Note: Descriptions are shown in the official language in which they were submitted.
WO 92/00~23 PCI-/GB91/01018
20~12~
Au~o~atlc carcass grading apparatus and ~ethod
This invention relates to methods of and apparatus for the
automatic grading of carcasses In abattoirs, processing plants
5 and the like. Hitherto, grading of carcasses has been performed
manually and thus has been subject to variations between and
within operators. The assessments made from such a manual
operation are totally subject~ve. ln the case of beef carcasses
there is an additlonal problem of perspective, the hind port~on
10 often belng several metres away from and above the grader. With
the intention of el~minating operator variabil1ty, an automatic
inspection system us~ng video cameras and itrtage analysis has
been devlsed.
Carcasses are graded according to officlally accepted
15 criterla, which vary from country to country. A transparent
system has been devised on wh~ch it is possible to superlmpose a
variety of grading strategies. Examlnation of the total carcass
as proposed in this method, gives information on speclfic sex,
dlstrlbution and other attributes ~o wh~ch weighting factors
20 ~such as carcass weight and size) can be applied, ~f necessary,
to determine an overall grade. Meat yield is calculated on the
basis of normalised measurements, that ls, by taking one or more
views of the carcass and shrinking or expanding the image to a
standard size and then compar~ng it with historic data from
25 previous carcass measurements for which the meat yield and nteat
yield distributlon has been determined. Initially this
informat~on may be less preclse than required, but as the
database on which such predictions are based expands, the
precision of yield predictions and the accuracy of yield
30 distrlbution will steadily improve. This enhancement may
progress e~ther in a passive manner by updates to the database
or, in an actlve way, by means of an intelligent, dynamic
database whtch continually expands through analysis of lts
acquired data.
Interpretation of the measurements has been based on an
.
. ' ~ ' '~ . ,
,~ . ~ . , - , :
,: . : . . , . ,
. , ., . .
.
.
WO 92/00523 PCI/GB91/01018
2n8S~2~
in-depth observation of carcass dressing operations. For
instance, there are three ways of taking the hide off an animal
- totally manual, semi automatic or fully automatic. A11 of
these operate in slightly different ways, with the manual
5 technique belng the most variable of the three. Th~s means that
the method of hide removal wlll determlne how much the fat and
thin muscles of the belly reg~on ~around the cutanus ~runc~i) is
actually ripped off. This, in turn, may indicate that the area
around the belly is fat-reduced or even fat-free but 1t is a
10 totally false indication because the fat has actually been
removed. System intelligence has been developed to account for
this and other aberrations that can be caused by dressing
practice.
European Patent Application No.0321981Al discloses a method
15 and apparatus for determinatlon of the conformation, fatness and
other properties of individual cattle classes. The silhouette
of a carcass or half carcass is recorded wlth a video camera ~n
a special light screening chamber and a calculatlon of the
parameters of the carcass made on the basis of an algorithm
20 derived using a number of subjective assessments made by manual
carcass graders. A second image may be recorded from the same
viewpoint using different illuminat10n.
One of the problems associated with this system is that it
requires every carcass to go ~nto a special viewing chamber. It
25 needs a special loading and unloading facility and is therefore
very limited in the number of carcasses that it can deal with.
In particular, it ~s restricted to slow slaughter lines, and, in
practice would not be usable for animals other than cattle.
Furthermore, it requires modification of a slaughter line to
30 enable it to be taken into use. It is restricted to
measurements taken from a single viewing point, and therefore
cannot cope with a widely varying population. It is apposlte
that the variability of beef carcasses in Denmark ~s very small
indeed. It comprises almost a single line or a single cross and
35 about 85~/~ of it is young bull-beef, produced for the Italian
.. . . . .
., . `
WO 92/00523 PCI/GB91tO1018
20~5124
market.
