CN106846354A - A kind of Book Inventory method on frame converted based on image segmentation and random hough - Google Patents

Abstract

The invention discloses a kind of Book Inventory method on frame converted based on image segmentation and random hough, belong to library management technical field.First, frame information is extracted according to image edge pixels accumulation, so as to divide the image into multiple subgraphs, each subgraph represents a spine image put in lattice on bookshelf；Then, by Rapid Accumulation of the books boundary straight line in parameter space in random hough Mapping implementations subgraph；Finally, local peaking is detected in parameter space, so as to realize that each counting for putting books volumes in lattice is checked on bookshelf, and the volumes of all books on bookshelf is finally given；The present invention preferably combines the advantage of image procossing and random hough conversion, and the auto inventory of books on frame is realized using computer disposal；Method is simply efficient, it is easy to implement, and greatly reduces the workload of Book Inventory, with stronger practicality and application value.

Description

A kind of Book Inventory method on frame converted based on image segmentation and random hough

Technical field

The invention belongs to library management technical field, and in particular to one kind is converted based on image segmentation and random hough Frame on Book Inventory method.

Background technology

Book Inventory is the basic data in books statistical indicator, is a difficulties of library management.By clear Point books can accurately determine the existing books property in library, for objectively reflect school run a school strength, strive for more Input has important reference value.Therefore, Book Inventory Work is very necessary for the daily management in library, but Traditional artificial mode of checking not only has taken but also laborious, have that labour demand is big, information updating not in time the shortcomings of.For books It is an important research direction for the management of shop, traditional Book Inventory Work to be digitized, automate improvement, in recent years Come, image processing techniques has obtained fruitful application in terms of taking care of books, for Book Inventory provides a kind of new side Method.

[Liu your emerald green is based on the automatic Book Inventory method of spine characteristics of image and airspace filter to background technology:China, ZL201410383904.7,2016.05.04], spine image is processed using the method for characteristics of image and airspace filter, The auto inventory of single books is realized, improve the efficiency of Book Inventory, the method is mainly comprised the following steps：

1) the spine image in collection stack room, it is ensured that row's books are comprised only in image；

2) the orientation inclination angle of books is obtained, edge pixel is projected, obtain spine word projection oscillogram；

3) edge projection oscillogram is filtered, obtains smooth edge projection oscillogram；

4) number of edge projection waveform medium wave peak is detected, the volumes comprising books as in image.

The method that the above method employs image procossing realizes checking for single books, improves operating efficiency；But its Mainly for the image procossing of simple single books, i.e., the books that can only be put to single books on bookshelf in lattice are carried out clearly Point, and the collected books quantity in library is very huge now, has multiple books to put lattice on each bookshelf, the above method is difficult to efficiently Check these magnanimity collected books；On the other hand, the above method calculates the accumulation of spine word using the method that edge projection is detected Peak value, the method is higher to the requirement of spine background situation；When word is less on books spine or distinguishes little with background, hold It is easily caused books missing inspection；And it is more to work as word, when being in two column distributions on spine, books false is easily caused.Above mentioned problem Occur, limit the practicality and range of application of background technology.

The content of the invention

1. the technical problem to be solved

The purpose of the present invention is to propose to a kind of Book Inventory method on frame converted based on image segmentation and random hough, So as to overcome the deficiencies in the prior art, the efficiency of Book Inventory is greatly improved.

2. technical scheme

The invention provides a kind of Book Inventory method on frame converted based on image segmentation and random hough, using skill Art protocol step is as follows：

Step 1：Image preprocessing

Coloured image of the collection comprising bookshelf framework and spine, is modified to picture, obtains the consistent figure of horizontal direction Piece, and ensure that bookshelf framework takes entire image, the image is designated as P；With the image lower left corner as origin, x-y rectangular co-ordinates are set up System, wherein coordinate system x-axis direction overlaps with image lower boundary, and coordinate system y-axis direction overlaps with left picture boundary；I and j is made to distinguish It is image in the number of pixel cells in coordinate system x and y directions, then image P can be expressed as the form of picture element matrixWherein p (i, j) is the pixel value at pixel cell (i, j) place, and I is figure P in the total pixel in x directions Number, J is the total number of pixels in y directions；

Step 2：Image segmentation

(1) coloured image P is converted into 256 grades of gray-scale map F, the gray-scale map can be represented in x-y rectangular coordinate systems It is imaged the form of prime matrixWherein f (i, j) is the pixel value at pixel cell (i, j) place；For any Pixel cell f (i, j), the gradient f that differential process calculate x directions is carried out by neighborhood unit pixel_xWith the gradient f in y directions_y

Then gradient magnitude of the image at coordinate unit (i, j) place isGradient direction is α (i, j) =arctan (f_y/f_x), wherein arctan () represents tan of negating；

(2) local extremum at edge is extracted by non-maxima suppression, so as to reject the influence of part non-edge point；It is first First, according to the value of α (i, j), it is four sectors interval that the gradient direction at pixel cell f (i, j) place is evenly dividing, respectively generation The gradient of table horizontal direction, left-leaning direction, Right deviation direction and vertical direction：

