Contour based Split and Merge Segmentation and Pre-classiﬁcation of

Zooplankton in Very Large Images

Enrico Gutzeit

, Christian Scheel

, Tim Dolereit

and Matthias Rust

Fraunhofer Institute for Computer Graphics Research IGD, Joachim-Jungius-Str. 11, 18059 Rostock, Germany

University of Rostock, Institute for Computer Science, Albert-Einstein-Str. 22, 18059 Rostock, Germany

Arivis AG, Kroepeliner Str. 54, 18055 Rostock, Germany

Keywords:

Segmentation, Contour, Split & Merge, Pre-classiﬁcation, Shape Features, Zooplankton, Large Images.

Abstract:

Zooplankton is an important component in the water ecosystem and food chain. To understand the inﬂuence of

zooplankton on the ecosystem a data collection is necessary. In research the automatic image based recognition

of zooplankton is of growing interest. Several systems have been developed for zooplankton recognition on

low resolution images. For large images approaches are seldom. Images of this size easily exceed the main

memory of standard computers. Our novel automatic segmentation approach is able to handle these large

images. We developed a contour based Split & Merge approach for segmentation and, to reduce the non-

zooplankton segments, combine it with a pre-classiﬁcation of the segments in reference to their shape. The

latter includes a detection of quasi round segments and a novel one for thin segments. Experimental results on

several large images show that we are able to handle them satisfactorily.

1 INTRODUCTION

Zooplankton and phytoplankton are a signiﬁcant part

of the food chain of every aquatic ecosystem. Con-

sequently, they represent a fundamental parameter of

the ecosystem’s structure and act as an indicator of

its structural changes. The analysis can be done in

situ or in the laboratory with ﬁxed and ﬁltered plank-

ton samples. Expected results are the identiﬁcation

of plankton in these samples and also their classiﬁ-

cation into the appropriate species categories. The

usual approach starts with the deployment of plank-

ton nets followed by manual evaluation of the gath-

ered samples by a taxonomic expert. This is a labor

intensive and time consuming procedure which heav-

ily depends on rare expert knowledge. Hence, the au-

tomatic recognition of plankton is of growing interest

in biology. Many good reasons for automatic plank-

ton recognition are discussed by Gaston and O’Neill

(Gaston and O’Neill, 2004). A full solution for an

automatic or semi-automatic system should consist of

an image acquisition stage (laboratory or in situ) fol-

lowed by an image processing stage.

In this paper we concentrate on zooplankton sam-

ples acquired in a laboratory setup. We focus on the

segmentation of zooplankton and brieﬂy describe the

image acquisition as well as the zooplankton recog-

nition. Our acquisition method is realized with low

cost equipment producing high resolution images in

size of about 60000×60000 pixels with a pixel size

of about 0.4 µm. Images of this size easily exceed

the main memory of standard computers. Our follow-

ing novel automatic segmentation approach is able to

handle these large images. It comprises an adjusted

segmentation of all the objects in the image and a pre-

classiﬁcation to eliminate non-zooplankton segments.

The paper is organized as follows. Section 2

presents related work and in section 3 the problems

to solve are discussed. Our approach to image acqui-

sition and image processing is explained in section 4.

In section 5 we discuss our results and we ﬁnish with

conclusions and future works in section 6.

2 RELATED WORK

The recognition and quantiﬁcation process can be

supported by different techniques. The approaches of

(Sheldon and Parsons, 1967), (Fawell, 1976), (Her-

man and Dauphinee, 1980), for instance, do not in-

volve image processing methods, but electronic, col-

orimetric or ultrasonic measuring methods. Image

processing is a powerful tool for automatic recog-

nition and classiﬁcation of zooplankton. A typi-

cal analysis pipeline starts with the image acquisi-

tion and consists of segmentation, feature extraction,

training and classiﬁcation stages. Some ﬁrst exam-

417

Gutzeit E., Scheel C., Dolereit T. and Rust M..

Contour based Split and Merge Segmentation and Pre-classiﬁcation of Zooplankton in Very Large Images.

