Fig. 4. Schematic of the fish pond in which fish training experiments were conducted. a Concrete pool; b feeding station with blinking light; c partitions; d Plexiglas channel.
The training of the fish was carried out in three phases. During the first phase fish were trained to move to the feeding platforms in response to a visual signal. Each
training session lasted 1 min. Food was placed on one platform at a time, and a blinking blue light was operated above that platform for 20 s prior to food placing,
for 30 s during feeding and 10 s after feeding.
Four training sessions were carried out every day. The total amount of food provided was 1 of the fish biomass, and no food was given outside the training
sessions. During the second phase, fish were trained to approach the feeding platforms by
passing through a small opening in a mesh partition. The pool was divided into two equal sections by two identical affixed wire partitions each of which had a small
20 × 20 cm opening. The openings of the two partitions coincided so that they formed a single passage which fish could use to move from one section of the pool
to the other. During the third phase training continued with the two partitions Fig. 4c moved away from each other and connected by a 100 cm-long transparent
Plexiglas channel, through their 20 × 20 cm rectangular openings Fig. 4d.
3. Results and discussion
3
.
1
. Species recognition Fig. 5 shows the scattergram of the whole-body MIs of the three species. The
grey mullet MI cluster is almost completely separable from those of the other two
species. In the single case in which the MI of a grey mullet was unexpectedly low, its image included a dark shadow which may have been the result of a sudden
movement of the fish at the instant when its image was grabbed. This shadow was not detected by the algorithm as a background; the fish appeared wider than
normal in its binary image and its MI was reduced. The MI values of the grey mullet were in general smaller than those calculated for dead fish in our previous
work Zion et al., 1999. This could be due to the fact that the grey mullet used in the present study were much smaller average mass, 40 g than those used in the
previous one average mass, 485 g and perhaps their elongated shape had not fully developed. The MIs of carp and St. Peter’s fish have completely overlapping ranges
of values, because the fish sometimes open their fins and other times keep them close to their body, greatly varying the shapes of their silhouettes. The boundary
threshold value between the grey mullet cluster and those of the other two species was set to the average plus two standard deviations of the MI values of carp and
St. Peter’s fish.
All grey mullet of one test set and 19 out of 20 of the other set were correctly classified. The average classification error for grey mullet was 2.
Fig. 5. Distribution of whole-body moment-invariant f
1
of the three fish species.
Fig. 6. Distribution of moment-invariants F
t
= sqrtf
1t 2
+ f
2t 2
of the tail ends 10 of fish length of carp and St. Peter’s fish. A few MI above 1.0 are omitted in order to illustrate the region of interest.
As explained above, the two ends of fish images which were not identified as grey mullet were extracted by determining the total fish length and ‘segmenting out’ 10
of that length from each end. The MIs of both ends were calculated and the larger one was used by the classification procedure. Though carp have a very typical tail
shape, their MI values vary significantly between tails which were widely open when their images were acquired and those which were less open. The more widely a tail
was open, the higher its MI value. The ‘openness’ of St. Peter’s fish tails had less effect on their MI values and it was, therefore, decided to set the boundary between
the MI values of the tails of the two species to the average of the MI value for the St. Peter’s fish tails plus two standard deviations.
Fig. 6 shows a scattergram of F
t
of the tail segments whose lengths were 10 of the fish lengths. Use of the threshold values calculated for F
t
, enabled 79 and 91 of carp and St. Peter’s fish, respectively, to be correctly identified average errors of
the two test sets. Fig. 7 shows a scattergram of df the percentage deviation of the MI of a given
tail from the MI of a rectangle with the same aspect ratio. Use of the threshold values calculated for df, enabled 91 of both carp and St. Peter’s fish to be
correctly identified average errors of the two test sets. Using df as the feature for classification proved to be better than F
t
. Though the typical tail shape of carp and St. Peter’s fish are different, there may be coincidences
which result in similar MI values. This can happen as a result of the fact that tail segments were defined to be 10 of fish length regardless of their width how open
the tail was at the time the image was acquired. By comparing tail MI value to that of a rectangle, which is much closer in shape to the tail of St. Peter’s fish
than that of carp, the difference between those tail shapes was enhanced. Some of the errors may be attributed to the fact that parts of the fins and
tails semi-transparent and if the background light is too strong the contrast between the fish image and the background may be good but parts of the tail
may be lost in the segmentation process. In future experiments it is intended to vary the light intensity in order to determine the best lighting conditions,
which generate good contrast between the fish and the background and yet keep the semi-transparent fins dark enough to be included in the segmented
object. In one case, the tail of a St. Peter’s fish was damaged and therefore lost its typical shape. In another case, a St. Peter’s fish image was acquired,
in which the tail looked like a carp’s tail, possibly because of a backward motion.
3
.
2
. Fish handling and training Fish training results are presented here in a general descriptive way, since the
experiment was carried out without replicates, merely testing and demonstrating the feasibility of manipulating a group of fish in a manner that renders them
suitable for sorting by a computer vision system. During the first phase of
Fig. 7. Distribution of df, the percentage deviation of the MI of a given tail from the MI of a rectangle with the same aspect ratio. A few MIs above 300 are omitted in order to illustrate the region of interest.
Fig. 8. Actual view of a fish swimming from one side of the pool to the other, through a narrow transparent Plexiglas channel mounted between the two mesh partitions. Another fish is entering the
channel from the right.
training, the fish rapidly learned to associate the blinking blue light with food; after 4 days of training, they would move towards the feeding tables as soon as
the blinking started before any pellets were provided. During the second phase, as soon as the light signal was given the fish at-
tempted to approach the feeder. Some succeeded immediately and all learned within 3 days to use the rectangular opening 20 × 20 cm to reach the feeding
platforms. During the third phase, it took a few days for the fish to learn to use the
narrow passage effectively to move from one section of the pool to the other; the tunnel was then gradually narrowed from 20 × 20 cm to 20 × 5 cm to allow
the fish to swim only in single file and restrict their distance from a sensing device e.g. a CCD camera mounted at the side to a very narrow range Fig. 8.
The fish learnt to negotiate the narrow channel within 7 days.
4. Summary and conclusions