九州工業大学学術機関リポジトリ Ttle Identfy a Specfed Fsh speces b Confuson Matrx Author(s) Yan, Zyue; Zhang, Lfeng; Hu, Xuel Issue Date 2015-03-29 URL http://hdl.handle.net/10228/5835 Rghts 産業応用工学会 Kyushu Insttute of Technology Academc Re
Proceedngs of the 3rd Internatonal Conference on Industral Applcaton Engneerng 2015 Identfy a Specfed Fsh speces by the Co-occurrence and Confuson Matrx Zyue Yan 1,2, Lfeng Zhang 1, Xuelong Hu 2, Sech Serawa 1 1 Department of Electrcal Engneerng and Electroncs, Kyushu Insttute of Technology, 1-1 Sensu-cho,Tobata-u, Ktayushu 804-8550,Japan 2 School of Informaton Engneerng, Yangzhou Unversty, 198 Huayangx Road, Hanang, Yangzhou 225127, Chna *Correspondng Author: yanzyue07@gmal.com Abstract Nowadays, nvasve speces threaten natve speces has become a global problem. Invasve speces mght be carryng pathogenc mcroorgansms, reduce bologcal speces and even threat to human health. Therefore, n ths study, we proposed a method of co-occurrence matrx to texture analyss of three speces of fsh. We catch the body pattern, and mae a udgment based on confuson matrx. Smulaton results show that three speces of fsh can be classfed from each other reasonable. Keywords: nvasve speces, co-occurrence matrx, confuson matrx. 1. Introducton In recent years, many foregn creatures destroy the ecosystem of natve speces, and ths becomes a serous worldwde problem now. Foregn creature was called specalzed nvasve alen speces, and Japanese government ssued the "nvasve alen speces act'' n 2004. In Japan, there are 14 nds of alen speces about fsh. Invasve speces can destroy the ecologcal envronment. When a stranger creature allowed nto ecosystem, natve speces would loss of lvng space and food, stranger creatures may be release chemcals and attac natve speces. Ultmately, natve speces wll become extnct f they are not be protected tmely. In recent years, the man way of catchng stranger creature s manual control. But t wll cost much money and manpower. Therefore, t s necessary to get rd to them at the begnnng of breedng. Natural texture has sophstcated structure. Now there s no unfed standard to descrbe ts characterstcs. How to characterze the natural texture and classfcaton s an mportant drecton n the feld of computer vson research. The way of texture characterstcs can be dvded nto statstcs method and structural method. Statstcal method s more sutable for processng natural texture (1). In statstcal method, the co-occurrence matrx s classcal way to descrbe the texture. The generated parameters can descrbe the statstcal characterstcs of the texture n many aspects. In ths study, we use co-occurrence matrx for texture analyss to three nds of fshes, crap, sunfsh and blac bass. We get a part of fsh body surface mage, mae a co-occurrence matrx and then calculate four nds of characterstcs form co-occurrence matrx. Fnally we use confuson matrx to evaluate the smulaton result of the proposed method. From the evaluaton value, t s proved that the proposed method can dentfy the fsh speces effectvely. ln secton 2, co-occurrence matrx s ntroduced. In secton 3, confuson matrx method s selected for ths study. Smulatons for fsh texture analyss wth two methods are revealed n secton 3. 2. Co-occurrence matrx 2.1 Co-occurrence matrx defnton In 1973, Haralc proposed co-occurrence matrx to descrbe the texture feature (2). Co-occurrence matrx s defned over an mage to be the dstrbuton of co-occurrng values at a gven offset. G, (, 0,1,2, 1) (1) d DOI: 10.12792/cae2015.006 12 2015 The Insttute of Industral Applcatons Engneers, Japan.