DOI: 10.5220/0004648604170424

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 417-424

ISBN: 978-989-758-003-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

ples are the works of (Katsinis, 1979) or (Jeffries

et al., 1984). Common shape features for classiﬁca-

tion are extracted from the threshold-segmented im-

ages of low resolution (256×256 px). The images in

(Chehdi et al., 1986), taken through a microscope, are

already slightly bigger (512×512 px). Our work con-

centrates on images many times larger and requires

advanced methods for automatic segmentation.

More recent and also technically more mature ap-

proaches mostly contain a special imaging device in

combination with a speciﬁc image processing system.

Imaging devices can be specialized to work either in

situ or in a laboratory. In situ devices are able to

image ﬂowing particles in a volume of water over

time. They are mostly constructed as mooring buoy,

towed body or as an installation on a ROV or AUV.

A ﬁrst example for an in situ device is the Video

Plankton Recorder (VPR) (Davis et al., 1992). An

identiﬁcation Software, the Visual Plankton, to rec-

ognize plankton in images taken by the VPR is de-

scribed in (Davis et al., 2005). The image acqui-

sition is adjustable between a resolution of 10 µm

and 300 µm per pixel. The segmentation is done af-

ter binarization and regions of interest (ROIs) are ex-

tracted. Another system is the Shadowed Image Par-

ticle Proﬁling Evaluation Recorder (SIPPER) (Sam-

son et al., 2001) with the Plankton Image Classiﬁca-

tion and Extraction Software (P.I.C.E.S.) (Luo et al.,

2004). Binary images with a pixel resolution of 50

µm are produced and hence, a segmentation is al-

ready given. The Underwater Vision Proﬁler (UVP)

(Picheral et al., 2010) is suited for real-time, in situ

particle counting, sizing and zooplankton detection

(Website, 2013b). This is done by integrating a dedi-

cated software within an intelligent camera, which is

able to remove a background image. The pixel reso-

lution of the images amounts to 174 µm. Further sys-

tems are the HAB-Buoy (Culverhouse et al., 2006)

coupled to the Dinoﬂagellate Categorisation by Arti-

ﬁcial Neural Network (DiCANN) software (Toth and

Culverhouse, 1999), the In Situ Ichtyoplankton Imag-

ing System (ISIIS) (Tsechpenakis et al., 2007; Tsech-

penakis et al., 2008), the ImagingFlowCytobot (Ol-

son and Sosik, 2007) and CytoBuoy (Dubelaar and

Jonker, 2000).

In a laboratory setup, like our scenario, water sam-

ples can be prepared and analyzed. The commer-

cial system Flow Cytometer And Microscope (Flow-

CAM) (Sieracki et al., 1998) is equipped with the

software Visual Spreadsheet (Website, 2013a). The

software is able to extract sample particle sizes of 2

µm to 2 mm after segmentation and stores the ROIs.

No details on the process are given. The system most

comparable to ours is ZooScan (Gorsky and Gros-

jean, 2003) with the software Zooprocess and Plank-

ton Identiﬁer (Website, 2013c). Zooscan is a water-

proof scanner with a maximal resolution of 2400dpi

resulting in a pixel size of 10.6 µm. On Zooscan’s

website is stated: ”if all available image dimensions

can be acquired, there is still some limitation for the

largest images processing. We are working on new

macros that will overcome these limitations soon”.

In Zooprocess a background image is substracted to

segment and extract individual objects. Zooprocess

and Plankton Identiﬁer are also able to process im-

ages from UVP.

Existing systems are not able to produce or handle

large images as required in our scenario. Particularly,

the published post-processing and segmentation algo-

rithms require the images to ﬁt into the computer’s

main memory.

3 DISCUSSION OF PROBLEMS

In our scenario, zooplankton species inside a counting

chamber with water (diameter 25 mm) have to be rec-

ognized. Images of the chamber are taken stepwise

by a microscope camera.

Figure 1: The counting chamber (left) and a part of it (right)

with small and large species are shown. The stitched tiles

are illustrated through lines and species with circles.