According to the present invention there is provlded a
method of grading carcasses after slaughter comprising the steps
of checking for the presence of a carcass in the field of view
5 of a camera, checking that the orlentat70n ol said carcass with
respect to the camera is in accordance w~th a predetermined
arrangement, exposing an l~age of the carcass from a plurality
of different viewpoints , determin~ng a plurality of dimenslons
of sa~d carcass from said 1mages and comparlng said dimensions
10 with stored values to determine the overall grading of said
carcass.
There is also provided apparatus for the gradlng carcasses
after slaughter comprising a plural~ty of video cameras adapted
to be positioned adJacent to a slaughter l~ne to expose an image
15 of a carcass on said slaughter line from a plurality of
different v1ewpo~nts, signal processing means to derive from
said images a plurality o~ parameters characteristic of satd
carcass, storage means to store a correspond~ng plural~ty of
parameters derived from prior measurement of reference
20 carcasses, comparator means to compare said plurality of
parameters wlth said corresponding plurality of parameters to
derive a further parameter indicative of the grade of said
carcass and indicator means to provide an indication of the
magn~tude of said further parameter.
The invention will now be particularly described with
reference to the accompanying draw~ngs and photographs, in which:
Figure 1 is a lateral view of a well formed carcass, side
F~gure 2 ~s a dorsal v~ew of the same carcass half,
Flgure 3 ls the posterior/anterior view of the same
carcass,
Figure 4 is a lateral view of a less well developed less
well fleshed carcass,
Figure 5 the dorsal view of that same carcass,
Fi~ure 6 is the posterior anterior view of that less well
fleshed carcass and
. . ~ ; , : .
. . . ,. :
. . .
WO 92/0û~;23 PCI/GB91/01018
2o~5t24
-- 4 --
Figures 7a and 7b are a schematic plan and side view of
inspection apparatus set up in an abattoir.
Referring now to Figures 7a and 7b of the drawings,
carcasses proceed along a rail R past a viewing position P. At
5 the viewlng position are three cameras Cl,C2,C.I . One camera is
positioned above the rail to provide a posterior/anterior view
of the carcass. The other two are positioned laterally so that
their l~nes of view are at 90 to one another and 45 to the
rail. It is therefore possible to obta~n lateral, dorsal and
10 posteriorlanterior views simultaneous1y. The same ob~ect~ve may
be achieved by positioning the cameras at a 90 bend in the
rail. In this case the cameras are again positioned above the
rail and laterally at 90 to one another, but in this instance
it is not necessary to orientate the carcass at 45 to the
15 rail. A third option is to have the cameras staggered
linearly. Such an arrangement would require the carcass to be
rotated through 90 after leaving the lateral viewing position
~C2) and before entering the dorsal v1ewing position (C3). The
posterior/anterior camera (Cl~ can be suitably positioned above
20 the dorsal viewing position.
It would be possible with the second arrangement, to
eliminate one lateral camera if ~t were posit~oned at the bend,
but care would have to be taken to ensure that ~mages of other
carcasses do not enter the field of view. In order to maintain
25 line speed, it may be prudent to introduce a preview camera C4.
Thls will enable the system to ensure that the carcass is
correctly aligned and orientated prior to entering the grading
station. If not, suitable remedial action to reposition the
carcass can be taken without affectiny the the continuity of the
30 automatic grading operation.
With all animals, carcasses or split sides are supported by
means of suspension from a hook or hooks running above a Fixed
rail. Wlthin any one factory, the des~gn o~ that hook will be
speciflc as their rail will only accept one type of hook, the
35 system being able to compensate for all variations in hook
.... . . . .
!; . .
WO 92/00523 PCr/GB91/01018
20~5124
-- 5 --
length.
The carcasses are viewed when in a (1) warm, (2) semi-warm
or (3) cold state; the preferred states being (2) or (3) which
occur some thirty minutes onwards after slaughter. Prior to
5 that the bulk of the fat ls translucent and any fat cover and
distribution information must be interpreted in a completely
dlfferent way because as the fat becomes cold ~t becomes more
reflective and less absorbent to light, this providing greater
accuracy on relatlve thickness data.
