An Automated Fish Counting Algorithm in Aquaculture Based on Image Processing

Transcription

1 Inernaional Forum on Mechanical, Conrol and Auomaion (IFMCA 06) An Auomaed Fish Couning Algorihm in Aquaculure Based on Image Processing Jiuyi Le,a, Lihong Xu,b College of Elecronics and Informaion Engineering, Tongji Universiy, Shanghai, China a lejiuyi9@63.com, bxulhk@63.com Keywords: Fish Couning; Free-swimming; Overlap; Skeleon Exracion Absrac. A new algorihm based on endpoins of skeleon is presened o efficienly ge he number of fish in his paper. Considering he complexiy of underwaer environmen like lack of ligh, his paper presens an improved adapive hresholding mehod o segmen he fish image beer. In addiion, he objec of our research is free-swimming fish. The overlapped fish in he image makes he couning resul inaccurae ofen. So afer segmenaion and morphological processing, his paper adops image hinning mehod o exrac he skeleon of fish. Afer ha, we ge he fish number according o he number of corresponding endpoins in he image. The experimenal resuls show ha he mehod can accuraely coun he fish populaion even under high overlapped degree. Inroducion Wih he inroducion and popularizaion of he concep of marine ranching, he offshore aquaculure indusry has developed rapidly in recen years. The precise couning of he fish socks provides he basis for he effecive managemen of scienific feeding, sale, ransporaion, and breeding densiy conrol. The radiional couning mehod uses conainers like ne bag o sample, which brings los of disadvanages like low efficiency, limied arificial experience and so on []. Since he 80s of he 0h cenury, a variey of research insiues invened a number of fish couners, which was difficul o apply o he acual producion because i can be easily affeced by environmenal consrains, fish size and oher facors. Besides, i is so expensive ha few fishermen wan o use i in acual producion []. Wih he improvemen of he compuer vision and image processing echnology, hey are applied o more indusries o improve he level of auomaion of producion. A he same ime, fish couning mehod based on image processing is aracing more and more aenion due o is high efficiency and accuracy [3]. Some research mehods, such as conneced area couning, area couning, daa fiing [4], curve evoluion [5] and neural nework [6], did no ake he overlapped problem ino consideraion. So hey will no give an accurae number of fish when some fishes ge ogeher. In his paper, we propose an algorihm based on endpoins of exraced skeleon, which aenuaes he segmened images so ha he fish images are refined ino some one-pixel wide skeleons. And hen he fish populaion is accuraely couned using he number of endpoins. Experimenal Plaform The experimenal plaform consiss of hree pars. The whole experimen is carried ou in a ransparen glass aquarium, he size of he fish ank is.m 0.6m 0.5m. The second par is an underwaer image acquisiion module, which consiss of a waerproof camera and an underwaer cloud plaform. The camera capure he video verically and we can adjus he bracke o obain he appropriae focal lengh. The camera is SHARP /4 CCD model, surrounded by a circle of LED compensaing underwaer ligh in he dark condiion. The hird par of he sysem is he hos compuer. Copyrigh 07, he Auhors. Published by Alanis Press. This is an open access aricle under he CC BY-NC license (hp://creaivecommons.org/licenses/by-nc/4.0/). 358

2 The analog video mus be convered o digial signal o be saved in he compuer. The resoluion of all he video in his experience is The overall sysem diagram is shown in Figure : Figure. Experimenal Plaform Image Preprocessing Due o he diversiy and complexiy of he underwaer environmen, he underwaer images usually suffer from severe noise, which reduces he qualiy of underwaer images and affecs he accuracy of image analysis. In order o make he underwaer image easier o analyze, we need o carry on appropriae prereamen o he underwaer picure [7]. For he objecs o be sudied, we only focus on he conour informaion of he fish in he image. In order o remove redundan informaion, he auhor firs convered he underwaer color image o grayscale image. In he RGB model, if R = G = B, hen he color informaion represens a grayscale color, where he value of R = G = B is called he grayscale value. So every pixel of he grayscale image only need one bye o sore he grayscale value (also called inensiy value or brighness value) which range from 0 o 55. The auhor uses he weighed average mehod o gray he image. Every pixel ge is grayscale value according o he formula: f (i, j ) = 0.30 R(i, j ) G (i, j ) + 0.B(i, j ). Then he median filer is used o remove he noises. The resul afer one preprocessing operaion is saisfacory. The sample grayscale image is as follows: Figure. Grayscale image Image Segmenaion Image segmenaion is an imporan sep in he analysis of image daa, and in all image segmenaion algorihms, hreshold segmenaion has been widely used for a long ime for is exreme simpliciy and high pracicaliy. For a beer view, all he images appear in his aricle are images ha reverse he black and whie pixels of original images. 359