The images (tiles) are stitched together to a large

image (approx. 60000×60000 px), which is illus-

trated in ﬁgure 1. The zooplankton is located only

inside the counting chamber. Hence, it is helpful for

further processing to detect the contour of this cham-

ber. For recognition, a segmentation of the zooplank-

ton is needed. The following problems must be solved

to segment the zooplankton. Different zooplankton

species can vary in size and in positioning. Some

species can overlap multiple tiles due to a position

on the border and large species can be even bigger

than a single tile. On the contrary, small species can

be similar in size of an unimportant particle (non-

zooplankton). A further problem is the size of the

image, which is too big for processing at once. Also,

there are visible side effects after the stitching process

due to illumination variances that lead to high gradi-

ents at the borders of the acquired tiles.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

418

To overcome the problems we developed a con-

tour based Split & Merge approach for segmentation.

We do not split the image in homogenous regions

and merge similar regions like in (Kelkar and Gupta,

2008). In our approach we split the image in regular

tiles and merge the segmented contours in an itera-

tive way. Furthermore, to reduce the non-zooplankton

segments we implemented a pre-classiﬁcation of the

segments in reference to their shape. Our novel seg-

mentation and pre-classiﬁcation approach will be de-

scribed in following section.

4 OUR APPROACH

Our approach to segment and pre-classify zooplank-

ton is divided into ﬁve steps, which is illustrated in

ﬁgure 2.

Figure 2: Our segmentation and pre-classiﬁcation proce-

dure for very large images.

In step A the images are acquired with an in-

verted dissected light microscope setup. The im-

ages are stitched together and post processed into a

large grayscale input image. Thereafter, in step B,

the contour of the counting chamber is detected to

reduce the region to be processed in the following

steps. The region inside the chamber is segmented

in step C with a Split & Merge approach. As result

we get segments representing zooplankton and non-

zooplankton of different kinds. In step D we execute

the pre-classiﬁcation to simplify the recognition pro-

cess. Thereby, we can reduce the non-zooplankton

set in our scenario. Finally, a zooplankton recogni-

tion with a support vector machine (SVM) is done in

step E. In this paper we set the focus on step B, C

and D. An overview of the main notations used in this

paper is given in table 1.

Table 1: Main notation used in this paper.

large grayscale input image

C, P contour and point in general

the counting chamber contour

a zooplankton candidate contour

a contour in a quadtree leaf

∗

a complete and incomplete contour

Q, Q

quadtree, quadtree leaf

right and bottom quadtree neighbor

, P

start and end contour merging point

Ω

zooplankton candidate segment set

Ω

, Ω

set of quasi round and thin segments

Ω

reduced set of zooplankton candidate

4.1 Image Acquisition

We used counting chambers with a glass bottom

(Utermoehl, Kolkwitz) and an inverted dissected light

microscope setup (Olympus IX50). Due to the re-

quired resolution and the thickness of the sample we

added a motorized xyz-stage (Maerzhaeuser) to sup-

port tiled acquisition (xy) in multiple focus planes (z).

Sufﬁcient results were achieved with a 10× magni-

ﬁcation lens and a total number of 5 focus planes,

which leads to a data size in the gigabyte range.

The acquisition is controlled by the imaging software

arivis Browser with an acquisition module that is con-

nected to the microscope components and allows the

deﬁnition of the required acquisition strategies and

parameters. The single tiles are acquired as RGB

images with a conﬁgurable overlap for compensating

small variations in the motor stage positioning. Af-

ter acquiring the single images an automated stitching

and intensity compensation are performed to create a

large plane from the single tiles. Color information

is not utilized for subsequent steps, so all planes are

converted to one greyscale plane.To further reduce the

amount of data, an edge blending algorithm was ap-

plied that creates one merged image from the single

focus planes by selecting pixels with the highest con-

trast.

4.2 Chamber Segmentation

Zooplankton is positioned only inside the counting

chamber (see ﬁgure 1). For this reason, the detection

of the chamber contour C

is required. We have scaled

the large image to a width of 2048 pixel. The follow-

ing methods were implemented and tested for their

ability to extract the contour of the counting chamber

properly.

ContourbasedSplitandMergeSegmentationandPre-classificationofZooplanktoninVeryLargeImages

419

4.2.1 Longest Contour

The region inside the chamber is brighter and many

edges are in the darker outside region. A contour

seems to be directly calculable by binary segmenta-

tion and a strengthening of the edges. Firstly, a binary

segmentation with Otsu’s Method (Otsu, 1979) is per-

formed on the image. After a gaussian smoothing a

canny edge detection is applied, followed by some

morphological operations. Finally, the longest con-

tour in the resulting binary image is extracted.