Wherein, s (i, j) is the sector belonging to coordinate unit (i, j) place gradient direction, and π is the corresponding radian number in 180 ° of angles； Then according to s (i, j) value, two 8 neighborhood coordinate units of the same gradient direction in coordinate unit (i, j) place are searched

Wherein,WithRespectively number of pixel cells of the two 8 neighborhood coordinate units of (i, j) in coordinate system x and y directions； Finally, by the neighborhood territory pixel of the gradient magnitude at coordinate unit (i, j) place and same gradient directionCompare, extract gradient magnitude Local maximum：

(3) binary conversion treatment is carried out to gradient magnitude using dual threshold thresholding, if Low threshold thresholding is σ_low, high threshold door It is limited to σ_high, rim detection is carried out to image and obtains binary imageWherein b (i, j) is pixel cell The binaryzation pixel value at (i, j) place：

For gradient magnitude between σ_lowAnd σ_highBetween coordinate unit (i, j), the value of b (i, j) is according to 8 neighborhood pictures The gradient of plain unit determines

Wherein, M_maxIt is the maximum of the neighborhood territory pixel unit gradient of coordinate unit (i, j) place 8；

(4) edge of each lattice bookshelf framework is straight line both horizontally and vertically, and the edge of books is then deposited in framework In more diagonal oblique line, therefore treatment is filtered to binary image B using a diagonal matrix, obtains new pixel Value：

B (i, j)=max [b (i, j), b (i-1, j-1), b (i+1, j+1), b (i-1, j+1), b (i+1, j-1)], wherein, Max [] is represented and is taken maximum computing；Then bianry image is obtained into image in level respectively to both horizontally and vertically projecting The accumulation parameter Q in direction_xThe accumulation parameter Q of (i) and vertical direction_y(j)：

According to the horizontal detection threshold γ of accumulation parameter_xWith vertical detection thresholding γ_y, extract Q_xThe corresponding sequence of (i) valley Number it is stored in array D_x, extract Q_yJ the corresponding sequence number of () valley is stored in array D_y：

Wherein, arg [] represents the sequence number for meeting constraints [], then D_xData correspondence bookshelf framework is in level side To coordinate, D_yCoordinate of the data correspondence bookshelf framework in vertical direction；

(5) original image is next divided into multiple subgraphs using the information of framework, in each subgraph correspondence bookshelf One rectangle puts the spine image in lattice；First, the position coordinates D of bookshelf framework is extracted_xAnd D_y, so as to realize in bookshelf Each rectangle puts the positioning of lattice；Then, the border that lattice are put according to rectangle is split to coloured image P, obtains multiple subgraphs Picture, realizes the dividing processing of original image：

Wherein N is the total number of subgraph, P_nRepresent n-th subgraph.

Step 3：Stochastical sampling

(1) for n-th subgraph P_n, remember I_nIt is it in the total number of pixels in x directions, J_nIt is it in the total pixel in y directions Number；Pre-processed, obtained the gradient direction α at pixel cell (i, j) place_n(i,j)；Then by α_nIt is four that (i, j) is evenly dividing Individual sector is interval；Use s_n(i, j) represents subgraph P_nIn the sector belonging to coordinate unit (i, j) place gradient direction；Carry out edge inspection Measure binary imageWherein b_n(i, j) is subgraph P_nIn the two-value at pixel cell (i, j) place Change pixel value；

(2) in binary image B_nIn randomly select a pixel cell (i for nonzero element₁,j₁) so that 0.3J_n≤j₁ ≤0.7J_nAnd b_n(i₁,j₁) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₁,j₁)；

(3) in binary image B_nSecond pixel cell (i of nonzero element of middle selection₂,j₂)≠(i₁,j₁) so that 0.3J_n≤j₂≤0.7J_nAnd b_n(i₂,j₂) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₂,j₂)；If s_n(i₁, j₁)≠s_n(i₂,j₂), resampling, until s_n(i₁,j₁)=s_n(i₂,j₂), by two pixel cell (i₁,j₁) and (i₂,j₂) group Into a stochastical sampling pixel pair；；

(4) repeated sampling, obtains L stochastical sampling pixel pair.