The temperature is not usually measured because the
conditions within the slaughterhouse and the cooling rooms are
generally fairly well controlled. Each system will be
calibrated to a defined set of environmental charaeteristics
specific to that plant or abattoir. If conditions change
15 radically, a new set of parameters will have to be installed.
Other information utilised in the grade calculations but not
provided directly by the system are cold carcass weight or
dressed carcass weight, that is when the hide and innards are
removed and any miscellaneous pieces of fat such as channel fat
20 and kidney knob fat are taken out. Such data are readily
available in most establishments via an automatic weighing
system.
As the carcass progresses through the system, it is first
correctly ori0ntated wlth respect to the predetermined position
25 of the cameras. If the orientation is not correct, but line
speed and operation maintalned, then computer transformations
and re-drawing would have to take place. This would require a
lot of computing power and take a considerable time. Such a
technique is possible on a beef grading line due to slower ltne
30 speeds. ~ith present technology, it i 5 impractical for pig,
lamb and all white meat lines.
The sequence is predetermined by the way the views are to be
interpreted. Preferably, three views are taken, although two of
these can actually be generated from the data of the other
35 five. Provision has been made for another camera (C5) to
.
'' ' ' ' .~ '' ' ~ ,' ' ' ',,'~' ' ' ' ' ' "'' "
, , ~ ' ~ . ' ~ "' ' ,
W O 92/005~3 PCT/GB91/~1018
2085~ 2~
provide data on the content and distribution of fat overlying
the ribs inside the dressed body cavity from a medial carcass
view. Data provided from such a view will only be required lf
fat distribution data provided by the other defined camera
5 positions is unclear, ambiguous or insuffic~ent for grading to
be accomplished. It is possible to magnify and zoom in on a
particular image. Although the saving in equlpment costs is
marginal, it does reduce the processing time, which on fast l~ne
throughputs can be a signlficant advantage, although it may be
lO accompanied by a marginal decrease in accuracy of yleld
prediction.
Three general vlews of the carcass are taken to prov~de
basic data on overall carcass characteristics such as length,
extent of minimum and maximum width, area distribution curves,
15 etc. Each of the views are taken with a standard,
non-interlaced high def~nition scan rate of l/50th second, up to
800 lines, although faster shutter speeds can be accommodated if
necessary. For most applications alternate line sampling is
sufficlent and permlts the picture information to be processed
20 much faster.
Three overall views provide the general carcass
information. The information in these views is then processed
by hardware, with a range of separation capabilities such as a
differentiating clrcuit as an edge detector, or by dynamic
25 software that shifts thresholds according to the type of carcass
being viewed.
~ ith certain species of animal, useful information can be
obtained from an interior view of the carcass. However, due to
si2e variations, it is advantageous to have a dynamic viewing
30 window9 the size of which is set in accordance wlth the
dimensions of the carcass which have been determined prevlously.
Very thick fat exhibits a high reflectance value. ~ith
sheep in particular there are some problems wlth the translucent
nature of the fat. ~ith normal direct lighting, because at
35 specific carcass locations there is a film of tissue, and as
.
WO 92/00!;23 PCI/GB91/01018
208512~
there is structure underneath it, although one is on top of the
other, lt acts on occasions like a mirror, particularly when the
carcass is st~ll warm. When it is at a particular angle it
bends the light causing it to reflect along this collagenous
5 material, glv~ng rise to reduced contrast. The problem can be
overcome by uslng d~ffuse or indirect llght.
In making yield predictions, lt is important to determine
the sex of the animal, because a helfer of identical
conformation and weight to a steer wlll have more saieable meat
10 on it than the steer just because the bone structure is
lighter. Neck muscle is very pronounced and developed in a
bull; the fat around udder region is smooth in cows and helfers
but rough in bulls and steers. This and other d~stingulshing
sex characteristics are determined by this system.