4.2.2 Active Contours

The chamber is nearly round and enclosed by a con-

tour with high gradients. However, there is much

noise in the image leading to high gradients too. An

active contour (Kass et al., 1988) applied on a pre-

processed image seems to be appropriate. Therefore,

we ﬁrst perform a binary segmentation with Otsu’s

Method on a intensively smoothed image. After a

morphological closing, a Hough transform for circle

detection is used to determine the starting contour

points. The active contour algorithm is executed on

the binary image. We choose an equal weighting of

the internal and external energy terms, without exter-

nal constrains.

4.2.3 Region Growing

We also implemented a region growing approach,

since the chamber is positioned at the image center,

it is large in size and its region is even in brightness.

Consequently, a seed point inside the chamber can be

calculated reliably. We choose a center rectangular

area of 5% of the image size and use the pixel with

the lowest gray value as seed point. It can happen,

that the pixel is enclosed by noise and the segmented

region will get too small. In this case, different seed

points are chosen in the rectangular area until the seg-

mented region is bigger than 35% of the image. A

general problem of region growing is leaking. Leak-

ing is improbable in our case, because the chamber is

surrounded by a darker area and there are high gra-

dients from the inside to the outside area. After a

successful region growing with a ﬂood ﬁll algorithm

the longest contour is extracted. The resulting con-

tour can have some unwanted protrusions, because of

species or image artifacts nearby or on the chamber

contour. We eliminate the protrusions by using our

own polygon approximation. Therefore we analyze

the angle changes of the contour points. Our elimina-

tion process works with two points P

and P

. We use

and P

, an angle and a minimal/maximal distance

from P

to span a searching area. The point in the

searching area with the shortest distance to the con-

vex hull on the extracted contour is used as new point

. The old point P

is set to P

. The process is re-

peated until the starting point is reached.

4.3 Contour based Split & Merge

Candidate Segmentation

The input image I

can be larger than the available

main memory. Hence, we use a Split & Merge ap-

proach that ﬁrst, splits the large image into smaller

processable tiles and merges the results afterwards.

Additionally, to compensate for the high gradients at

the tile borders resulting from illumination variances

during acquisition, we choose a tilesize that matches

the size of the acquired tiles (including stitching pa-

rameters).

Figure 3: From left to right the leaf notes Q

of the quadtree,

not merged and merged contours of the zooplankton candi-

dates are illustrated.

To process the large image data in a way that

is structured and easy to handle, we make use of a

quadtree data structure Q. The quadtree is used for

splitting the image top down and managing the re-

sulting tiles and their neighborhood. A quadtree node

lying inside of or on the chamber contour C

is further

split, until a node has the size of a single tile (leaf node

). Only the leaf nodes Q

are processed further in

our approach (see left in ﬁgure 3).

After the construction of Q, a segmentation and

merging process is done, which will be described in

the following subsections. A set of segments Ω

resulting and available for further processing in step

4.3.1 Tile segmentation

The segmentation is done separately for every quad-

tree leaf Q

in Q. Zooplankton inside the grayscale

tiles typically varies in shape and texture, but consists

of pixels with similar high gray values. Hence, a sim-

ple and fast threshold segmentation is appropriate. A

single threshold is used for all tiles to avoid unaligned

fragments in the case of specimen located on tile bor-

ders. The threshold is determined on the whole input

image I

by Otsu’s Method. To reduce the inﬂuence

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

420

of noise, the image is smoothed and a binary segmen-

tation followed by a morphological ﬁltering (closing

and opening) is performed. If a tile in Q

is overlap-

ping the chamber contour C

, then the irrelevant out-

side contours are removed. Finally, all the contours

of the remaining segments C

are stored in the corre-

sponding leaf Q

4.3.2 Contour Merging

A single contour C

in Q

can be complete C

∗

incomplete C

. A complete contour has no further

neighboring contours. For each incomplete contour

a connected contour is searched for and a merg-

ing procedure follows. It is executed only on the leaf

node level in the quadtree and starts with the upper

left node. From left to right and top to bottom, con-

nectable contours are searched for in the node’s right

and bottom neighbors Q

. The following pseudo

code shows the procedure in short.

