在接下来的滤波处理过程中均假设F(u,v)表示中心化后的频谱。通过步骤S41中找出的含有噪声的频率分量所有可能的频率分量在(ui1,vi1)和(ui2,vi2)处，其中i＝1...Np。进一步地，如图9所示，具体的滤波步骤包括以下步骤：
S421：判断图像的频谱中，(ui1,vi1)和(ui2,vi2)处的频率分量是否真的包含噪声频率分量，若判断为包含，进行接下来的滤波处理，若判断为不包含，对该处的频率分量不做任何处理。
本发明中，在步骤S41中找到的噪声频率分量不一定是真的包含有噪声的频率分量，因此，在滤除之前，首先判断这一对点处的频率分量是否真的包含噪声频率分量的信息。
本发明中判断的方法是，先求得图像中心化后的幅度谱FA(u,v)＝|F(u,v)|。由于幅度关于中心对称，以(ui1,vi1)为中心，选取大小为(2m+1)(2n+1)的邻域为SFAi，同时以(ui1,vi1)为中心，选取大小为(2m+3)(2n+3)的邻域为SFAEi。其中邻域SFAi和SFAEi定义如下：
Figure PCTCN2016082581-appb-000018
Figure PCTCN2016082581-appb-000019
则是否需要对该频点进行噪声滤波的判定阈值TF由下式计算得到：
TF＝med{FA(u,v)|FA(u,v)∈SFAEi-SFAi}
其中SFAEi-SFAi表示所有属于集合SFAEi但不属于集合SFAi的元素的集合，符号med{·}表示求集合中的中值，即将集合中的元素按照由大到小的顺序排列，排在最中间的那个元素的值。本发明中一般初始选取m＝n＝2，如果滤除效果不好可根据实际情况增大邻域大小。计算集合SFAi的元素极大值，
T＝max{FA(u,v)|FA(u,v)∈SFAi}
其中{FA(u,v)|FA(u,v)∈SFAi}表示所有属于邻域SFAi元素的集合，符号max{·}表示求集合中的最大值。
若T＞TF，则认为(ui1,vi1)和(ui2,vi2)处的频率分量包含噪声频率分量，需要进行滤除；否则，若T＜TF，则认为(ui1,vi1)和(ui2,vi2)处的频率分量不包含的噪声频率分量，无需进行滤除处理。然后，考察下一对可能包含噪声频率分量的频率成分。
S422：针对包含噪声频率分量的所有频率分量分别设计巴特沃斯陷波带阻滤波器，将噪声频率分量滤除，得到滤波后的子图Bimf和Cimf。
对于每对噪声频率分量所在的位置，一般其幅值都各不相同，所以针对不同的噪声频率分量，应设计不同的陷波滤波器。本发明中，对于不同的噪声分量，设计陷波滤波器的思路是一样的，接下来将以其中一对噪声频率分量为例，来说明设计陷波滤波器的过程。
假设第i对频率分量，即(ui1,vi1)和(ui2,vi2)处的频率分量包含噪声频率分量，现为其设计滤波器进行滤除。进一步的，如图9所示，设计滤波器的步骤包括以下几步：
S4221：选择要处理的噪声频率分量，计算各点到该噪声频率分量对应的坐标的距离，即确定距离Dki1(u,v)和Dki2(u,v)。
其中(ui1,vi1)和(ui2,vi2)为滤波器的中心，此处滤波器的中心已经是中心化后的结果，根据公式(6)和公式(7)计算距离如下：
Dki1(u,v)＝[(u-ui1)2+(v-vi1)2]1/2
Dki2(u,v)＝[(u-ui2)2+(v-vi2)2]1/2
根据公式(3)和公式(4)可知，
Figure PCTCN2016082581-appb-000020
M为频率矩形的行数，N为频率矩形的列数，δ为步骤S41中求出的噪声各频率分量之间的间距。距离确定之后，然后确定滤波器的截止频率D0ki。
S4222：对所选择要处理的噪声频率分量，确定截止频率D0ki。
本发明实施例中，确定滤波器的截止频率的方法是先初始化一个较小的截止频率，得到一个滤波函数对图像进行滤波，然后考察滤波后的结果来更新截止频率，直到滤波后的结果达到要求。进一步的，如图10所示，确定截止频率并对图像实现频域滤波的步骤包括以下几步：