Step 4：Parameter is accumulated

Accumulative element distance is calculated, hough parameter spaces (ρ, θ) are set up, ρ and θ is respectively distance and angle in space Parameter；By two pixel cell (i of stochastical sampling₁,j₁) and (i₂,j₂) parameter space is projected to by hough conversion, calculate Corresponding parameters unit (ρ_hough,θ_hough), ρ_houghAnd θ_houghDistance and angle after respectively hough conversion in parameter space Parameter；Ballot accumulation is carried out in corresponding accumulative element, the collinear sampled point of correspondence is fallen into identical accumulation single Unit, obtains parameter and accumulates matrix H (ρ, θ), and specific method is as follows：

(1) sliding-model control is carried out to parameter ρ and θ according to discrete interval Δ ρ and Δ θ, forms parameter space (ρ, θ), The central point parameter of (u, v) individual parameters unit is ρ_uAnd θ_v：

ρ_u=(u-1/2) Δ ρ, u=1,2 ..., U,

θ_v=(v-1/2) Δs θ-pi/2, v=1,2 ..., V,

Wherein, u and v are respectively the sequence number of parameter ρ and θ, and the parameters unit number V of parameters unit the number U and θ of ρ takes Value is respectively：V=Int (π/Δ θ), Int () represent rounding operation, and π is that 180 ° of angles are corresponding Radian number；Set up parameter and accumulate matrix H (ρ, θ), it is 0 to put each unit initial value；

(2) by first pixel of stochastical sampling to (i₁,j₁) and (i₂,j₂) parameter sky is projected to by hough conversion Between, the pixel is calculated to identified parameter ρ_houghAnd θ_hough：

θ_hough=arctan [- (i₁-i₂)/(j₁-j₂)]；

ρ_hough=0.5 (i₁+i₂)cos(θ_hough)+0.5(j₁+j₂)sin(θ_hough)；

(3) corresponding unit (ρ is searched in parameter space (ρ, θ)_u,θ_v) so that | | ρ_hough-ρ_u||≤Δρ,|| θ_hough-θ_v||≤Δθ；If (ρ_u,θ_v) exist, then in the memory cell (ρ of parameter accumulation matrix H (ρ, θ)_u,θ_v) in tired out Plus：H(ρ_u,θ_v)=H (ρ_u,θ_v)+1；Otherwise, by (ρ_hough,θ_hough) insertion parameter space (ρ, θ), and take initial value H (ρ_hough, θ_hough)=1；

(4) repeat to accumulate, until L stochastical sampling pixel is to being mapped to parameter space, obtain parameter accumulation matrix H (ρ,θ)。

Step 5：Peakvalue's checking

(1) according to detection threshold G₁, matrix H (ρ, θ) accumulating to parameter and is detected, value is less than G in making H (ρ, θ)₁'s Element value is zero, and detection threshold is set to G₁=0.3J_n；Consider books inclination angle, angle restriction value θ is set_T, make θ in H (ρ, θ) Absolute value is less than θ_TElement value be zero；

(2) parameters unit (ρ that any one is not zero in H (ρ, θ)_u,θ_v), define its local search area Ω (ρ_u, θ_v) be：Ω(ρ_u,θ_v)={ (ρ, θ) | [| | ρ-ρ_u| |≤5, θ=θ_v,H(ρ,θ)≠0]}；

(3) next according to Ω (ρ_u,θ_v) judged：If Ω (ρ_u,θ_v) it is sky, this parameters unit (ρ_u,θ_v) it is one Individual local peaking；If Ω (ρ_u,θ_v) it is not sky, if H (ρ_u,θ_v) value more than or equal to region Ω (ρ_u,θ_v) in all units Value：H(ρ_u,θ_v) >=H (ρ, θ), (ρ, θ) ∈ Ω (ρ_u,θ_v), then parameters unit (ρ_u,θ_v) it is a local peaking；

(4) H (ρ, θ) all of local peaking's unit is extracted, the number t of local peaking is calculated_n, then subgraph P_nMiddle books Volumes be t_n-1；

Step 6：Counting is checked

When the value of n changes from 1 to N, for all of subgraph P_n, the process of repeat step 3 to step 5, detection Obtain the peak value number t in all subgraph parameter accumulation matrixes_n, n=1,2 ..., N, then on whole bookshelf books volumes T For:

3. beneficial effect

Compared with background technology, beneficial effects of the present invention explanation：

(1) Book Inventory method on a kind of frame converted based on image segmentation and random hough that the present invention is used, first Multiple subgraphs are divided the image into by image segmentation, Book Inventory then is carried out to each subgraph, magnanimity can be solved Books check problem on frame, with good actual application value；

(2) being converted by random hough carries out Rapid Accumulation to books borderline region, and detection determines the number of border peak value Amount, can reduce the amount of calculation for the treatment of, the accuracy that raising is checked.

Brief description of the drawings

Accompanying drawing 1 is overall flow figure of the invention；

Accompanying drawing 2 is book image on frame to be processed in the embodiment of the present invention；

Accompanying drawing 3 is the image after binary conversion treatment in the embodiment of the present invention；

Accompanying drawing 4 is the image after filtering process in the embodiment of the present invention；

Accompanying drawing 5 is the gentle vertical direction pixel accumulation result figure of embodiment of the present invention reclaimed water；

Accompanying drawing 6 is the subgraph obtained after image segmentation in the embodiment of the present invention；

Accompanying drawing 7 is first sub- Image Edge-Detection result in the embodiment of the present invention；

Accompanying drawing 8 is first sub- image parameter accumulation result in the embodiment of the present invention；

Accompanying drawing 9 is the parameter accumulation figure after Threshold detection and angle restriction in the embodiment of the present invention；