1: for each C

in Q

and C

in Q

2: if C

and C

are neighbored

3: search P

and P

in C

and split C

, C

4: merge C

and C

to C

5: if C

is a complete contour C

∗

6: put C

into Ω

7: remove C

and C

in Q

8: else

9: replace C

and C

with C

10: end if

11: end if

12: end

First, two merging points are calculated for one in-

complete contour C

in the current leaf node Q

(gray

quadrant in ﬁgure 4). The merging process is exe-

cuted for this incomplete contour and the incomplete

contours of the neighbor leafs Q

one at a time. The

merging start point P

and end point P

are calculated

by walking clock wise through the points on the con-

tour C

. P

represents the most left and bottom point

in Q

that has a matching point on the tile border in

. In a similar way P

represents the most right and

upper point. The points P

and P

split the contour

and C

into the contour parts C

and C

, respec-

tively C

and C

. A new contour C

is created by

merging C

in Q

with C

in Q

. C

and C

are elimi-

nated. The old contours C

and C

are removed and

the nodes are updated by linking the merged contour

to Q

and Q

. If a contour is complete, all links

will be removed and the ﬁnal contour becomes part

of Ω

. Before adding a contour to Ω

, contours

that are smaller than a minimum expected zooplank-

ton size have been already eliminated. An illustration

of the contour merging can be found in ﬁgure 4.

Figure 4: From left to right the contour merging part is il-

lustrated. The gray tile denotes the current leaf node Q

4.4 Pre-classiﬁcation

The segmented set Ω

contains zooplankton and

non-zooplankton. The non-zooplankton segments are

for example eggs, rotten plankton or other particles.

The aim of this step is to classify the elements in the

candidate set Ω

as one of the three sub-sets:

• reduced zooplankton candidates Ω

• quasi-round segments Ω

• and thin segments Ω

In our scenario the sets Ω

and Ω

are typical non-

zooplankton. The reduced set Ω

is produced for

the zooplankton recognition resulting in a more stable

classiﬁcation result.

Ω

= Ω

\ {Ω

, Ω

} (1)

The very fast working algorithms of our approach will

be described in the following subsections.

4.4.1 Detection of Quasi-round Segments

Our detection of quasi round segments is simple and

fast. Quasi round segments are characterized by a

contour shape going from round to quadratic. Eggs

and bubbles in the water are quasi round and should

not be passed into the recognition step. To be quasi

round, an object must exceed a threshold for one of

the two following similarities. The similarity to a

round shape is calculated from the zooplankton can-

didate area A

and its enclosed circle area A

. The

similarity to a quadratic shape is determined by using

the bounding box, whereby the side with the maximal

length l is important. The condition for a quasi round

object is:

Ω

= {C|∀C ∈ Ω

: (

> T

)∨(

> T

)} (2)

Good thresholds for the detection are T

= T

= 0.6.

4.4.2 Our Novel Detection of thin Segments

In our case, a thin segment C

characterizes a ﬁlamen-

tous object without big blobs on the inside. Two key

ContourbasedSplitandMergeSegmentationandPre-classificationofZooplanktoninVeryLargeImages

421

ideas play an important role. The ﬁrst one is an anal-

ysis of the pixel distances to the segment border in a

histogram representation. A thin object has a peak in

the low distance range of the histogram. The second

key idea is to normalize the histogram relative to the

radius of a circle with the same area. In comparison to

other shapes with the same area, a circle contains the

highest distances to the border. An exemplary illus-

tration of the detection approach is pictured in ﬁgure

Figure 5: From left to right: Thin objects and a circle with

identical areas followed by the corresponding normalized

histogram relative to the circle radius with the empty bin

amount E are illustrated.

At ﬁrst, a distance transform is applied on C

and

the area A is calculated. All distances are normalized

relative to the radius r of the circle. A second nor-

malization is done, so that r = 255. The distances

are represented in a histogram with 256 bins. Begin-

ning at the maximal histogram value (255), the empty

bins E are counted in direction of the histogram ori-

gin. Finally, a threshold T

is used for the detection

as follows:

Ω

= {C|∀C ∈ Ω

255

> T

}; relative to r =

(3)

A good threshold for the detection is T

= 0.6.