S42221：针对所选择要处理的噪声频率分量，初始化滤波器的截止频率D0ki。
本发明实施例中，首先将截止频率D0ki初始化为一个较小的值，初始截止频率D0ki＝1。
S42222：根据已确定的距离和截止频率设计含有一个陷波对的滤波器对图像中的(ui1,vi1)和(ui2,vi2)处的噪声频率分量进行滤波，得到滤波后的图像。
本发明实施例中，设计的含有一个陷波对的二阶巴特沃斯陷波帯阻滤波器HNRi(u,v)的表达式如下：
Figure PCTCN2016082581-appb-000021
其中，Dki1(u,v)和Dki2(u,v)是步骤S4221中确定的距离。
频率域滤波的过程是将图像的中心化后频谱与滤波函数相乘，得到滤波后的结果
Figure PCTCN2016082581-appb-000022
Figure PCTCN2016082581-appb-000023
S42223：对滤波后的图像的幅度谱进行考察，判断(ui1,vi1)和(ui2,vi2)处，及其邻域内的噪声频率分量幅值是否减小到可接受的范围。若减小到可接受的范围，则处理下一对噪声频率分量，若没有，更新截止频率重新设计滤波器进行滤除，直到新的滤波器将噪声频率分量减小到可接受的范围。具体为：
先求滤波后的结果
Figure PCTCN2016082581-appb-000024
的幅度谱
Figure PCTCN2016082581-appb-000025
定义以(ui1,vi1)为中心，大小为(2m+1)(2n+1)的邻域为
Figure PCTCN2016082581-appb-000026
如下：
Figure PCTCN2016082581-appb-000027
令，
Figure PCTCN2016082581-appb-000028
其中
Figure PCTCN2016082581-appb-000029
表示所有属于邻域
Figure PCTCN2016082581-appb-000030
元素的集合，符号max{·}表示求集合中的最大值。
比较TH与TF的大小，其中TF为步骤S421中的TF。若TH<TF，则认为该对噪声频率分量已被滤除，此时的截止频率即为该陷波滤波器的最终截止频率；若TH>TF，则以一定的步长更新截止频率的值，本发明实施例中用D0ki＝D0ki+5进行更新，然后转到步骤S42222，重新设计滤波器 HNRi(u,v)，并对该对噪声频率分量重新进行滤波处理。当多次更新截止频率后，对图像进行滤波能使得TH<TF，则认为此时的截止频率为最终的截止频率。
S4223：根据上述确定的距离和确定的截止频率得到二阶巴特沃斯陷波帯阻滤波器，对所选择要处理的噪声频率分量进行滤除。
S5：将滤波后的两个子图Bimf和Cimf进行亮度调整及融合，得到最终的滤波后图像Aimf。其中，本发明将滤波后的两个子图Bimf和Cimf进行融合的方法是：重叠区域取均值，非重叠区域取有效值。在融合之前首先对子图进行亮度调整。
设子图Bim和Cim经过频率域滤波之后的频谱分别为
Figure PCTCN2016082581-appb-000031
和
Figure PCTCN2016082581-appb-000032
由于子图之间的噪声以及有用信息的差异，滤波之后子图中的直流分量不一致，也就是说两个子图的亮度不统一。因此，在融合子图之前，因调整两子图的直流分量。具体的做法是取
Figure PCTCN2016082581-appb-000033
和
Figure PCTCN2016082581-appb-000034
的直流分量的均值作为各自的直流分量。即令：
Figure PCTCN2016082581-appb-000035
然后令：
Figure PCTCN2016082581-appb-000036
直流分量调整后，再通过傅里叶逆变换将图像还原到空间域。假设
