CN106056588A - Capsule endoscope image data redundancy removing method - Google Patents

Abstract

The invention discloses a capsule endoscope image data redundancy removing method comprising the steps as follows: (1) taking a first image in an image sequence as a key image; (2) taking an image next to the key image as a contrast image, and calculating the global characteristic similarity between the key image and the contrast image; (3) judging whether normalized cross-correlation coefficients are all greater than a threshold, if the normalized cross-correlation coefficients are all greater than the threshold, determining that the global characteristic similarity is greater than a threshold, further calculating the local characteristic similarity between the key image and the contrast image, and going to step (4), or, taking the current contrast image as a new key image and returning to step (2); and (4) judging whether the local characteristic similarity between the key image and the contrast image is greater than a preset threshold, if the local characteristic similarity between the key image and the contrast image is greater than the preset threshold, marking the current contrast image as a redundant image and deleting the contrast image, or, taking the current contrast image as a new key image and returning to step (2). Through the method, the total number of images needing to be checked by doctors can be reduced, and the diagnosis efficiency and accuracy can be improved.

Description

Capsule endoscope view data de-redundancy method

Technical field

The present invention relates to technical field of image processing, in particular to a kind of capsule endoscope view data de-redundancy method.

Background technology

Whole digestive tract can be checked by Wireless capsule endoscope in painless noninvasive mode, be one revolutionary Technological break-through.Patient swallow one seed lac capsule, capsule advances along digestive tract, and captured image data, and capsule endoscope is wanted altogether Working 6～8 hours, gather about 50000 width images, substantial amounts of view data makes the inspection work of doctor become arduous and consumption Time.Medical practitioner typically requires about 2 hours just can complete the inspection work to whole image sequence.The most gastral different Often only exist in a few two field picture, it addition, in the image that capsule endoscope collects, be usually present substantial amounts of redundancy, doctor Raw probably due to for a long time browsing pictures inspection visual fatigue there is and mistaken diagnosis.

The Chinese patent of Publication No. CN 103747270 A " a kind of de-redundancy method of capsule endoscope video image and System ", this technology first defines a key images frame, is calculated one by one based on gray scale with picture frame behind by key images frame The cross-correlation coefficient of image, by judging more than threshold value, whether cross-correlation coefficient judges whether image is redundant frame.This technology Judge the similarity of image by calculating cross-correlation coefficient based on gray level image, and actual capsule collects is cromogram Picture, a kind of common disease (digestive tract hemorrhage) is mainly judged by colouring information, ignores colouring information and be easily caused image Delete by mistake.Additionally what cross-correlation coefficient characterized is the global characteristics of image, and this technology only make use of the global characteristics of image, do not examines Considering the local feature of image, the size sometimes of the focus in image can be smaller, and what undersized focus caused is local between image The difference of feature, therefore judges image whether redundancy may result in delete only by global characteristics by mistake.

Paper " capsule endoscope redundant image data Automatic sieve removes method " list of references (periodical): Sun Yuqian, Lv Qingwen, Liu Zhe magnitude. capsule endoscope redundant image data Automatic sieve removes method [J]. and computer utility is studied, and 2012,29 (6): 2393-2396, this technology first by image at HSV (Hue tone, Saturation saturation, Value lightness) space quantization, To HSV rectangular histogram, it is then based on normalized mutual information and calculates the similarity of adjacent image, finally according to screening ratio to figure As data screen out.This technology only considered the global characteristics of image, does not accounts for the difference of image local feature.Still deposit In the possibility deleted by mistake.

Paper " research of capsule endoscope image sequence redundant data screening method " list of references (thesis for the doctorate): penning. glue Capsule endoscopic image sequence redundancy data screening method research [D]. the Central China University of Science and Technology, 2013, in propose based on camera motion The redundant data screening method estimated, the motion of capsule is estimated by the method by image registration, by motion size Judge view data whether redundancy.This technology uses image registration techniques, it is contemplated that global motion model and local motion mould Type, computationally intensive, it is unfavorable for that the capsule endoscope applying to reality assists in diagosis system.

Summary of the invention

The present invention is aiming at above-mentioned technical problem, it is provided that a kind of capsule endoscope view data de-redundancy method, the party Method reduces the amount of images that doctor needs to browse, thus improves diagnosis efficiency and accuracy rate.

