Machine Vision System Finds Application in Chicken Inspection
Tyson Food's poultry processing plant in New Holland, PA, got a look at the chicken plant of the near future earlier this spring.
That's when Yud-Ren Chen with USDA's Agricultural Research Service tested a prototype for automated chicken inspection. The prototype system, developed by Chen, "sees" the chickens with a visible and near infrared light probe and four cameras fitted with filters.
Developed over the past seven years, the prototype consists of four spectral cameras, a light probe, and a spectrophotometer which are all linked to computers. When the chickens, on hooks dangling from a moving chain, pass through a light beam, the interruption triggers a fraction of a second photo opportunity. One pair of cameras takes photos of the chicken's front, and the other pair, takes photos of its back. One camera of each pair uses a red filter; the other, a green one. This obtains images of the bird's front and back in two colors.
The light probe scans chickens at line speeds up to 140 a minute. A computer decides instantly whether a chicken shows signs of defects or disease, indicating with a red light that the bird requires close examination by a human inspector. If the prototype becomes commercially available, the computer would redirect rejected carcasses to a separate "re-inspect line."
Chen's group developed computer software that compares the images at different wavelengths, to determine if the bird is wholesome or not. Color differences can be caused by improper bleeding during slaughter or by blood-related diseases like septicemia. Skin textural differences can be caused by tears, bruises, or tumors.
"The same physical condition involving surface color and texture shows up differently under different wavelength filters," says Chen. "We use two wavelengths for comparison, to be sure we don't miss anything."
The cameras also detect body size. An abnormally small chicken requires closer inspection because disease may have stunted the bird's growth.
After a chicken passes the cameras, it crosses another light beam, this one triggering a scan from about an inch away.
A light probe illuminates a portion of the chicken with both near-infrared and visible light. The chicken absorbs some of the light, but any that is reflected is analyzed by the spectrophotometer and computer using software developed by Chen's team.
Differences between light shining on the bird and light reflection are due to variations in external skin color and texture and to internal blood color and tissue composition. In the prototype, the probe can analyze properties deep beneath the chicken's skin, stopping only at the abdominal cavity.
A red light on the frame near the computer setup indicates rejection.
"The prototype has an average accuracy rate of over 95%. We are continually improving this, and we achieved 100% accuracy in a recent test comparing the system's conclusions with those of a veterinarian," says Chen.
"To maintain accuracy, the system occasionally needs retraining with special software. This adjusts the computer to recognize the normal skin color of different chicken breeds or chickens fed different rations, for example. Processing plant employees would do this retraining by running self-learning software while flipping switches to show the system chickens that are normal and chickens that are not."
Though this prototype can spot unwholesome birds, it can tell the reasons for condemnation only in cases of septicemia or improper bleeding.
"But," notes Chen, "these two conditions account for over half of the carcasses removed from the processing lines."
Chen aims to expand the system's capabilities, along with incorporating advances in computer and sensor technology. He is also planning to test a new probe that will explore the whole chicken still without touching it and take color photographs of the abdominal cavity as well as the viscera. The color images would also be analyzed by the computer.
Almost eight billion chickens go through federally inspected plants annually, compared to less than three billion 30 years ago, Chen says. "If you are going to increase productivity without sacrificing the accuracy of meat and poultry inspection, you have to use machine vision and other automated sensors."
For more information contact Yud-Ren Chen, USDA-ARS Instrumentation and Sensing Laboratory, 10300 Baltimore Ave., Beltsville, MD 20705. Tel: 301-504-8450; Fax: 301-504-9466.