Figure PCTCN2016082581-appb-000037
和
Figure PCTCN2016082581-appb-000038
经过亮度调整并通过傅里叶逆变换还原到空间域后的子图Bimf和Cimf。
在步骤S3中我们知道，原始图像Aim大小为M×Nn，子图Bim为原始图像的前N列，子图Cim为原始图像的后N列。同样，子图Bimf包含了前N列的信息，子图Cimf包含了后N列的信息。最终滤波后图像Aimf的大小为M×Nn。如图11所示，融合的具体方法是：最终的滤波后图像Aimf的第1列到第(Nn-N)列的值取子图Bimf的第1列到第(Nn-N)列对应的值，图像Aimf的第(Nn-N+1)列到第N列的值取子图Bimf的第(Nn-N+1)列到第N列和子图Cimf的第1列到第(2N-Nn)列对应的平均值，图像Aimf的第N+1列到第Nn列的值取子图Cimf的第(2N-Nn+1)列到第N列对应的值。如图12所示是本发明实施例中图像融合的示意图。具体实现的表达式如下所示：
Figure PCTCN2016082581-appb-000039
其中，1≤i≤M，(i,j)表示数字图像的像素点的坐标。
图13(a)-(h)给出了本发明实施例中几例红外图像滤波前后的对比图。其中，左列图像为原始图像，右列图像为滤波后的结果。
另外，本发明除了可以对单张图像进行滤波处理外，也可以处理大量的视频数据。本发明中在对整个视频数据进行滤波处理的方法是先将红外视频数据按照噪声强度的不同剪辑为一小段一小段的 视频。通常，在同一时间段，由于器件本身状况以及当时环境类似，在这一段时间内，连续帧的图像中存在的噪声也认为是相同的，因此，对于这一段时间内的连续帧图像，我们认为最优的截取长度是相同的，各子图中的噪声频率分量也是相同的。本发明在处理同一时间段内的图片时，通过利用步骤S1-S5所述的方法来分析其中一帧图像，得到最优截取长度以及对应的陷波梳状滤波器，然后直接处理该段视频中的其他帧图像。这样，将减少计算量，也不会降低滤波的质量。
图14示出了本发明实施例提供的红外图像条纹噪声的去除方法的系统结构图。为了便于说明，仅示出了与本发明相关的部分。
本发明提供的红外图像中条纹噪声的去除系统包括：估计模块1、查询模块2、分割模块3、滤波模块4和融合模块5；其中估计模块1用于选取待滤波图像Aim的特定行数据，并通过特定行数据的频谱信息获得条纹的宽度N1；查询模块2用于根据所述条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；分割模块3用于根据最优的截取长度N将大小为N×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim；滤波模块4用于采用陷波梳状滤波器分别对两个最佳待滤波子图Bim和Cim进行频率滤波后获得第一子图Bimf和第二子图Cimf；融合模块5用于将第一子图Bimf和所述第二子图Cimf进行融合处理，获得图像Aimf。
本发明提出的红外图像中条纹噪声去除的具体方法是先根据噪声的周期性选择周期对齐的长度，将图像分为两个有重叠区域的子图，之后对两个子图进行频域滤波。对于频域滤波，首先根据有限长周期序列的离散傅里叶变换的特点找到频域中的噪声频率分量，之后，对于每对噪声频率分量，设计自适应巴特沃斯陷波帯阻滤波器进行滤波。最后，将两个滤波后的子图进行融合，得到最后的结果。相对于现有的条纹噪声去除方法，人工参与少，批量处理效率高，滤波效果好，特别适合于处理大量包含条纹噪声的红外视频数据。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分步骤是可以通过程序来控制相关的硬件完成，所述的程序可以在存储于一计算机可读取存储介质中，所述的存储介质，如ROM/RAM、磁盘、光盘等。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (9)