4.5 Zooplankton Recognition

The recognition is not the major part of our contri-

bution. Nevertheless, we explain it in short to com-

plete the argumentation. In our scenario, the set Ω

is used for zooplankton species recognition. For ev-

ery segment 52 features are extracted. The best fea-

tures are selected by an automatic selection approach.

Classiﬁcation is done with an SVM. The classiﬁca-

tion result is improved manually by an zooplankton

expert with few user interactions. The improved result

is stored and used for training an improved classiﬁer.

5 RESULTS

We evaluated the steps chamber segmentation (B),

candidate segmentation (C) and pre-classiﬁcation (D)

separately. The steps (C) and (D) are performed on

the results produced with the best method of step (B).

All results are compared with ground truth data gener-

ated by a zooplankton expert, who analyzed the sam-

ple, marked all existing specimen with a bounding

box and classiﬁed them. For our evaluation ﬁve large

images with ground truth data were available with fol-

lowing properties:

1. image 1: 62520×62000 pixels, 30×40 tiles

2. images 2,3,4: 64604×66650 pixels, 31×43 tiles

3. image 5: 89612×89900 pixels, 43×58 tiles

5.1 Chamber Segmentation

To evaluate the chamber segmentation, we segmented

the chamber in all images manually and compared it

to the automatic chamber segmentation results. For

the three chamber segmentation methods, we calcu-

late precision, recall and f-score. A value of 1 indi-

cates a perfect segmentation. The standard deviations

according to the evaluation measures are illustrated in

brackets. The evaluation of the three chamber seg-

mentation methods is presented in table 2.

Table 2: Chamber segmentation evaluation of the longest

contour (lc), active contours (ac) and region growing (rg)

approach.

precision recall f-score

lc 0.889 (13.5) 0.991 (0.36) 0.927 (8.66)

ac 0.993 (0.14) 0.997 (0.11) 0.995 (0.09)

rg 0.996 (0.17) 0.995 (0.67) 0.996 (0.27)

As can be seen in table 2, the active contour

and region growing approach lead to the best re-

sults. However, region growing is slightly better. The

longest contour approach leads to the worst results

with an extracted contour that can be a little bit frayed.

Images of the results can be seen in ﬁgure 6.

Figure 6: The images show the results of our three ap-

proaches for chamber contour detection. From left to right:

The longest contour, the active contour and the region grow-

ing approach are pictured.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

422

5.2 Zooplankton Segmentation

The segments of the set Ω

are compared to the

ground truth (manually marked bounding boxes). In

Ω

are zooplankton and non-zooplankton segments.

We only calculate the number of missing zooplankton

segments in Ω

, because the separation in zooplank-

ton and non-zooplankton segments is done in a later

step. In summary, only 0.5% of the ground truth were

missing after segmentation.

5.3 Zooplankton Pre-Classiﬁcation

We evaluate the pre-classiﬁcation algorithm by com-

paring the ﬁnal set Ω

with the ground truth zoo-

plankton data. For the sets Ω

and Ω

an evaluation is

only possible by subjective manual estimation. Evalu-

ated by our visual judgment, the sets contain correctly

classiﬁed thin and quasi round objects (see ﬁgure 7).

Figure 7: The left images picture quasi round and the right

thin objects. The contour of the segment is drawn in white.

The amount of classiﬁed segments per image is

shown in table 3.

Table 3: Number of segments pre-classiﬁed as thin (Ω

quasi-round (Ω

) and zooplankton candidate (Ω

image input output

|Ω

| |Ω

1 342 61 27 254

2 361 103 30 228

3 416 71 37 308

4 337 45 39 253

5 644 156 67 421

sum 2100 436 200 1464

The ﬁnal set Ω

contains zooplankton and non-

zooplankton segments. We evaluate the binary classi-

ﬁcation case with the measures speciﬁcity, precision

and recall (see table 4), whereby the best possible

value is 1. As can be seen, the speciﬁcity and pre-

cision of our pre-classiﬁcation approach is very good,

Table 4: Speciﬁcity (sp), precision (pr) and recall (re) of the

reduced zooplankton candidate set Ω

1 2 3 4 5 All

sp 0.94 0.97 0.97 0.95 0.95 0.96

pr 0.96 0.96 0.98 0.97 0.94 0.96

re 0.48 0.43 0.40 0.49 0.43 0.44

but the recall is not. In other words, the set Ω

con-

tains nearly all zooplankton segments, but contains

some non-zooplankton segments, too.