Accompanying drawing 10 is local peak detection result in first subgraph in the embodiment of the present invention；

Specific embodiment

Main adopting of the invention is experimentally verified that all steps, conclusion are verified all on Matlab2010a. Below in conjunction with Figure of description 1 to Book Inventory side on a kind of frame converted based on image segmentation and random hough of the invention Method is described in detail.As shown in Figure 1, processing procedure of the present invention is segmented into 3 modules：M1 is image segmentation module, main Bookshelf framework is extracted according to image information, so as to be split according to the lattice of putting of bookshelf, divide the image into multiple subgraphs； M2 is random hough conversion process module, mainly by edge pixel in hough Mapping implementation subgraphs in parameter space Accumulation；M3 is peak detection block, and the main detection by parameter space local peaking realizes that the counting of books volumes is checked.

Specific handling process of the invention is as follows：

Step 1：Image preprocessing

Coloured image of the collection comprising bookshelf framework and spine, is modified to picture, obtains the consistent figure of horizontal direction Piece, and ensure that bookshelf framework takes entire image, the image is designated as P, as shown in Figure 2；With the image lower left corner as origin, x- is set up Y rectangular coordinate systems, wherein coordinate system x-axis direction overlap with image lower boundary, and coordinate system y-axis direction overlaps with left picture boundary； I and j respectively image are made in the number of pixel cells in coordinate system x and y directions, then image P can be expressed as the form of picture element matrixWherein p (i, j) is the pixel value at pixel cell (i, j) place, and I is figure P in the total pixel in x directions Number, J is the total number of pixels in y directions；

Step 2：Image segmentation

(1) coloured image P is converted into 256 grades of gray-scale map F, the gray-scale map can be represented in x-y rectangular coordinate systems It is imaged the form of prime matrixWherein f (i, j) is the pixel value at pixel cell (i, j) place；For any Pixel cell f (i, j), the gradient f that differential process calculate x directions is carried out by neighborhood unit pixel_xWith the gradient f in y directions_y, then Gradient direction of the image at coordinate unit (i, j) place is α (i, j)=arctan (f_y/f_x), gradient magnitude is

(2) local extremum at edge is extracted by non-maxima suppression, so as to reject the influence of part non-edge point；It is first First, according to the value of α (i, j), it is four sectors interval s (i, j) that the gradient direction at pixel cell f (i, j) place is evenly dividing, The gradient of horizontal direction, left-leaning direction, Right deviation direction and vertical direction is represented respectively；Then according to s (i, j) value, search and sit Mark two 8 neighborhood coordinate units of the same gradient direction in unit (i, j) placeFinally, by the gradient at coordinate unit (i, j) place The neighborhood territory pixel of amplitude and same gradient directionCompare, extract the local maximum of gradient magnitude；

(3) binary conversion treatment is carried out to gradient magnitude using dual threshold thresholding, if Low threshold thresholding is σ_low, high threshold door It is limited to σ_high, rim detection is carried out to image and obtains binary imageWherein b (i, j) is pixel cell The binaryzation pixel value at (i, j) place；For gradient magnitude between σ_lowAnd σ_highBetween coordinate unit (i, j), b's (i, j) takes Value determines according to the gradient of 8 neighborhood territory pixel units；The binary image obtained after treatment is as shown in Figure 3, it is seen that bookshelf framework picture There is nuance between element and books pixel, but discrimination is little；

(4) edge of each lattice bookshelf framework is straight line both horizontally and vertically, and the edge of books is then deposited in framework In more diagonal oblique line, therefore treatment is filtered to binary image B using a diagonal matrix, obtains new pixel Value；The image obtained after filtering is as shown in Figure 4, it is seen that after processing after filtering, effectively reduces the influence of background, increased frame Otherness between frame pixel and books pixel；

(5) by bianry image respectively to both horizontally and vertically projecting, image accumulation pixel Q in the horizontal direction is obtained_x The accumulation pixel Q of (i) and vertical direction_y(j), as shown in figure 5, pixel accumulation figures of the wherein Fig. 5 (a) for horizontal direction, Fig. 5 B () is the pixel accumulation figure of vertical direction；Comparison diagram 2 and Fig. 5 can be seen that the valley of pixel accumulation in Fig. 5 just corresponding diagram 2 The frame position of middle bookshelf；

(6) according to the horizontal detection threshold γ of accumulation parameter_xWith vertical detection thresholding γ_y, extract Q_xI () valley is corresponding Sequence number is stored in array D_x, extract Q_yJ the corresponding sequence number of () valley is stored in array, then D_xData correspondence bookshelf framework is in level The coordinate in direction, D_yCoordinate of the data correspondence bookshelf framework in vertical direction；Original image is divided into using the information of framework many Individual subgraph, rectangle puts the spine image in lattice in each subgraph correspondence bookshelf, realizes the segmentation portion of original image Reason, obtains image segmentation result as shown in Figure 6；As can be seen that by after image segmentation, original image puts lattice according to bookshelf Position has accurately been divided into 4 subgraphs, and segmentation effect is preferable；