  1. 一种红外图像中条纹噪声的去除方法，其特征在于，包括下述步骤：
    S1：选取待滤波图像Aim的特定行数据，并通过所述特定行数据的频谱信息获得条纹的宽度N1；
    S2：根据所述条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；
    S3：根据最优的截取长度N将大小为M×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim；
    S4：采用陷波梳状滤波器分别对两个最佳待滤波子图Bim和Cim进行滤波后获得第一子图Bimf和第二子图Cimf；
    S5：将所述第一子图Bimf和所述第二子图Cimf进行亮度调整及融合处理，获得图像Aimf。
  2. 如权利要求1所述的去除方法，其特征在于，在步骤S1中，所述条纹的宽度
    Figure PCTCN2016082581-appb-100001
    符号
    Figure PCTCN2016082581-appb-100002
    表示向下取整，即去掉小数部分，对于整数保持不变；Nn表示一行数据的长度；km表示中高频处的第一个峰值所对应的幅度谱坐标。
  3. 如权利要求1所述的去除方法，其特征在于，步骤S2包括以下步骤：
    S21：将待滤波图像Aim中长度为Nn的特定行数据作为一个Nn点的离散序列x(n)，根据条纹的宽度N1将所述离散序列x(n)截取为N1+1种不同长度的新序列，并计算不同长度的新序列的幅度谱；
    S22：提取步骤S21中各个不同长度的新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中代表噪声分量的第一个峰值作为一组特征值即{|XNn-0(k＝km0)|,|XNn-1(k＝km1)|,|XNn-2(k＝km2)|,......,|XNn-N1(k＝kmN1)|},km0,km1,...,kmN1分别为新序列的幅度谱|XNn-0(k)|、|XNn-1(k)|、|XNn-2(k)|、……、|XNn-N1(k)|中第一峰值所对应的幅度谱坐标km，计算这一组特征值的极值，并获得最优的截取长度N。
  4. 如权利要求3所述的去除方法，其特征在于，在步骤S21中，所述根据条纹的宽度N1将所述离散序列x(n)截取为不同长度的新序列具体为：
    以序列的第一个点为起点，依次截取前Nn点、前Nn-1点、前Nn-2点、……、前Nn-N1点作为新序列，记为xNn-0(n)、xNn-1(n)、xNn-2(n)、……xNn-N1(n)。
  5. 如权利要求3所述的去除方法，其特征在于，在步骤S22中，采用遍历局部区间[0.5km 1.5km]找出局部最大值的位置，获得噪声频率分量的第一个峰值的位置。
  6. 如权利要求1-5任一项所述的去除方法，其特征在于，在步骤S4中，通过检测最佳待滤波子图Bim和Cim中的噪声频率分量所在位置来确定陷波梳状滤波器的参数。
  7. 如权利要求6所述的去除方法，其特征在于，所述确定陷波梳状滤波器的参数步骤具体为：
    选择待处理的噪声频率分量，并计算各点到该噪声频率分量对应的坐标的距离Dki1(u,v)和Dki2(u,v)；
    根据所选择待处理的噪声频率分量来获得截止频率D0ki；
    根据所述距离和所述截止频率获得陷波梳状滤波器的参数。
  8. 如权利要求1-7任一项所述的去除方法，其特征在于，在步骤S5中，所述亮度调整具体为：取所述第一子图Bimf和所述第二子图Cimf的直流分量的均值作为各自的直流分量。所述融合处理具体为：在所述第一子图Bimf和所述第二子图Cimf的重叠区域取均值，在所述第一子图Bimf和所述第二子图Cimf的非重叠区域取有效值。
  9. 一种红外图像中条纹噪声的去除系统，其特征在于，包括：
    估计模块，用于选取待滤波图像Aim的特定行数据，并通过所述特定行数据的频谱信息获得条纹的宽度N1；
    查询模块，用于根据所述条纹的宽度N1将图像的特定行序列截取为(N1+1)种不同长度的新序列，并通过分析各个新序列的幅度谱信息获得最优的截取长度N；
    分割模块，用于根据最优的截取长度N将大小为M×Nn的待滤波图像Aim切分为一对大小为M×N的最佳待滤波子图Bim和Cim；
    滤波模块，用于分别对两个最佳待滤波子图Bim和Cim进行频率滤波后获得第一子图Bimf和第二子图Cimf；以及
    融合模块，用于将所述第一子图Bimf和所述第二子图Cimf进行融合处理，获得图像Aimf。
PCT/CN2016/082581 2016-05-19 2016-05-19 一种红外图像中条纹噪声的去除方法及系统 WO2017197618A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
PCT/CN2016/082581 WO2017197618A1 (zh) 2016-05-19 2016-05-19 一种红外图像中条纹噪声的去除方法及系统
CN201680000899.4A CN106462957B (zh) 2016-05-19 2016-05-19 一种红外图像中条纹噪声的去除方法及系统
US15/838,404 US10282824B2 (en) 2016-05-19 2017-12-12 Method and system for eliminating stripe noise in infrared images