Where and are the mage ntensty values of the mage, d s the relatonshp between pxels of locaton. When two pxel locatons d s selected, we can get matrx wth d (3). G d 0,0 0,1 0, 1 1,0 1,1 1, 1 Gd Gd Gd Gd Gd Gd G d Gd Gd 1,0 1,1 1, 1 2.2 Co-occurrence matrx characterstc parameters (2) Haralc defnes 14 nds of feature parameters for texture analyss of co-occurrence matrx (4). Ulaby's study found that based on 14 nds texture feature, ust 4 characterstc parameters are rrelevant (5). These values are easy to calculate and they deserve a more accurate classfcaton. Entropy: the amount of nformaton whch s the mage measurement. ENT G ( )log ( ) (3) 1 1 Energy: the sum of the squares of the gray level co-occurrence matrx elements, so also nown as energy, reflects the mage grey dstrbuton unformty degree of thcness and texture. ASM 1 1 G(, ) 2 (4) Contrast: reflects the mage of the degree of clarty and texture groovng depth. If the value of contrast s small, the pcture s fuzzy. 1 2 CON n G (, ) (5) n0 n Relevancy: t s used to measure the smlarty degree of the elements of gray level co-occurrence matrx on the row or column, therefore, the relatve value reflects the local gray correlaton of mage sze. (6-7) COR 1 1 G(, ) (6) 1 1 1 1 ss G, G, 2 1 1, 2 s G 2.3 Co-occurrence matrx characterstc values Table 1~3 show the four characterstc value of co-occurrence matrx for each fsh. We capture one part of each fsh. And we can fnd dfferent fshes have dfferent values. Table 1: Statstcal results of sunfsh body texture Sunfsh1 6.610716 0.216256 0.72475 0.30549 Sunfsh2 6.682494 0.221526 0.758724 0.287059 Sunfsh3 6.567916 0.222732 0.692758 0.34549 Sunfsh4 5.958765 0.347993 0.672726 0.196863 Sunfsh5 6.613791 0.215509 0.74275 0.320784 Sunfsh6 6.464438 0.261066 0.680734 0.326667 Sunfsh7 6.884852 0.148534 0.734582 0.417647 Sunfsh8 6.662622 0.241086 0.786118 0.249804 Sunfsh1 Sunfsh2 Sunfsh3 Sunfsh4 Sunfsh5 Sunfsh6 Sunfsh7 Sunfsh8 Fg 1:Body texture pattern of sunfsh Table 2: Statstcal results of sunfsh body texture Crap1 7.179772 0.12919 0.862893 0.349412 Crap2 7.231622 0.0834 0.649161 0.900784 Crap3 7.07714 0.113908 0.676716 0.736078 Crap4 7.157248 0.100221 0.657833 0.921569 Crap5 6.513132 0.220037 0.619326 0.389804 Crap6 6.661643 0.203593 0.725724 0.313333 Crap7 7.100701 0.107043 0.738975 0.555686 Crap8 7.305372 0.082896 0.840736 0.673333 Crap9 7.328107 0.081013 0.69722 1.030196 2 1 1, 2 s G 13
3. Smulaton Crap1 Crap2 Crap3 Crap4 Crap5 Crap6 Crap7 Crap8 Crap9 Fg 2:Body texture pattern of crap Table 3: Statstcal results of bass body texture Bass1 6.381112 0.309316 0.739679 0.20627 Bass2 6.211198 0.348428 0.806412 0.14823 Bass3 6.182986 0.337062 0.667224 0.20745 Bass4 6.572694 0.242015 0.764562 0.27841 Bass5 7.028404 0.106752 0.602086 0.72784 Bass6 6.412045 0.266177 0.666517 0.27764 Bass7 5.779659 0.510017 0.64922 0.13176 Bass8 6.407621 0.300795 0.764388 0.20039 Bass9 6.857924 0.117146 0.476464 0.77843 Bass10 6.834359 0.128222 0.551338 0.641176 Bass11 6.566195 0.244329 0.701262 0.33215 Bass12 6.569027 0.274583 0.700281 0.38745 Bass13 6.414443 0.260578 0.465271 0.46352 Bass14 6.412397 0.214351 0.572996 0.4201 Bass15 6.547202 0.216138 0.525349 0.53725 Bass16 6.751274 0.138794 0.600531 0.56784 Bass1 Bass2 Bass3 Bass4 Bass5 By adoptng the co-occurrence matrx n ths experm ent, we can quantfy three nds of fsh body