Image processlng for carcass evaluat~on is a
computer-intensive operation because the outline curves involved
are complex spline curves, that is, they are composite curves
made up of a large number of elementary components. The
angularlty, degree of curvature, length and other mathematical
20 and geometric descriptions at a specif~c point or region on the
carcass can be used objectively to define a shape at that point
~~ ~ compariscin ~n the grading procedure and quantitatively to
ct muscularity and thus lean meat yield.
mage analysis may be used in conjun~~ion with other
2S t Iniques and devices such as multiple w iength infra-red
measurement to measure or predlct moisture content and so to
predict the chemical liquid content.
~ hen the carcass is fresh, the water content can be
predicted with some accuracy; with butchered meat, however, the
30 water status needs to be determined. Certain surface cuts of
meat will lose water rapidly and lnterior cuts will lose water
very slowly. The water content will effect the apparent lipid
content. ~ith on-line production techniques, the ability to
measure fat and water content will enable accurate predictions
35 of lean content to be made. This is necessary, for examp1e, in
. ' ,.' ' ', ' '
:,
.. . . . : :
WO 92/OQ523 PCI/GB91/01018
. .~ . .
2 ~ 8 5 ~ 2 ~ - 8 -
the manufacture of a calorie-controlled product.
With pig grading it is possible to obtain information on fat
depth much more easily with other technologies rather than by
image analysis. There are practical difficulties in looking at
5 the cut surface of a pig, arislng from factors such as blood
splatter and fat smear. However, electrical resistance and
reflectance probes are unable to measure chan~es in confcrmation
and therefore t~nd to underestimate the lean content of well
conformed animals and over-estimate saleable meat yleld in
10 carcasses of poorer conformation. Also, because their pigs are
much more variable, European predictlon equations are not
suitable for the American pig population. In such applications,
the use of image analysis to pre~ict shape and muscularity,
together with other technologies for fat depth informatlon, will
15 provide more accurate predictions of percentage yield. For
beef, image analysis is well developed as a standalone
technique. For sheep, because there are different fat
distributions, particularly inter-muscular fat, other
techniques, such as ultrasound scanning can provide valuable
20 supplementary information for both grading and yield prediction.
With poultry, fat content and distribution around the hind
quarters of the animal, particularly the cloacal reg~on or
around the neck fat are good indicators. These display
themselves as changes in the carcass conformation. Therefore,
25 with appropriate adaptation the techniques described above can
be utilised to grade poultry carcasses w~th or without the
addition of techniques to quantify fat content at the positions
defined.
There are a number of objectives with this image analysis
30 system. These include the ability to grade carcasses, which is
important for producer payment, and predict yield, not solely
for the total carcass but for some specific parts, such as the
primals. The whole system not only has an attraction to the
abattoir producer or the legislature, but it becomes a large
35 marketing tool for meat purchasers such as the major
'. '
. . .
:'
WO 92/00523 PCr/GB91/0111~$ : ~
.
20~124 :
supermarkets, who at present de~ermine, from data at point of
sale, the quantity and type of meat being sold, but can only
subjectively determine wholesale meat purchases. This techni~ue
will enable them to make bulk meat purchases based on accurate
5 assessments of saleable meat. In addition, carcasses may be
bought on a quant~tative prediction of hindquarter to
forequarter meat distrlbution.
In order to achieve an accurate prediction of primal y~eld,
a cutting grid based on primitive measurements of length and
10 width may be drawn and superimposed on each of the basic
two-dimensional views. Since butchering methods vary from
country to country, the advantage of this system is a new
cutting grid can be superimposed to mirror the changes in
butchery techniques.
The v~deo cameras are specially constructed to achieve a
predetermined spectral response. This may be ~odified,
according to the specific circumstances, by use of filters,
structured lighting, or by selectton of a charge-coupled image
sensing device with the desired response. In some instances it
20 is not desirable to have filters in front of the cameras because
it degrades the quallty of the image. The spectral response is
chosen to increase and enhance the separation between the colour
of the components of the meat, particularly the fatness. For
example, in applications where the meat may be different shades
25 of red, purple or pink or the fat may be different shades of
whlte or yellow, selection of the green output (from an imaging
device with separate RGB output) gives the best separation for
the lean meat colour in beef. Any one of the other RGB outputs
may be used for the fat because the fat is reflecting across the
30 whole spectral wavelength range. Alternatively, for certain
applications data can be taken from each of the RGB outputs and
a polygonal database can be built up together wlth or in
addition to luminance/chrominance data on the way in which the
colours are changing relative to compositional variations in
35 component materials or alterations in physical or chemical
. . . . : .