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2016/082581 WO2017197618A1 (zh) 2016-05-19 2016-05-19 一种红外图像中条纹噪声的去除方法及系统

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/838,404 Continuation US10282824B2 (en) 2016-05-19 2017-12-12 Method and system for eliminating stripe noise in infrared images

Publications (1)

Publication Number Publication Date
WO2017197618A1 true WO2017197618A1 (zh) 2017-11-23

Family

ID=58215609

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/082581 WO2017197618A1 (zh) 2016-05-19 2016-05-19 一种红外图像中条纹噪声的去除方法及系统

Country Status (3)

Country Link
US (1) US10282824B2 (zh)
CN (1) CN106462957B (zh)
WO (1) WO2017197618A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109509158A (zh) * 2018-11-19 2019-03-22 电子科技大学 基于幅度约束红外图像条纹去除的方法
CN110838093A (zh) * 2019-11-05 2020-02-25 安徽大学 一种高效、高质量的fMOST或MOST显微图像条纹噪声去除方法
CN111383196A (zh) * 2020-03-13 2020-07-07 浙江大华技术股份有限公司 红外图像条纹消除方法、红外探测器及存储装置
CN112017127A (zh) * 2020-08-21 2020-12-01 珀乐(北京)信息科技有限公司 一种基于光谱滤波的图像条带噪声去除方法及系统
CN112862816A (zh) * 2021-03-15 2021-05-28 太原理工大学 一种hrtem图像中煤芳香烃晶格条纹的智能提取方法
CN117830141A (zh) * 2024-03-04 2024-04-05 奥谱天成(成都)信息科技有限公司 红外图像竖条纹噪声去除方法、介质、设备及装置

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107507203B (zh) * 2017-05-02 2020-04-14 大连理工大学 一种基于热红外图像的服务器设备边界直线角度自动提取方法
CN107454963B (zh) * 2017-06-16 2021-11-30 深圳市汇顶科技股份有限公司 指纹图像处理方法及光学指纹辨识系统
CN108305220B (zh) * 2017-12-29 2020-06-02 华中科技大学 一种机载红外退化图像校正方法
CN109360168B (zh) * 2018-10-16 2021-02-12 烟台艾睿光电科技有限公司 红外图像去条纹的方法、装置、红外探测器及存储介质
CN109447920B (zh) * 2018-11-12 2021-02-02 中国人民解放军国防科技大学 基于陷波滤波器的条带噪声自动去除方法
CN111199166B (zh) * 2018-11-16 2023-04-18 首都师范大学 基于频域和空域特征的视频拉条的检测和恢复方法
CN109767403B (zh) * 2019-01-23 2022-12-16 西安科技大学 一种基于场景的红外焦平面阵列条状噪声消除方法
CN110310237B (zh) * 2019-06-06 2020-08-18 武汉精立电子技术有限公司 去除图像摩尔纹、显示面板子像素点亮度测量、Mura缺陷修复的方法及系统
CN110400271B (zh) * 2019-07-09 2021-06-15 浙江大华技术股份有限公司 一种条纹非均匀性校正方法、装置、电子设备及存储介质
CN110796621B (zh) * 2019-10-29 2022-06-07 浙江大华技术股份有限公司 红外图像去横纹处理方法、处理设备和存储装置
CN111311510B (zh) 2020-01-22 2023-08-25 京东方科技集团股份有限公司 图像中条纹噪声的去除方法及装置
CN111289106B (zh) * 2020-03-26 2022-11-11 广西科技大学 一种基于数字滤波的光谱降噪方法
CN112924956B (zh) * 2021-02-05 2022-05-13 中国人民解放军国防科技大学 一种降低低频噪声级的背景均衡方法
CN114998126B (zh) * 2022-05-24 2024-05-24 中国人民解放军国防科技大学 一种低波段星载sar图像电离层幅度闪烁条纹去除方法
CN114881895B (zh) * 2022-07-08 2022-09-30 托伦斯半导体设备启东有限公司 一种基于帧间差异的红外图像条纹噪声处理方法
CN115272137B (zh) * 2022-09-28 2022-12-20 北京万龙精益科技有限公司 基于fpga的实时固定模式噪声去除方法、装置、介质及系统
CN116228596B (zh) * 2023-05-06 2023-06-30 北京师范大学 一种基于即插即用先验的条纹噪声去除方法
CN116738153B (zh) * 2023-08-15 2023-10-17 中国科学院东北地理与农业生态研究所 基于光谱分析的有机肥利用效果评价方法
CN117314798B (zh) * 2023-11-29 2024-05-03 国药新光医疗科技有限公司 一种去除纤维镜图像网格的方法