3 Osu Mehod and Relaed Derivaives The maximum beween-class variance is an imporan basis for saisical unsupervised clusering in saisical paern recogniion. The image segmenaion mehod based on maximum beween-class variance, which is also called Osu mehod, is proposed by famous Japanese scholar NOBUYUKI [8]. Among he global hresholding echniques, he Sahoo e al. sudy [9] concluded ha he Osu mehod was one of he beer hreshold selecion mehods for general real world images wih respec o uniformiy and shape measures. This mehod selecs hreshold values ha maximize he beween-class variances of he hisogram. Le he pixels of a given image be represened in L gray levels, and le he probabiliy of he occurrence of a pixel wih a grayscale i o be approximaed by is frequency. The normalized hisogram of he image is: H = { p 0, p... pl } pi =. () i=0 Now suppose ha we dichoomize he pixels ino wo classes B and O (background and objec, or vice versa) by a hreshold a level ; B denoes pixels wih levels [, L, ], and O denoes pixels wih levels [ + L, L]. The probabiliy of background and objec is PB ( ) and PO ( ), respecively. The average gray level of he background and objec is µ B ( ) and µo ( ), respecively. So he overall average gray level of he image is: µ= ip i. () According o he definiion of variance, he oal variance of levels of his image is: σ = E[(i µ ) ] = pi (i µ ). (3) i=0 Then we can easily find ha for a given image, he average gray level µ and variance σ are boh consan. So hey are independen of he hreshold discussed here. The wihin-class variance of he background class and he objec class is defined as: σ B = (4) (i µ B( )) * pi. PB ( ) (5) σ O = (i µo ( )) * pi. PO ( ) i = + Le σw ( ) represen he wihin-class variance, which is defined as: σ w ( ) = PB ( ) σ B + PO ( ) σ O. (6) Le σ ( ) represen he beween-class variance, which is defined as: OB σ OB ( ) = PO ( ) ( µo ( ) µ ) + PB ( ) ( µ B ( ) u ). (7) In fac, L L σ w + σ OB = ( i µt ) + i µt µt pi = ( i µt ) pi = σ. i = i = (8) Therefore, he oal variance of he image is equal o he sum of he wihin-class variance and beween-class variance of he objec class and he background class: σ = σ W + σ OB. (9) Therefore, he hreshold seleced by Osu mehod is a value which maximizes he σ OB ( ). In oher words, he formula o selec opimal hreshold * according o Osu crierion is ha: * = Arg max σ OB ( ). 0 ( ) (0) 360

4 . So ha equaion According o he formula (9), we can easily ge he equaion ha σ W = σ σ OB (0) is equivalen o: () * = Arg min PB ( ) σ B ( ) + PO ( ) σ O ( ). 0 Equaion () shows ha maximizing beween-class variance and minimizing wihin-class variance can be equivalen o ge he opimal hreshold based on OTSU crieria. However, Hou and oher people in he Research Insiue of Communicaions in Singapore have poined ou in documen [0] ha The Osu mehod works well only when he images o be hresholded have clear peaks and valleys in oher words, i works for images whose hisograms show clear bimodal or mulimodal disribuions. In order o overcome he problem ha he opimal hreshold seleced in Osu always lean o he class wih larger variance, a new mehod based on minimum class variance o selec hreshold was presened in documen [0]. In his documen, he minimized objecive funcion is: * = Arg min σ B ( ) + σ O ( ). () 0 I has been proved in documen [0] ha his mehod can buffer he endency of he hreshold moving o he class wih larger variance. Wu poined in he documen [] ha Osu is an opimal approximaion based on L-norm. In his paper, an image segmenaion mehod based on L-norm is proposed, and is opimal hreshold is seleced by: = Arg min {S w ( )} = Arg min {w0 ( ) d 0 ( ) + w ( ) d ( )}. (3) 0< < Where w0 ( ) and w ( ) represen he class probabiliy of he background and objec, respecively; d 0 ( ) d ( ) represen absolue difference of he background and objec, respecively. They are shown as d0 ( ) = pi ω0 ( ) i µ0 ( ), d ( ) = pi ω ( ) i µ ( ). i = + (4) An Exponenial Improved Algorihm Based on Osu Wheher he original Osu mehod or Hou s and Wu s improved mehod, all of hem have some resriced condiions on gray hisogram if hey wan o ge a good resul. However, he gray hisogram of he underwaer image is varied due o various facors such as illuminaion, waer urbidiy and so on. According o he improvemen research of Hou and Wu, we can find ha class probabiliy or hisogram gray inerval is helpful for Osu mehod o segmen he grayscale image beer. Based on all he above research, his paper presens an improved Osu algorihm which akes boh class probabiliy and hisogram gray inerval ino consideraion. The crierion funcion of hreshold selecion is: σ wq ( ) = pi i µ B ( ) q pi i µo ( ) q. (5) q q i = + PB ( ) PO ( ) Where σ wq ( ) is he absolue wihin-class difference wih exponenial parameer. And is he bes + hreshold when i minimizes σ wq ( ). q and q are wo adjusable parameers, which are used as an index of wihin-class absolue difference and class probabiliy. As can be seen from he formula, he hreshold selecion crieria in his paper fully consider he class probabiliy and gray level difference, so ha i can adap o differen siuaions for image segmenaion. When q and q ake some special value, he mehod presened in his paper is equivalen o he Osu algorihm, Hou s mehod and Wu s improved mehod respecively. They are proved as followed: When q = and q =0, formula (5) is ransformed ino he following cases: 36