.
WO 92/00~23 PCI'/G~91/0101~
~ S~
- 10 -
parameters during production. This i5 valuable for process
control operations. Particular products can be examined a
specific points in its process and from a database knowledge of
the components at those points may be interpreted to show that
5 there is, for example, too much fat in the product, or too much
liquor, the pastry is not the correct cons~stency and so on.
This information may be transmitted backwarcls and forwards down
a process line to modify the product in accordance with an ~deal
market requirement. For most applications, however, lt is
10 sufficient just to keep to luminance, grey scale, or composite
video information.
Because the efficiency of the system is influenced by a
number of parameters including the luminance of the llghts, the
reflectivity of each carcass, electronic drift in the cameras or
15 the system hardware, it is necessary for the system to be
dynamically self-calibrating on system startup and
self-compensating dur7ng operation. Self-compensation during
operation is achieved by a combination of camera auto-iris,
signal auto-gain and dynamic modification to system parameters
20 via the system software.
The system described herein is also capable of
self-~iagnostics by means of which it is continually monitoring
the performance of both the hardware and the software and will
visually and audibly warn the operator of failures/errors and
25 their degree of severity.
~ ith such a proliferation of shape and tissue distribution
information available from a number of two-dimensional views,
such a system i5 capable of generating three~dimensional
projections for individual carcasses. Using a process of
30 normalisation it is thus possible to create and modify cutting
pathways for robotic and automated systems from known and
defined shapes present in the database based on the measurements
quantified by the image analysis system. A similar approach can
also be applied to carcass dressing procedures.
By the inclusion of an intelliyent knowledge-based database,
,, . -
WO 92/00523 PCI/GB91/01018
-
20~12~
the system described above is capable of objective meat
inspection and other allied tasks.
In summary, image analysis is used to compile carcass data
based on historic experlence of what factors contribute to
S yield. A multiple thresholding techn~que enables fat
distribut~on and content to be derived. Complex shape
information can be stored as simple elements of spline curves.
Other useful data, including carcass weight and specimen sex,
can be used to bias the data as necessary. For particular
10 curves and shapes tolerance databases can be set up. This shape
information will be a mixture of area measurement, connectivity
analysis, length, widths, boundary points and edge detectlon.
Predetermined dlvisions can be set for each of these
measurements and a bias introduced to those measurements to
15 enable accurate grade and precise yield lnformat'ion to be
generated. In conjunction with other information, accurate
grading and yield pred~ction can be made for the three major red
meat species. With some modification this can be extended to
poultry.
This system also provides the basis for rapid automation of
butchery and dressing techniques to be developed. The addition
of artificial intelligence extends system use into areas of
carcass welfare such as meat inspection.
A flow diagram for a preferred method of carcass grading is
25 shown in Table 1.
. ' . - '
: :. : ~ : ,. . .
WO 92/ûO~Z3 PCI-/GB91/01018
. .
2 0 ~ S ~ 2 4 - 12 -
Table 1
1. Check calibration
2. Carcass present/absent
5 3. Check carcass orientation
4. Measure overal1 dimensions for each v~ew (fat areas/bulk)
5. Re-scale grid and superimpose
6. Obtain area data w1th different thresholds wlthin each cell
(area)
10 7. Qbtain shape conformation/information withln each cell (edge
connectivity)
8. Determine sex characteristic if necessary
9. Determine fat distribution within each cell for each view
10. Determine conformation/muscular1ty within each cell
15 11. Compute grade (reference to database/algorithm on raw data)
12. Compute fat/conformation interaction
13. Compute overall yield
14. Compute yleld within each tell (primals)
lS. Output data ~stamp carcass/sort carcass~
',.
. . :
. . .
.
: . ' .
.