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103679643A (zh) * 2013-06-03 2014-03-26 哈尔滨工程大学 一种多条纹噪声定位滤除方法
CN104580937A (zh) * 2015-01-21 2015-04-29 中国科学院上海技术物理研究所 一种红外成像系统条纹噪声去除方法

Family Cites Families (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8180168B2 (en) * 2007-01-16 2012-05-15 Hewlett-Packard Development Company, L.P. One-pass filtering and infrared-visible light decorrelation to reduce noise and distortions
US7920746B2 (en) * 2007-04-23 2011-04-05 Aptina Imaging Corporation Compressed domain image summation apparatus, systems, and methods
US9451183B2 (en) * 2009-03-02 2016-09-20 Flir Systems, Inc. Time spaced infrared image enhancement
US9756264B2 (en) * 2009-03-02 2017-09-05 Flir Systems, Inc. Anomalous pixel detection
US8520970B2 (en) * 2010-04-23 2013-08-27 Flir Systems Ab Infrared resolution and contrast enhancement with fusion
US9237284B2 (en) * 2009-03-02 2016-01-12 Flir Systems, Inc. Systems and methods for processing infrared images
US8224112B2 (en) * 2009-06-08 2012-07-17 Xerox Corporation Fuzzy method to detect thin lines in scanned image
CA2838992C (en) * 2011-06-10 2018-05-01 Flir Systems, Inc. Non-uniformity correction techniques for infrared imaging devices
US9819879B2 (en) * 2011-07-12 2017-11-14 Samsung Electronics Co., Ltd. Image filtering apparatus and method based on noise prediction using infrared ray (IR) intensity
JP5871571B2 (ja) * 2011-11-11 2016-03-01 株式会社Pfu 画像処理装置、矩形検出方法及びコンピュータプログラム
JP5854774B2 (ja) * 2011-11-11 2016-02-09 株式会社Pfu 画像処理装置、直線検出方法及びコンピュータプログラム
KR101328741B1 (ko) * 2012-01-03 2013-11-11 중앙대학교 산학협력단 색상 채널을 이용한 영상 개선 장치 및 방법
US9213883B2 (en) * 2012-01-10 2015-12-15 Samsung Electronics Co., Ltd. Method and apparatus for processing depth image
US9990730B2 (en) * 2014-03-21 2018-06-05 Fluke Corporation Visible light image with edge marking for enhancing IR imagery
US10002434B2 (en) * 2014-07-31 2018-06-19 Hewlett-Packard Development Company, L.P. Document region detection
US9674466B2 (en) * 2015-02-26 2017-06-06 Dual Aperture International Co., Ltd. Hybrid image correction for dual-aperture camera
CN105184740A (zh) * 2015-05-20 2015-12-23 华中科技大学 一种红外焦平面图像非均匀性条带校正方法
JP6647956B2 (ja) * 2016-04-28 2020-02-14 任天堂株式会社 情報処理システム、画像処理装置、情報処理プログラムおよび情報処理方法
US10931901B2 (en) * 2016-08-31 2021-02-23 Flir Systems Ab Method and apparatus for selectively correcting fixed pattern noise based on pixel difference values of infrared images

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103679643A (zh) * 2013-06-03 2014-03-26 哈尔滨工程大学 一种多条纹噪声定位滤除方法
CN104580937A (zh) * 2015-01-21 2015-04-29 中国科学院上海技术物理研究所 一种红外成像系统条纹噪声去除方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
XU, HUAYAN ET AL.: "Strip Noise Removal Method for Image Based on Frequency Domain", SILICON VALLEY, 31 March 2012 (2012-03-31), pages 194, ISSN: 1671-7597 *