texture. Accordng to the values of co-occurrence matrx, we adopt least square method to ft coeffcent. After settng a fxed output fgure of the same fsh, we can ft coeffcent through four statstcal results of every sngle fsh speces. Calculate the actual value, after gettng the coeffcent and compare t wth the settng fgure. Accordng to the results of comparson and analyss, we can use confuson matrx to deal wth. In mage accuracy assessment, t s manly used for comparatve classfcaton results and actual measured value. lt shows the result n confuson matrx about the precson of the classfcaton. Table 4:Cofuson matrx Name Sample1 Sample2 Sample3 Sum Sample1 a b c A Sample2 d e f B Sample3 g h C Sum D E F G A a b c D=a+d+g B =d e f E=b+e+h C =g h F=c+f+ G=A+B+C (7) Kappa: a nd of calculaton method of classfcaton accuracy. A appa calculaton result ranges from -1 to 1, but usually appa falls between 0 to 1. It can be dvded nto fve groups to represent dfferent levels of consstency. Pa Pe Kappa= 1 Pe (8) Bass6 Bass7 Bass8 Bass9 Bass10 Bass11 Bass12 Bass13 Bass14 Bass15 Bass16 Fg3:Body texture pattern of bass a e Pa G A D B E C F Pe GG Table 5:The result of appa Interval Results 0.0-0.20 low consstency(slght) 0.21-0.40 general consstency (far) 0.41-0.60 medum consstency (moderate) 14
0.61-0.80 a hgh degree of consstency (all) 0.81-1 almost perfect Table 6: Statstcal results of Confuson matrx Name Crap Sunfsh Bass Sum Crap 7 1 1 9 Sunfsh 0 6 2 8 Bass 0 4 12 16 Sum 7 11 15 33 97 8111615 Pe 0.359045 3333 7+6+12 Pa =0.757576 33 appa=0.621777 From Table 6, we calculate the value of accuracy. Accordng to Table 5, the result of appa s n a hgh degree of consstency. segmentaton va wavelet transform,vol.1997,no. 05,2010 (5) Ulaby FT, Kouyate F'Brseo B, et a1: Textural nformaton n SAR Images, IEEE Transactons on Geo scence and Remote Sensng, Vol.24, No.2, pp.235-245,1986 (6) BharatM H, Lu J J,MAC G: Image texture analyss: Methods and comparsons, Chemo Metrcs and Intellgent Laboratory Systems, Vol.72,No.1, pp.57-65, 2004 (7) Yao H Y, L B C : Generalzed co-occurrence matrx method for content-based mage retreval[j]. Computer Engneerng Applcatons, Vol.34, No.6, pp.98-101, 2004 4. Conclusons ln ths study, we propose a partcular method to dentfy specfed fsh speces by co-occurrence and confuson occurrence matrx. Due to the lmted space, ths study ust selects three nds of fshes. It may not applcable to all fshes, but the four characterstc value of co-occurrence are effectve parameters to descrbe the body texture. Ths concluson can be used as a reference for the classfcaton. But the method based on co-occurrence also has ts own dsadvantages. It requres a great amount of calculaton, not to menton that t's tme-consumng. In future, we want to ncrease the amount of fshes and verfy the study n a practcal condton. References (1) Ba xuebng, Wang eq, Wang hu: Research on the classfcaton of wood texture based on gray level co-occurrence matrx, Journal of harbn nsttute of technology,vol.37,no.12,pp.1167-1170,2005 (2) Haralc R M,Shanmugam K: Texture features for mage classfcaton, IEEE Tmns on Sys, Man, and Cyb,Vol.3,No.6,pp.610-621, 1973 (3) GAO Cheng-Cheng, HUI Xao-We: GLCM Based Texture Feature Extracton, BEIJING SURVEYING AND MAPPING, Vol.19,No.6,pp.195-198, 2010 (4) C.S.Lu, P.C.Chung;C.F.Chen : Unsupervsed texture 15