6 CONCLUSION AND FUTURE

WORKS

We presented a novel approach to segment and pre-

classify zooplankton located in a counting chamber

in very large images. It is the ﬁrst approach in the

ﬁeld of zooplankton recognition that is able to deal

with images larger than the available main mem-

ory of standard computers. We outlined how three

different methods can be used to detect the count-

ing chamber and one method to segment the zoo-

plankton with a contour based Split & Merge ap-

proach. The segmentation result contains zooplank-

ton and non-zooplankton segments. To reduce the

non-zooplankton species, we presented a solution for

pre-classifying the segments. The classiﬁcation algo-

rithm can detect thin and quasi round objects. For

the thin object detection, a novel approach was intro-

duced allowing a detection independently of the ob-

ject shape and rotation. The chamber detection eval-

uation shows very good results. The evaluation of

the zooplankton segmentation and pre-classiﬁcation

show that there are non-zooplankton segments that re-

main undetected.

Our current work focuses on detecting these non-

zooplankton segments by using the segmentation and

pre-classiﬁcation results. We are training one or

more classes representing non-zooplankton and we

are working on a multi resolution segmentation by

using the implemented quadtree. For the recognition

we plan to adapt and to develop special shape fea-

tures like in (Seng, 2013). To extend the system and

to be able to segment every zooplankton species with-

out exception an user-driven semi-automatic segmen-

tation like active contours in (Kass et al., 1988) and

(Ahmed et al., 2013) is planned.

ACKNOWLEDGEMENTS

We thank Joerg Voskamp, Arjan Kuijper, An-

dreas Fricke and Martin Feike. This paper con-

tains results of the research project ZooCount (no.

KF2212904KM0). ZooCount was funded by the Cen-

tral Innovation Program for SMEs, Federal Ministry

of Economics and Technology, Germany.

ContourbasedSplitandMergeSegmentationandPre-classificationofZooplanktoninVeryLargeImages

423

REFERENCES

Ahmed, F., Le, H., Olivier, J., and Bon, R. (2013). An ac-

tive contour model with improved shape priors using

fourier descriptors. VISAPP 2013 - Proceedings of the

International Conference on Computer Vision Theory

and Applications, 1:472–476.

Chehdi, K., Boucher, J., and Hillion, A. (1986). Automatic

classiﬁcation of zooplancton by image analysis. In

Acoustics, Speech, and Signal Processing, IEEE In-

ternational Conference on ICASSP ’86., volume 11,

pages 1477 – 1480.

Culverhouse, P., Williams, R., Simpson, B., Gallienne, C.,

Reguera, B., Cabrini, M., Fonda-Umani, S., Parisini,

T., Pellegrino, F., Pazos, Y., Wang, H., Escalera,

L., Moroo, A., Hensey, M., Silke, J., Pellegrini, A.,

Thomas, D., James, D., Longa, M., Kennedy, S., and

del Punta, G. (2006). Hab buoy: a new instrument for

in situ monitoring and early warning of harmful al-

gal bloom events. African Journal of Marine Science,

28(2):245–250.

Davis, C. S., Gallager, S. M., Bermann, N. S., Haury,

L. R., and Strickler, J. R. (1992). The video plank-

ton recorder (vpr): design and initial results. In Arch.

Hydrobiol. Beith., volume 36, pages 67–81.

Davis, C. S., Thwaites, F. T., Gallager, S. M., and Hu, Q.

(2005). A three-axis fast-tow digital video plankton

recorder for rapid surveys of plankton taxa and hy-

drography. Limnology and Oceanography-Methods,

3:59–74.

Dubelaar, G. and Jonker, R. (2000). Cytobuoy: a step

forward towards using ﬂow cytometry in operational

oceanography. Sci. Mar., 64(2):255–265.