Step 3：Stochastical sampling

(1) for the n-th=1 subgraph, pre-processed, obtained the gradient direction α at pixel cell (i, j) place_n(i, j)；Then by α_nIt is four sectors interval that (i, j) is evenly dividing；Use s_n(i, j) represents subgraph P_nAt coordinate unit (i, j) place Sector belonging to gradient direction；Carry out rim detection and obtain binary imageAs shown in Figure 7；Can be with Find out, after treatment, the edge of books is more obvious；

(2) in binary image B_nIn randomly select a pixel cell (i for nonzero element₁,j₁) so that 0.3J_n≤j₁ ≤0.7J_nAnd b_n(i₁,j₁) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₁,j₁)；

(3) in binary image B_nSecond pixel cell (i of nonzero element of middle selection₂,j₂)≠(i₁,j₁) so that 0.3J_n≤j₂≤0.7J_nAnd b_n(i₂,j₂) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₂,j₂)；If s_n(i₁, j₁)≠s_n(i₂,j₂), resampling, until s_n(i₁,j₁)=s_n(i₂,j₂), by two pixel cell (i₁,j₁) and (i₂,j₂) group Into a stochastical sampling pixel pair；

(4) repeated sampling, obtains L stochastical sampling pixel pair；

Step 4：Parameter is accumulated

Accumulative element distance is calculated, hough parameter spaces (ρ, θ) are set up, ρ and θ is respectively distance and angle in space Parameter；By two pixel cell (i of stochastical sampling₁,j₁) and (i₂,j₂) parameter space is projected to by random hough conversion, Calculate corresponding parameters unit (ρ_hough,θ_hough)；Ballot accumulation is carried out in corresponding accumulative element, makes correspondence collinear Sampled point can fall into identical accumulative element, obtain parameter and accumulate matrix H (ρ, θ), as shown in figure 8, wherein Fig. 8 (a) is three The three-dimensional histogram of dimension, Fig. 8 (b) is H (ρ, θ) to the two-dimentional display figure obtained after ρ-θ plane projections；It can be seen that by Accumulated in random hough conversion is employed, the parameter accumulation matrix for obtaining is sparse, and the method need to only carry out limited adopting The mapping of sample pixel can be achieved with the sparse accumulation on books border, therefore its amount of calculation is lower than traditional hough conversion；

Step 5：Peakvalue's checking

(1) according to detection threshold G₁=0.3J_n, matrix H (ρ, θ) is accumulated to parameter and is detected that value is small in making H (ρ, θ) In G₁Element value be zero；Consider books inclination angle, angle restriction value θ is set_T=10 °, θ absolute values are less than θ in making H (ρ, θ)_T Element value be zero, obtain result as shown in Figure 9；Understood with Fig. 8 (b) contrasts, after over-threshold detection and angle restriction, Fig. 9 In eliminate a large amount of unrelated accumulative elements；

(2) parameters unit (ρ that any one is not zero in H (ρ, θ)_u,θ_v), define its local search area Ω (ρ_u, θ_v) be：Ω(ρ_u,θ_v)={ (ρ, θ) | [| | ρ-ρ_u| |≤5, θ=θ_v,H(ρ,θ)≠0]}；

(3) next according to Ω (ρ_u,θ_v) judged：If Ω (ρ_u,θ_v) it is sky, parameters unit (ρ_u,θ_v) it is one Local peaking；If Ω (ρ_u,θ_v) it is not sky, if H (ρ_u,θ_v) value more than or equal to region Ω (ρ_u,θ_v) in all units Value：H(ρ_u,θ_v) >=H (ρ, θ), (ρ, θ) ∈ Ω (ρ_u,θ_v), then parameters unit (ρ_u,θ_v) it is a local peaking；

(4) H (ρ, θ) all of local peaking's unit is extracted, as shown in Figure 10, wherein white rectangle represents local peaking Position, carrying out counting to number of peaks can obtain t₁=48, then subgraph P₁The volumes of middle books is t₁- 1=47, this and the subgraph The volumes of books is consistent as in；

Step 6：Counting is checked

When the value of n changes from 1 to N, for all of subgraph P_n, the process of repeat step 3 to step 5, detection The peak value number in all subgraph parameter accumulation matrixes is obtained, as shown in table 1：

The Book Inventory result of table 1

Image category	Actual volumes
					47	47	47
	37	37	37
					44	44	44
	38	35	38
				General image P	166	163	166

For the ease of comparing, table 1 is in angle restriction value θ_TValue be respectively 5 °, verified in 10 ° of environment, can be with Find out θ_TValue it is larger to the performance impact of method；For example, in subgraph P₄In, because the inclination angle of part books is larger, when θ_TWhen value is smaller, missing inspection can be caused；By adjusting θ_TValue, the accuracy of Book Inventory can be improved, if books pendulum Being tried one's best when putting is disposed vertically books, and the effect that this method is checked is more preferable.