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109509158A (zh) * 2018-11-19 2019-03-22 电子科技大学 基于幅度约束红外图像条纹去除的方法
CN109509158B (zh) * 2018-11-19 2021-06-11 电子科技大学 基于幅度约束红外图像条纹去除的方法
CN110838093A (zh) * 2019-11-05 2020-02-25 安徽大学 一种高效、高质量的fMOST或MOST显微图像条纹噪声去除方法
CN110838093B (zh) * 2019-11-05 2022-06-10 安徽大学 一种fMOST或MOST显微图像条纹噪声去除方法
CN111383196A (zh) * 2020-03-13 2020-07-07 浙江大华技术股份有限公司 红外图像条纹消除方法、红外探测器及存储装置
CN111383196B (zh) * 2020-03-13 2023-07-28 浙江大华技术股份有限公司 红外图像条纹消除方法、红外探测器及存储装置
CN112017127A (zh) * 2020-08-21 2020-12-01 珀乐(北京)信息科技有限公司 一种基于光谱滤波的图像条带噪声去除方法及系统
CN112862816A (zh) * 2021-03-15 2021-05-28 太原理工大学 一种hrtem图像中煤芳香烃晶格条纹的智能提取方法
CN112862816B (zh) * 2021-03-15 2024-03-15 太原理工大学 一种hrtem图像中煤芳香烃晶格条纹的智能提取方法
CN117830141A (zh) * 2024-03-04 2024-04-05 奥谱天成(成都)信息科技有限公司 红外图像竖条纹噪声去除方法、介质、设备及装置
CN117830141B (zh) * 2024-03-04 2024-05-03 奥谱天成(成都)信息科技有限公司 红外图像竖条纹噪声去除方法、介质、设备及装置

Also Published As

Publication number Publication date
CN106462957B (zh) 2019-03-08
CN106462957A (zh) 2017-02-22
US20180174274A1 (en) 2018-06-21
US10282824B2 (en) 2019-05-07

Similar Documents

Publication Publication Date Title
WO2017197618A1 (zh) 一种红外图像中条纹噪声的去除方法及系统
Chandel et al. Image filtering algorithms and techniques: A review
CN108399611B (zh) 基于梯度正则化的多聚焦图像融合方法
JP2009516172A5 (zh)
US9443286B2 (en) Gray image processing method and apparatus based on wavelet transformation
Shukla et al. Generalized fractional filter-based algorithm for image denoising
Satya et al. Stripe noise removal from remote sensing images
US20150311050A1 (en) Method for Determining a Spectrum from Time-Varying Data
CN108921170B (zh) 一种有效的图像噪声检测和去噪方法及系统
Fan et al. Noise suppression and details enhancement for infrared image via novel prior
KR20160044979A (ko) 영상 처리 방법 및 장치
Gilles et al. Atmospheric turbulence restoration by diffeomorphic image registration and blind deconvolution
US9036889B2 (en) Method and apparatus for stain separation using vector analysis
Campagna et al. Modification of TV-ROF denoising model based on Split Bregman iterations
Gowri et al. 2D Image data approximation using Savitzky Golay filter—Smoothing and differencing
Lv et al. Convex blind image deconvolution with inverse filtering
US10521888B2 (en) Aerothermal radiation effect frequency domain correction method
US11145033B2 (en) Method and device for image correction
JP6584083B2 (ja) 画像処理装置および画像処理プログラム
Abraham Sundar et al. Multi-frame super-resolution using adaptive normalized convolution
Xiao et al. Pixel-level image fusion
Li et al. Super–Resolution Restoration and Image Reconstruction for Passive Millimeter Wave Imaging
Šimberová et al. Analysis of dynamic processes by statistical moments of high orders
KR101775273B1 (ko) 이미지들 사이의 상관관계 획득 방법 및 시스템
Šimberová et al. Statistical moments of active-region images during solar flares

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 16902003

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 29.03.2019)

122 Ep: pct application non-entry in european phase

Ref document number: 16902003

Country of ref document: EP

Kind code of ref document: A1