For achieving the above object, a kind of capsule endoscope view data de-redundancy method designed by the present invention, it includes Following steps:

Step 1: after patient swallow's capsule endoscope, capsule endoscope gathers view data in digestive tract, collects The total M width image according to shooting time arrangement of image sequence in view data, by image sequence first shooting moment Corresponding image, as key images, enters step 2；

Step 2: take the image image as a comparison of the adjacent subsequent time of described key images, calculates key images with right Ratio global characteristics similarity between image, extract the most respectively key images three components of R, G, B (red, green, blue is trichroism) and Tri-components of R, G, B of contrast images, calculate tri-components of R, G, B of key images and R, G, B tri-of contrast images respectively The normalized-cross-correlation function NCC of component_R、NCC_GAnd NCC_B；

Step 3: judge the normalized-cross-correlation function NCC that step 2 obtains_R、NCC_GAnd NCC_BThe most all more than threshold value T₁, the most then the similarity S of global characteristics_gMore than threshold value T₁, now, calculate further between key images and contrast images The similarity S of local feature_g, enter step 4；Otherwise, the similarity S of global characteristics_gIt is not more than threshold value T₁, by currently take out Contrast images is labeled as nonredundancy image, and using the contrast images currently taken out as new key images, returns step 2；

Step 4: judge the similarity S of local feature between key images and contrast images_gWhether more than the threshold preset Value T₂, the contrast images the most currently taken out is labeled as redundant image, and this redundant image is deleted；Otherwise, will be current The contrast images taken out is labeled as nonredundancy image, and the contrast images currently taken out is walked as new key images, return Rapid 2.

Present invention have the advantage that

1, by the removable substantial amounts of similar pictures of the method, so doctor is made to inspect picture lighter；For exception Picture, the most gastral exception only exists in a few two field picture, and doctor is probably due to browse a large amount of picture for a long time and occur regarding Feel tired and ignore these a few frame abnormal images and cause mistaken diagnosis, reduce the picture sum needing to check, diagnosis effect can not only be improved Rate, the most also can improve the accuracy rate of diagnosis.Therefore, the present invention can reduce the picture sum that doctor needs to check, Improve diagnosis efficiency and accuracy rate.

2, the present invention has considered global characteristics similarity and the similarity of local feature of image, reduces lesion image The probability by mistake deleted.

3, the method mentioned in paper " research of capsule endoscope image sequence redundant data screening method " passes through image registration Estimate the motion of capsule, get rid of redundant image according to the size of motion.The method that the present invention proposes is by calculating image Global and local feature, gets rid of redundant image by the similarity size of feature.The feature extraction of image and characteristic similarity Judgement on time loss much smaller than image registration techniques, the present invention does not use image registration, method efficiency to compare employing The method of image registration techniques is higher.

Accompanying drawing explanation

Fig. 1 is the flow chart of the present invention；

Fig. 2 is calculating and the decision flow chart of global characteristics similarity in view data de-redundancy method of the present invention；

Fig. 3 is the calculation flow chart of local feature similarity in view data de-redundancy method of the present invention.

Detailed description of the invention

The present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.

Capsule endoscope view data de-redundancy method as described in Fig. 1～3, it comprises the steps:

Step 1: after patient swallow's capsule endoscope, capsule endoscope gathers view data in digestive tract, collects Total M width image (coloured image) according to shooting time arrangement of image sequence in view data, by image sequence first Individual shooting image corresponding to moment, as key images, enters step 2；

Step 2: take the image image as a comparison of the adjacent subsequent time of described key images, calculates key images with right Ratio global characteristics similarity between image, extracts tri-components of R, G, B and R, G, B of contrast images of key images the most respectively Three components, tri-components of R, G, B calculating key images respectively are mutual with the normalization of tri-components of R, G, B of contrast images Close coefficient NCC_R、NCC_GAnd NCC_B；

Step 3: judge the normalized-cross-correlation function NCC that step 2 obtains_R、NCC_GAnd NCC_BThe most all more than threshold value T₁, the most then the similarity S of global characteristics_gMore than threshold value T₁, now, calculate further between key images and contrast images The similarity S of local feature_l, enter step 4；Otherwise (normalized-cross-correlation function NCC_R、NCC_GAnd NCC_BOne is had to be less than In threshold value T₁), the similarity S of global characteristics_gIt is not more than threshold value T₁, the contrast images currently taken out is labeled as nonredundancy image, And using the contrast images currently taken out as new key images, return step 2；