Problem is checked from what embodiment the result can be seen that the present invention solves books on frame well：First, root Frame information is extracted according to image edge pixels accumulation, so as to divide the image into multiple subgraphs, each subgraph is represented on bookshelf One spine image put in lattice；Then, it is empty in parameter by books boundary straight line in random hough Mapping implementations subgraph Between in Rapid Accumulation；Finally, local peaking is detected in parameter space, so as to realize each on bookshelf putting books volume in lattice Several countings are checked, and finally give the volumes of all books on bookshelf；The present invention preferably combines image procossing and random The advantage of hough conversion, the auto inventory of books on frame is realized using computer disposal；Method is simply efficient, it is easy to implement, greatly The big workload for reducing Book Inventory, with stronger practicality and application value.

The part that the present embodiment is not described in detail belongs to conventional means commonly understood in the industry, does not describe one by one here.

Claims (5)

1. a kind of Book Inventory method on frame converted based on image segmentation and random hough, it is characterised in that including following step Suddenly：

Step 1：Image preprocessing

Coloured image of the collection comprising bookshelf framework and spine, is modified to picture, obtains the consistent picture of horizontal direction, and Ensure that bookshelf framework takes entire image, the image is designated as P；With the image lower left corner as origin, x-y rectangular coordinate systems are set up, its Middle coordinate system x-axis direction overlaps with image lower boundary, and coordinate system y-axis direction overlaps with left picture boundary；I and j is made to be respectively figure As coordinate system x and y directions pixel cell sequence number, then P can be expressed as the form of picture element matrix Wherein p (i, j) is the pixel value at pixel cell (i, j) place, I be P in the total number of pixels in x directions, J is the total pixel in y directions Number；

Step 2：Image segmentation

Image P is converted into gray-scale map first, neighborhood differential treatment is carried out, the marginal information of image is extracted；Then, by pixel Integration detection obtains the position of bookshelf framework, so as to extract the boundary information that rectangle in bookshelf puts lattice；Finally, put according to rectangle The boundary information of lattice is put, image P is divided into N number of subgraph, during a rectangle puts lattice in each subgraph correspondence bookshelf Spine image, wherein n-th subgraph is designated as P_n；

Step 3：Stochastical sampling

For n-th subgraph P_n, remember I_nIt is it in the total number of pixels in x directions, J_nIt is it in the total number of pixels in y directions；It is first Image preprocessing is first carried out, the gradient direction α at pixel cell (i, j) place is extracted_n(i,j)；Then according to α_nThe value of (i, j) will It is four sectors interval that gradient direction is evenly dividing, and carries out rim detection and obtain binary image Finally, two pixel cell (i are randomly selected from the interval of same sector₁,j₁) and (i₂,j₂) as a pixel pair, wherein i₁ And j₁Respectively first sampled pixel coordinate system x and y directions unit number, i₂And j₂Respectively second sampled pixel In the pixel cell sequence number in coordinate system x and y directions；

Step 4：Parameter is accumulated

Accumulative element distance is calculated, hough parameter spaces (ρ, θ) are set up, ρ and θ is respectively distance and angle parameter in space； By two pixel cell (i of stochastical sampling₁,j₁) and (i₂,j₂) parameter space is projected to by hough conversion, calculate corresponding Parameters unit (ρ_hough,θ_hough), ρ_houghAnd θ_houghDistance and angle parameter after respectively hough conversion in parameter space； Corresponding accumulative element carries out ballot accumulation, the collinear sampled point of correspondence is fallen into identical accumulative element, obtains Parameter accumulates matrix H (ρ, θ)；

Step 5：Peakvalue's checking

According to detection threshold, accumulating matrix H (ρ, θ) to parameter carries out Threshold detection, obtains the number t of local peaking in matrix_n, Then subgraph P_nThe volumes of middle books is t_n-1；

Step 6：Counting is checked

When the value of n changes from 1 to N, for all of subgraph P_n, repeat step 3 to step 5 process, detection obtain institute There is the peak value number t that subgraph parameter is accumulated in matrix_n, n=1,2 ..., N, then the volumes T of books is on whole bookshelf:

$T={\Σ}_{n=1}^N(t_n-1),$

Wherein,Represent when the value of n changes from 1 to N to bracket in content carry out read group total.