5 ( ) = pi i µ B ( ) pi i µo ( ) i = + PB ( ) PO ( ) p [i µ B ( )] p [i µo ( )] = PB ( ) i + PO ( ) i = PB ( ) σ B + PO ( ) σ O = σ w ( ). (6) P P i = + B( ) O( ) Compared wih formula (6), we know ha he mehod of his paper is equivalen o he original Osu mehod a his momen. σ q w When q = and q = 0, we can also prove ha his mehod is equivalen o he minimum wihin-class absolue difference mehod proposed by Wu. The process of derivaion is similar o he derivaion of he above formula (6), which will no be repeaed here. 3 When q = q =, his mehod is equivalen o he minimum variance mehod proposed by Hou. From he above proof, we find ha he original Osu algorihm, Hou's minimum variance mehod and Wu s minimum wihin-class absolue difference mehod are only one of hose specific form of mehod presened in his paper []. These hree mehods have a solid heoreical foundaion of mahemaical saisics, which also shows he raionaliy of he mehod presened in his paper. This mehod is more adapive when you adjus he parameers. For he grayscale image in Figure, he resul using Osu and mehod presened in his paper is show as (b) and (c) in figures 3 respecively. (a) (b) (c) Figure.3 Segmenaion resul The figures above show ha he original Osu mehod is no good enough o process he deail of he images. The Osu mehod misaken a lo of fish ail pixel poins as he background poins. Neverheless, he improved mehod presened in his paper has a beer performance. The binary image has some noises because of he impuriies in he waer. In addiion, here are also some black pixels in he fish area and some fish area is spli. So we conduc a closed operaion (corrosion afer expansion) and hen a median filer on he binary image [3]. And finally, a very ideal binary image is shown in figure 4. 36

6 Figure.4 Binary image afer been filered Skeleon Exracion This paper is aimed o coun he fish precisely. So we are going o hin he fish image in his par o make fish couning easier, which is also called skeleon exracion. The so-called skeleon, can be undersood as he axis of he image, such as recangle s skeleon is he cener line of he long direcion, circle s skeleon is is cener poin [4]. Figure 5 is he objec pixel p and is 8 neighborhood. Every pixel is eiher 0 or. Our mehod for exracing he skeleon of a picure consiss of removing all he conour poins of he picure excep hose poins ha belong o he skeleon. In order o preserve he conneciviy of he skeleon, we divide each ieraion ino wo subieraions [5]. p9 p p3 p8 p p4 p7 p6 p5 Figure.5 8-neighborhood In he firs subieraion, he conour poin p is deleed from he digial paern if i saisfies he following condiions: (a) B( p ) 6, (b) A( p ) =, (c) p p4 p6 = 0, (d) p4 p6 p8 = 0. Where A( p ) is he number of 0 paerns in he order se p, p3... p8, p9 ha are he eigh neighbors of p. B ( p ) is he number of pixels which equals o. In he second subieraion, only condiions (c) and (d) are changed as follows: (c ) p p4 p8 = 0, (d ) p p6 p8 =

7 Figure.6 Flowchar of he hinning algorihm The soluion o he se of equaion (c) and (d) are p4 = 0 or p6 = 0 or ( p = 0 & & p8 = 0 ). So he poin p, which has been removed, migh be an eas or souh boundary poin or a norh-wes corner poin. Similarly, i can be proved ha he poin p deleed in he second subieraion migh be a norh-wes boundary poin or a souh- eas corner poin. By condiion (a), he endpoin of a skeleon line are preserved. Also condiion (b) preven he deleion of hose poins ha lie beween he endpoins of a skeleon line. A flowchar of he hinning algorihm is shown in figure 6. The original image is sored in marix IT and a couner C is se o 0. The resul of he processed picure is sored in marix M. The ieraion sops unil no poins can be deleed. The image of fish skeleon is shown in figure 7. Figure.7 Skeleon image 364