Fawell, J. (1976). Electronic measuring devices in the sort-

ing of marine zooplankton. In zooplankton ﬁxation

and preservation, pages 201 – 206. Ed. by H.F. Steed-

man, Paris:UNESCO Press.

Gaston, K. J. and O’Neill, M. A. (2004). Automated species

identiﬁcation: why not? Philosophical Transactions

of the Royal Society of London Series B-Biological

Sciences, 359(1444):655–667+.

Gorsky, G. and Grosjean, P. (2003). Qualitative and quanti-

tative assessment of zooplankton samples. GLOBEC

INTERNATIONAL NEWSLETTER APRIL, 9.

Herman, A. W. and Dauphinee, T. M. (1980). Continu-

ous and rapid proﬁling of zooplankton with an elec-

tronic counter mounted on a batﬁsh vehicle. Deep

Sea Research Part A. Oceanographic Research Pa-

pers, 27(1):79 – 96.

Jeffries, H., Berman, M., Poularikas, A., Katsinis, C.,

Melas, I., Sherman, K., and Bivins, L. (1984). Auto-

mated sizing, counting and identiﬁcation of zooplank-

ton by pattern recognition. Marine Biology, 78:329–

334.

Kass, M., Witkin, A., and Terzopoulos, D. (1988). Snakes:

Active contour models. INTERNATIONAL JOURNAL

OF COMPUTER VISION, 1(4):321–331.

Katsinis, C. (1979). Digital Image Processing and Identiﬁ-

cation of Zooplankton. University of Rhode Island.

Kelkar, D. and Gupta, S. (2008). Improved quadtree method

for split merge image segmentation. In Emerging

Trends in Engineering and Technology, 2008. ICETET

’08. First International Conference on, pages 44 –47.

Luo, T., Kramer, K., Samson, S., Remsen, A., Goldgof,

D. B., Hall, L. O., and Hopkins, T., editors (2004).

Active learning to recognize multiple types of plank-

ton, volume 6. IEEE.

Olson, R. J. and Sosik, H. M. (2007). A submersible

imaging-in-ﬂow instrument to analyze nano- and mi-

croplankton: Imaging FlowCytobot. Limnology and

Oceanography: Methods, 5:195–203.

Otsu, N. (1979). A Threshold Selection Method from Gray-

level Histograms. IEEE Transactions on Systems,

Man and Cybernetics, 9(1):62–66.

Picheral, M., Guidi, L., Stemmann, L., Karl, D., Iddaoud,

G., and Gorsky, G. (2010). The underwater vision pro-

ﬁler 5: An advanced instrument for high spatial reso-

lution studies of particle size spectra and zooplankton.

Limnology and Oceanography: Methods, 8:462–473.

Samson, S., Hopkins, T., Remsen, A., Langebrake, L., Sut-

ton, T., and Patten, J. (2001). A system for high-

resolution zooplankton imaging. Oceanic Engineer-

ing, IEEE Journal of, 26(4):671 –676.

Seng, L. (2013). Contour-based shape recognition using

perceptual turning points. VISAPP 2013 - Proceed-

ings of the International Conference on Computer Vi-

sion Theory and Applications, 1:487–491.

Sheldon, R. and Parsons, T. R. (1967). A practical manual

on the use of the Coulter counter in marine research.

Coulter Electronics Sales.

Sieracki, C. K., Sieracki, M. E., and Yentsch, C. S. (1998).

An imaging-in-ﬂow system for automated analysis of

marine microplankton. Marine Ecology Progress Se-

ries, 168:285–296.

Toth, L. and Culverhouse, P. F. (1999). 3d object recogni-

tion from static 2d views using multiple coarse data

channels. Image Vision Comput., 17(11):845–858.

Tsechpenakis, G., Guigand, C. M., and Cowen, R. K., edi-

tors (2007). Image Analysis Techniques to Accompany

a new In Situ Ichthyoplankton Imaging System. IEEE.

Tsechpenakis, G., Guigand, C. M., and Cowen, R. K.

(2008). Machine vision-assisted in-situ ichthyoplank-

ton imaging system. Sea Technology, 49(12):15–20.

Website (2013a). http://www.ﬂuidimaging.com.

Website (2013b). http://www.hydroptic.com/uvp.html.

Website (2013c). http://www.zooscan.com.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

424