2. Book Inventory method on a kind of frame converted based on image segmentation and random hough according to claim 1, its It is characterised by the image partition method described in step 2：

S21：Gray proces

Coloured image P is converted into 256 grades of gray-scale map, the gray-scale map is designated as F, it can be represented in x-y rectangular coordinate systems It is imaged the form of prime matrixWherein f (i, j) is the pixel value at pixel cell (i, j) place；

S22：Neighborhood differential treatment

For any pixel cell f (i, j), differential process are carried out by neighborhood unit pixel, calculate the gradient f in x directions_xWith y side To gradient f_y：

$\left\{\begin{array}{c}{f}_x=f(i,j)+f(i+1,j)-f(i,j+1)-f(i+1,j+1)\\ f_y=-f(i,j)+f(i+1,j)-f(i,j+1)+f(i+1,j+1)\end{array}\right.,$

Then gradient magnitude of the image at coordinate unit (i, j) place isGradient direction be α (i, j)= arctan(f_y/f_x), wherein arctan () represents tan of negating；

S23：Extract edge local extremum

The local extremum of image border is extracted by non-maxima suppression, so as to reject the influence of part non-edge point；First, According to the value of α (i, j), it is four sectors interval that the gradient direction at pixel cell f (i, j) place is evenly dividing, and is represented respectively The gradient of horizontal direction, left-leaning direction, Right deviation direction and vertical direction：

Wherein, s (i, j) is the sector belonging to coordinate unit (i, j) place gradient direction, and π is the corresponding radian number in 180 ° of angles；So Afterwards, according to s (i, j) value, two 8 neighborhood coordinate units of the same gradient direction in coordinate unit (i, j) place are searched

Wherein,WithRespectively pixel cell sequence number of the two 8 neighborhood coordinate units of (i, j) in coordinate system x and y directions；Most Afterwards, by the neighborhood territory pixel of the gradient magnitude at coordinate unit (i, j) place and same gradient directionCompare, extract gradient magnitude Local maximum：

S24：Rim detection

Binary conversion treatment is carried out to gradient magnitude using dual threshold thresholding, if Low threshold thresholding is σ_low, high threshold thresholding is σ_high, Rim detection is carried out to image and obtains binary imageWherein b (i, j) is pixel cell (i, j) place Binaryzation pixel value：

For gradient magnitude between σ_lowAnd σ_highBetween coordinate unit (i, j), the value of b (i, j) is according to 8 neighborhood territory pixel units Gradient determine

Wherein, M_maxIt is the maximum of the neighborhood territory pixel unit gradient of coordinate unit (i, j) place 8；

S25：Framework is extracted

The edge of bookshelf framework is straight line both horizontally and vertically, and in framework the edge of books then exist it is more diagonal oblique Line, therefore treatment is filtered to binary image B using a diagonal matrix, obtain new pixel value：

B (i, j)=max [b (i, j), b (i-1, j-1), b (i+1, j+1), b (i-1, j+1), b (i+1, j-1)],

Wherein, max [] is represented and is taken maximum computing；Then bianry image is obtained into figure respectively to both horizontally and vertically projecting As accumulation parameter Q in the horizontal direction_xThe accumulation parameter Q of (i) and vertical direction_y(j)：

$\left\{\begin{array}{c}{Q}_x\left(i\right)={\Σ}_{j=1}^{J}b(i,j)\\ Q_y\left(j\right)={\Σ}_{i=1}^{I}b(i,j)\end{array}\right.,$

According to the horizontal detection threshold γ of accumulation parameter_xWith vertical detection thresholding γ_y, extract Q_xI the corresponding sequence number of () valley is deposited Enter array D_x, extract Q_yJ the corresponding sequence number of () valley is stored in array D_y：

$\left\{\begin{array}{c}{D}_x=\underset{i\∈\[0,I\]}{\arg }\[Q_x\left(i\right)\≥{\mathrm{\γ}}_x\]\\ D_y=\underset{j\∈\[0,J\]}{\arg }\[Q_y\left(j\right)\≥{\mathrm{\γ}}_y\]\end{array}\right.$

Wherein, arg [] represents the sequence number for taking and meeting constraints [], then D_xData correspondence bookshelf framework is in the horizontal direction Coordinate, D_yCoordinate of the data correspondence bookshelf framework in vertical direction；

S26：Segmentation figure picture

Next original image is divided into multiple subgraphs, each subgraph correspondence bookshelf gets on during a rectangle puts lattice Spine image；First, the position coordinates D of bookshelf framework is extracted_xAnd D_y, so as to realize putting each rectangle in bookshelf the positioning of lattice； Then, the border that lattice are put according to rectangle is split to coloured image P, obtains multiple subgraphs, realizes dividing for original image Cut treatment：

Wherein N is the total number of subgraph, P_nRepresent n-th subgraph.

3. Book Inventory method on a kind of frame converted based on image segmentation and random hough according to claim 1, its It is characterised by the stochastical sampling method described in step 3：

S31：Image preprocessing

According to the method for step S22 to subgraph P_nPre-processed, obtained the gradient direction α at pixel cell (i, j) place_n(i, j)；Then according to the method for step S23 by α_nIt is four sectors interval that (i, j) is evenly dividing；Use s_n(i, j) represents subgraph P_n The sector belonging to gradient direction at coordinate unit (i, j) place；Method according to step S24 carries out rim detection and obtains binaryzation ImageWherein b_n(i, j) is subgraph P_nIn the binaryzation pixel value at pixel cell (i, j) place；

S32：Choose first pixel

In binary image B_nIn randomly select a pixel cell (i for nonzero element₁,j₁) so that 0.3J_n≤j₁≤0.7J_n And b_n(i₁,j₁) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₁,j₁)；