8 Fish Couning When he refinemen process is compleed, he fish populaion can hen be couned based on he number of skeleal endpoins. And hen we process he conneced area one by one. The skeleon shape go by he former algorihm is 8-conneced. So if here are only one pixel from he eigh neighborhood whose value is, hen we rea his pixel as he endpoin of he skeleon. Afer all seps, he number of fish in he conneced area is deermined by he number of endpoins: If here are endpoins in he conneced area, and hen we believe he number of fish is. If here are 3 endpoins in he conneced area, and hen we believe he number of fish is. If here are 4 endpoins in he conneced area, here may be or 3 fish. And hen we believe he number of fish is.5. If here are n (n > 4) endpoins in he conneced area, and hen we believe he number of fish is n/+. Finally, he number of fish in all conneced areas is summed up o ge he oal number of fish [6]. Conclusion In order o solve he problem of fish couning, his paper inroduces he compuer vision echnology o improve he auomaion level. In view of he characerisics of underwaer images, a more general adapive hresholding mehod is proposed, and he segmenaion of fish image is more accurae. Secondly, his paper adops skeleon exracion mehod o solve he overlapped-fish problem cleverly. Three ypical cases are shown as figure 8. From (a) o (c), more and more fish gaher ogeher which makes our resul less accurae. We ge hundreds of image from he videos, and experimens prove ha he average couning error is less han 6%, which is much beer han some radiional mehod like conneced area mehod and so on. Bu in he acual producion, he underwaer environmen is much complex. For example, low visibiliy and lack of ligh and oher issues are he urgen problem o be solved. When he fish densiy is oo large, los of hem aach ogeher. A ha ime, he resul will no be precise as before. In he fuure work, we plan o conduc more experimens wih he simulaion of underwaer environmen. (a) 00% (b) 00% (c) 9.% Figure.8 Accuracy of couning wih differen degrees of overlap 365

9 Acknowledgemens This work was suppored in par by he Naional High-Tech R&D Program of China under Gran 03AA0305, he Naional Naural Science Foundaion of China under Gran and , and in par by he U.S. Naional Science Foundaion's BEACON Cener for he Sudy of Evoluion in Acion, under cooperaive agreemen DBI References [] Assis J, Claro B, Ramos A, e al. Performing fish couns wih a wide-angle camera, a promising approach reducing divers' limiaions[j]. Journal of Experimenal Marine Biology & Ecology, 03, 445: [] Yi Jin-geng Huang Gui-lin Design of Opical-Elecric Compuer Conrolled Sysem for Muli-channel Fry Couner J. Transacions of he Chinese Sociey of Agriculural Engineering, 997, [3] CHATAIN B, DEBAS L, BOURDILLON A A phoographic larval fish couning echnique comparison wih oher mehods, saisical appraisal of he procedure and pracical use J Aquaculure, 995, 4 - : [4] Cong rong Zhu. A mehod for fish fry auomaic couning based on machine vision [J]. Fishery Modernizaion, Volume 36, Issue, 009. [5] WANG Shuo, FAN Liangzhong, LIU Ying. The research of urbo fry couning mehod based on compuer vision [J]. Fishery Modernizaion, 05, 4():6-9. [6] Zheng X, Zhang Y, Zheng X, e al. A Fish Populaion Couning Mehod Using Fuzzy Arificial Neural Nework[C]. IEEE Inernaional Conference on Progress in Informaics and Compuing. 00:5-8. [7] Jiang H F, Xu Y. Research Advances on Securiy Problems of Underwaer Sensor Neworks [J]. Advanced Maerials Research, 0, 37-39: [8] OTSU N. A hreshold selecion mehod from gray level hisograms [J]. IEEE Trans on SMC, 979, 9(): [9] Sahoo, P.K., Solani, S., Wong, A.K., Chan, Y.C., 988. A survey of hresholding echniques. Compuer Vision, Graphics and Image Processing 4, [0] Hou Z, Hu Q, Nowinski W. L. On minimum variance hresholding [J]. Paern Recogniion Leers, 006, 7(4): [] WU Yi-quan, PAN Zhe. The Image Thresholding Algorihm Based on Minimum Wihin-Cluser Absolue Difference and Maximum Difference [J]. Journal of Signal Processing. 008(06): [] Yang Shuhong. Sudy on he Adapive and Fas Algorihm of Gray Scale Image Thresholding[D]. Chongqing Universiy, 04. [3] Fu X, Zeng D, Huang Y, e al. A fusion-based enhancing mehod for weakly illuminaed images [J]. Signal Processing, 06, 9:8-96. [4] Liang Yu, Guangwei Wang. A Fas Thinning Algorihm for Binary Image [J]. Science & Technology Vision, 0(4):-. [5] Zhang T Y, Suen C Y. A fas parallel algorihm for hinning digial paerns [J]. Communicaions of he Acm, 984, 7(3): [6] Fan S, Liu J, Yang Y. Research and Realizaion of Image Recogniion Technology in Fry Couning [J]. Fisheries Science, 008, 7(4):