Step 4: judge the similarity S of local feature between key images and contrast images_lWhether more than the threshold preset Value T₂, the contrast images the most currently taken out is labeled as redundant image, and this redundant image is deleted；Otherwise, will be current The contrast images taken out is labeled as nonredundancy image, and the contrast images currently taken out is walked as new key images, return Rapid 2；

Step 5: repeat step 2～step 4, until all of image all judges complete in sequence, now, remains all Key images, eliminates all redundant images, and the de-redundancy i.e. achieving capsule endoscope view data processes.

In the step 1 of technique scheme, key images is first in described image sequence according to shooting time arrangement Width image.Then the contrast images in step 2 is the second width image according to shooting time arrangement.

In technique scheme, described threshold value T₁Span be 0.6≤T₁≤0.85；Described threshold value T₂Value model Enclosing is 0.5≤T₂≤0.9.Judge according to repetition test, above-mentioned threshold value T₁With threshold value T₂Span can obtain optimal going Redundancy effect.

In the step 2 of technique scheme, calculate tri-components of R, G, B of key images and R, G, B tri-of contrast images The normalized-cross-correlation function NCC of individual component_R、NCC_GAnd NCC_BSpecific formula for calculation as follows:

${\mathrm{NCC}}_n=\frac{\underset{x,y}{\mathrm{\Σ}}\[I_n(x,y)-{\stackrel{\‾}{I}}_n\]\[J_n(x,y)-{\stackrel{\‾}{J}}_n\]}{\sqrt{\underset{x,y}{\mathrm{\Σ}}{\[I_n(x,y)-{\stackrel{\‾}{I}}_n\]}^2\·\underset{x,y}{\mathrm{\Σ}}{\[J_n(x,y)-{\stackrel{\‾}{J}}_n\]}^2}}$

Wherein, n is R, G, B, represents Color Channel；X, y are respectively abscissa and the vertical coordinate of image；I_n(x, y) for closing The view data of the single Color Channel that key image is corresponding, J_n(x y) is the image of single Color Channel corresponding to contrast images Data,WithIt is respectively image I_n(x, y) and J_n(x, gray average y).

In technique scheme, the similarity S of the local feature between described key images and contrast images_lCalculating side Method is to calculate key images and histograms of oriented gradients (HOG, the histogram of Oriented of contrast images Gradient) (see reference characteristic vector document, Triggs N D B.Histograms of Oriented Gradients For Human Detection [J] .Cvpr, 2005,1 (12): 886-893.), by the histograms of oriented gradients of key images The normalized-cross-correlation function value of the characteristic vector of the histograms of oriented gradients of characteristic vector and contrast images is as local feature Similarity function value S_l, above-mentioned similarity function value S_lComputing formula be:

$S_l=\frac{\underset{z}{\mathrm{\Σ}}\[V_1\left(z\right)-\stackrel{\‾}{V_1}\]\[V_2\left(z\right)-\stackrel{\‾}{V_2}\]}{\sqrt{\underset{z}{\mathrm{\Σ}}{\[V_1\left(z\right)-\stackrel{\‾}{V_1}\]}^2\·\underset{z}{\mathrm{\Σ}}{\[V_2\left(z\right)-\stackrel{\‾}{V_2}\]}^2}}$

Wherein, V₁(z) and V₂Z () is respectively the characteristic vector of the histograms of oriented gradients of key images and contrast images, z Being characterized the coordinate of vector, the value of z is [1, m], and m is characterized the dimension of vector,WithIt is respectively V₁(z) and V₂(z) equal Value.

The content that this specification is not described in detail belongs to prior art known to professional and technical personnel in the field.