S33：Choose second pixel point

In binary image B_nSecond pixel cell (i of nonzero element of middle selection₂,j₂)≠(i₁,j₁) so that 0.3J_n≤j₂≤ 0.7J_nAnd b_n(i₂,j₂) ≠ 0, it is s that the corresponding gradient direction of the pixel cell is interval_n(i₂,j₂)；

If s_n(i₁,j₁)=s_n(i₂,j₂), carry out step S34；

If s_n(i₁,j₁)≠s_n(i₂,j₂), repeat step S33, until s_n(i₁,j₁)=s_n(i₂,j₂), carry out step S34；

S34：Composition stochastical sampling pixel pair

By two pixel cell (i₁,j₁) and (i₂,j₂) one stochastical sampling pixel pair of composition；

S35：Terminate sampling

Repeat step S32~S34, obtains L stochastical sampling pixel pair.

4. Book Inventory method on a kind of frame converted based on image segmentation and random hough according to claim 1, its It is characterised by the parameter accumulation method described in step 4：

S41：Set up parameter accumulation space：

Sliding-model control is carried out to parameter ρ and θ according to discrete interval △ ρ and △ θ, parameter space (ρ, θ) is formed, (u, v) is individual The central point parameter of parameters unit is ρ_uAnd θ_v：

ρ_u=(u-1/2) △ ρ, u=1,2 ..., U,

θ_v=(v-1/2) △ θ-pi/2, v=1,2 ..., V,

Wherein, u and v are respectively the sequence number of parameter ρ and θ, the parameters unit number V values point of parameters unit the number U and θ of ρ It is not：V=Int (π/△ θ), Int () represent rounding operation, and π is the corresponding radian in 180o angles Number；

S42：Set up parameter accumulation matrix

According to the parameter space (ρ, θ) of definition, set up parameter and accumulate matrix H (ρ, θ), it is 0 to put each unit initial value；

S43：Image border is mapped to parameter space

By first pixel of stochastical sampling to (i₁,j₁) and (i₂,j₂) parameter space is projected to by hough conversion, calculating should Pixel is to identified parameter ρ_houghAnd θ_hough：

θ_hough=arctan [- (i₁-i₂)/(j₁-j₂)]；

ρ_hough=0.5 (i₁+i₂)cos(θ_hough)+0.5(j₁+j₂)sin(θ_hough)；

S44：Ballot accumulation

Corresponding unit (ρ is searched in parameter space (ρ, θ)_u,θ_v) so that | | ρ_hough-ρ_u||≤△ρ,||θ_hough-θ_v||≤ △ θ, wherein | | | | absolute value is sought in expression；

If (ρ_u,θ_v) exist, then enter line parameter in the corresponding units of parameter accumulation matrix H (ρ, θ) and add up：H(ρ_u,θ_v)=H (ρ_u,θ_v)+1；

Otherwise, by (ρ_hough,θ_hough) insertion parameter space (ρ, θ), and take initial value H (ρ_hough,θ_hough)=1；

S45：Terminate accumulation

Repeat step S43~S44, until L stochastical sampling pixel is to being mapped to parameter space, obtains parameter accumulation matrix H (ρ,θ)。

5. Book Inventory method on a kind of frame converted based on image segmentation and random hough according to claim 1, its It is characterised by the peak-value detection method described in step 5：

S51：Threshold detection

According to detection threshold G₁, matrix H (ρ, θ) accumulating to parameter and is detected, value is less than G in making H (ρ, θ)₁Element value be Zero, detection threshold is set to G₁=0.3J_n；

S52：Angle restriction

Consider books inclination angle, angle restriction value θ is set_T, parameter θ absolute value is less than θ in obtaining H (ρ, θ)_TUnit, make the list First respective element value is zero；

S53：Local Search

Parameters unit (the ρ that any one is not zero in H (ρ, θ)_u,θ_v), define its local search area Ω (ρ_u,θ_v) be：

Ω(ρ_u,θ_v)={ (ρ, θ) | [| | ρ-ρ_u| |≤5, θ=θ_v,H(ρ,θ)≠0]}；

S54：Peak extraction

Next according to Ω (ρ_u,θ_v) judged：

If 1) Ω (ρ_u,θ_v) it is sky, then parameters unit (ρ_u,θ_v) it is a local peaking；

If 2) Ω (ρ_u,θ_v) it is not sky, if H (ρ_u,θ_v) value be more than or equal to Ω (ρ_u,θ_v) in all units value, i.e., Meet：

H(ρ_u,θ_v) >=H (ρ, θ), (ρ, θ) ∈ Ω (ρ_u,θ_v),

Then parameters unit (ρ_u,θ_v) it is a local peaking；

S55：Peak counting

Repeat step S53~S54, extracts H (ρ, θ) all of local peaking's unit, calculates the number t of local peaking_n, then subgraph As P_nThe volumes of middle books is t_n-1。