4. Antecedentes procedentes de la psicología
4.7. Viktor Krankl (1905-1997)
the line. The viscera are always discarded once removed from the line. The inspectors must also make sure that the evisceration has been done properly and otherwise remove the carcass from the line.
Heart
Lung
Liver
Gizzard
Fig. 2.2:Vital parts of the broiler viscera. Heart and liver are the most important organs for the visual health inspection as these are most frequently affected by diseases.
Computer vision could aid the inspectors at both stages. It could detect and remove unhealthy viscera from the conveyor and still leave the secondary inspection to a veterinarian, as it is today. Another solution would be a system with an extra conveyor. Of the many broilers that are slaughtered only about 0.7 % are unhealthy, so the vast majority will be healthy [18]. Samples that are clearly healthy would continue on the same conveyor and clearly unhealthy samples will be discarded automatically. The remaining samples would be side-tracked onto the extra conveyor for manual inspection. This way the veterinarians can focus on the borderline cases and they will have more time to do the inspection.
2.2
Related Research
As food production speeds are going up, so does the need for automation inspection. The availability of hyperspectral imaging, increasing frame rates and spatial resolution, has broadened the possibilities within food inspection. From detection of bruises on apples [20] to predicting the tenderness of red beef [21].
Inspection of broilers are no exception and it already happens at multiple stages in the production. Broilers are monitored in the rearing houses by the
Chapter 2. Background
farmers to keep track of the feed consumption to make sure they are follow- ing the ideal growth plan. This is a measure for the entire flock and then divided onto the number of birds. Additional systems have been developed to monitor weights on individual birds either mechanically or via computer vision systems [22, 23]. Even though only a small percentage of a flock is measured, it still gives a more varied image of the weight distribution.
Most research within this field has focused on the post mortem inspection of the carcasses. Many have investigated the problem of detecting skin tu- mours, which leaves a circular scar but often without immediate discoloura- tion. However using hyperspectral imaging many have been successful in detecting skin tumours [24–28]. Systems developed by [25, 26] uses PCA to reduce the large number of wavelengths into a small number of principal components. Both collected their samples in a laboratory environment us- ing a hyper spectral line scan camera. The two works reported classification rates on pixel level of 96 % and 97.18 %, respectively. [24, 28] employed fea- ture selection methods which can be used to reduce the number of recorded spectral bands and thereby speed up recording time. With classification at tumour level they achieved a true positive rate of 90.6 % and 80.0 %, respec- tively, but [24] also reports over twice the number of false positives compared to [28].
Fig. 2.3:Carcass with tumor captured at 12 different wavelengths. Source: [27]
Researchers at USDA Agricultural Research Service have long been work- ing with universities to develop a system for classifying wholesome and un-
2.2. Related Research
to systemically diseased chickens which are detectable by their smaller size or reddish skin. In [29] a dual camera system was used to capture images at 570 nm and 700 nm of every bird. The images were divided into 107 blocks and the average intensity of each block was fed to a neural network. The system achieved 94 % and 87 % accuracy for wholesome and unwholesome chicken, respectively, working in-line at a speed of 91 birds per minute.
Starting with an offline multispectral setup, [30] investigated sets of three and four wavelengths captured with an area scan CCD camera and achieved accuracies of 95.7 % and 97.7 %. Through several iterations, now using a line-scan camera, [31–33] developed a system capable of running at 140 birds per minute identifying 99.8 % of the wholesome birds and 97.1 % of the unwholesome. The system used fuzzy logic to score each pixel from the breast area and the average of all scores to classify the broiler.
A lot of the development in quality control is going on in the industry. The companies Marel, Meyn, Foodmate and Baader Linco all have products for quality grading of broiler carcasses, but the details are very limited. Typ- ically, they use a front or back camera or both. Marel reports that they detect broken wings, red and blue bruises, faecal stains, remaining feathers and skin damage [34]. Meyn looks for similar defects but also have a foot pad in- spection system for detecting burns in the footpad [35]. The feet are typically exported to Asia, but the burns are also a measure for broiler welfare. A high percentage of foot pad burns could indicate poor or inadequate litter in the rearing house. Foodmate specifies similar features as Marel and Meyn but also includes missing parts, one leg hanging, hock burns and bile stains [36]. Hock burns are like foot pad burns while bile stains typically occur due to a bad evisceration. Baader Linco detects defects similar to those by Food- mate [37]. Besides the usual front and back camera, they have the option to install two side cameras, which can increase the detection rates beneath the wing. None of the companies specifies any detection rates.
During the evisceration there is a risk that the broiler carcasses become contaminated with faeces, ingesta, bile, etc. These can carry bacteria or foul taste and infected carcasses must either be removed from the line or washed. [38] developed a system for detecting faeces and ingesta on the surface of broiler carcasses at a rate of 140 birds per minute. They used a 3-CCD camera with optical trim filters at 515.4 nm, 566.4 nm and 631 nm and achieved an accuracy of 96 % in detection rate but also reported "many false positives". [39] developed a fully automatic system for detecting contaminants on the carcass surface. The computer processing the two images, captured at 500 nm and 710 nm, also control a spray system to wash and disinfect the carcasses. The true positive rate of the system is 91.6 %.
Wooden breast is a relatively new phenomenon within poultry produc- tion. The abnormality affects the breast muscles where areas become pale and hard [40]. [41] developed a system for rapid on-line detection of wooden
Chapter 2. Background
Fig. 2.4:Image of a commercial quality inspection system. Source: [34]
breast using NIR spectroscopy. Measuring the protein content, they achieved a 100 % accuracy in detecting wooden breast from normal fillets on a test set of 79 samples. The work was done on chicken fillets after cut-up and with the skin removed.
Inspection of viscera has received less attention from both research in- stitutions and companies. The viscera are transported alongside the carcass after evisceration until it has passed the manual inspection, after which the edible organs are harvested from the viscera set. Quality inspection can hap- pen after harvest where the organs can be presented individually but food safety inspection must be done while the viscera travels along side the corre- sponding carcass. Here it can be a challenge as the organs are unorganised and can cover each other.
[42] developed a method to detect spleen enlargement in turkeys as this abnormality is related to virus diseases. The spleen and the liver show sim- ilar colours in regular RGB images but they found that the spleen was easy segmentable in UV lighting and achieved a detection rate of 95 %. [43] pho- tographed both heart and liver in an effort to classify the four classes: airsac- culitis, cadaver, normal, and septicaemia. Using white light, they found that the liver couldn’t be used to separate normal and septicaemia and the heart couldn’t be used to separate normal and cadaver. By combining the models for both organs, they were able to classify the birds into the four classes with an accuracy of 82.5 %. Similar discoveries were made by [44] who used NIR to capture images of chicken hearts to detect the diseases: airsacculitis, as- cites, normal, cadaver, and septicaemia. Cadaver proved must difficult again and scored 82% true positives in the five-class problem, whereas the other diseases were detectable with rates between 92 % and 100 % of the 125 test