Claims (7)

1. a capsule endoscope view data de-redundancy method, it is characterised in that it comprises the steps:

Step 1: after patient swallow's capsule endoscope, capsule endoscope gathers view data, the image collected in digestive tract The total M width image according to shooting time arrangement of image sequence in data, by corresponding in image sequence first shooting moment Image as key images, enter step 2；

Step 2: take the image image as a comparison of the adjacent subsequent time of described key images, calculates key images and comparison diagram Global characteristics similarity between Xiang, extracts tri-components of R, G, B and R, G, B tri-of contrast images of key images the most respectively Component, calculates the normalized crosscorrelation system of tri-components of R, G, B of key images and tri-components of R, G, B of contrast images respectively Number NCC_R、NCC_GAnd NCC_B；

Step 3: judge the normalized-cross-correlation function NCC that step 2 obtains_R、NCC_GAnd NCC_BThe most all more than threshold value T₁If, It is, then the similarity S of global characteristics_gMore than threshold value T₁, now, calculate the local between key images and contrast images further The similarity S of feature_l, enter step 4；Otherwise, the similarity S of global characteristics_gIt is not more than threshold value T₁, the contrast that will currently take out Image tagged is nonredundancy image, and using the contrast images currently taken out as new key images, returns step 2；

Step 4: judge the similarity S of local feature between key images and contrast images_lWhether more than threshold value T preset₂, The contrast images the most currently taken out is labeled as redundant image, and this redundant image is deleted；Otherwise, will currently take out Contrast images be labeled as nonredundancy image, and using the contrast images currently taken out as new key images, return step 2.

Capsule endoscope view data de-redundancy method the most according to claim 1, it is characterised in that: after described step 4 Also include step 5: repeat step 2～step 4, until all of image all judges complete in sequence, now, remain institute relevant Key image, eliminates all redundant images.

Capsule endoscope view data de-redundancy method the most according to claim 1, it is characterised in that: described threshold value T₁'s Span is 0.6≤T₁≤0.85。

Capsule endoscope view data de-redundancy method the most according to claim 1, it is characterised in that: described threshold value T₂'s Span is 0.5≤T₂≤0.9。

Capsule endoscope view data de-redundancy method the most according to claim 1, it is characterised in that: in described step 2, Calculate the normalized-cross-correlation function NCC of tri-components of R, G, B of key images and tri-components of R, G, B of contrast images_R、 NCC_GAnd NCC_BSpecific formula for calculation as follows:

${\mathrm{NCC}}_n=\frac{\underset{x,y}{\Σ}\[I_n(x,y)-{\stackrel{\‾}{I}}_n\]\[J_n(x,y)-{\stackrel{\‾}{J}}_n\]}{\sqrt{\underset{x,y}{\Σ}{\[I_n(x,y)-{\stackrel{\‾}{I}}_n\]}^2\·\underset{x,y}{\Σ}{\[J_n(x,y)-{\stackrel{\‾}{J}}_n\]}^2}}$

Wherein, n is R, G, B, represents Color Channel；X, y are respectively abscissa and the vertical coordinate of image；I_n(x y) is key images The view data of corresponding single Color Channel, J_n(x, y) is the view data of single Color Channel corresponding to contrast images, WithIt is respectively image I_n(x, y) and J_n(x, gray average y).

Capsule endoscope view data de-redundancy method the most according to claim 1, it is characterised in that: described key images And the similarity S of the local feature between contrast images_lComputational methods be to calculate the direction ladder of key images and contrast images Spend histogrammic characteristic vector, by straight to the characteristic vector of the histograms of oriented gradients of key images and the direction gradient of contrast images The cross correlation numerical value of the characteristic vector of side's figure is as similarity function value S of local feature_l。

Capsule endoscope view data de-redundancy method the most according to claim 6, it is characterised in that: above-mentioned similarity letter Numerical value S_lComputing formula be:

$S_l=\frac{\underset{Z}{\Σ}\[V_1\left(Z\right)-{\stackrel{\‾}{V}}_1\]\[V_2\left(Z\right)-{\stackrel{\‾}{V}}_2\]}{\sqrt{\underset{z}{\Σ}{\[V_1\left(Z\right)-{\stackrel{\‾}{V}}_1\]}^2\·\underset{z}{\Σ}{\[V_2\left(z\right)-{\stackrel{\‾}{V}}_2\]}^2}}$

Wherein, V₁(z) and V₂Z () is respectively the characteristic vector of the histograms of oriented gradients of key images and contrast images, z is special Levying the coordinate of vector, the value of z is [1, m], and m is characterized the dimension of vector,WithIt is respectively V₁(z) and V₂The average of (z).